From: 2025-06-16 11:00 to 22:00
Place: Seminar Room M 3170-73 in the M-building, LTH
Contact: emma [dot] tegling [at] control [dot] lth [dot] se

Date & Time: June 16th, 11:00-12:00
Location: Seminar Room M 3170-73 at Dept. of Automatic Control, LTH
Author: Karl Rydén
Title: Tell the Flows: GNN-Based Graph Optimization under Natural Language Soft Constraints
Supervisor: Pontus Giselsson, Erik Tegler, Daniel Perez (RISE)
Examiner:  Emma Tegling
Abstract: The development of 6G network management technology relies on the solution of large-scale combinatorial optimization (CO) problems on graphs. Many of these problems are NP-hard and must be solved approximately in practice. Despite decades of research and development, classical heuristics and approximation algorithms still struggle with accuracy and the computational demands of scaling to large problem instances. 

Recent advances in machine learning introduce the use of generative models that are able to outperform classical solvers after being trained to maximize task-specific reward functions on example problem instances. This reward function often has many modes, mirroring the tendency of CO problems to have multiple diverse solution candidates. In application, one might prefer certain solutions over others and want to specify that preference in natural language rather than by redefining the CO problem. 

In this thesis, we investigate the possibility of guiding one such generative solver toward desirable regions of the solution space by incorporating soft constraints expressed in text. Although results ultimately fall short, we investigate sources of error to motivate several promising directions for future improvement.