Author: Andoni Irazusta Garmendia

Thesis: Exploring Neural Combinatorial Optimization: Current State and Future Perspectives

Directors: Alexander Mendiburu / Josu Ceberio

Day: 16 December 2024
Time: 12:00h
Place: Ada Lovelace room (Faculty of Computer Science)

Abstract:

"Problem solving in combinatorial optimization often relies on hand-crafted heuristics, which, while fast and practical, are constrained by domain expertise and fail to leverage historical data. Neural Combinatorial Optimization (NCO) is an emerging field that uses deep learning to automate heuristic generation by learning from data, demonstrating early success by outperforming some heuristics. However, its broader applicability remains unexplored and there is a gap in understanding why and under what conditions the use of deep learning is most effective. This thesis investigates NCO across various combinatorial optimization problems, benchmarking its performance against a wide set of conventional approaches, and identifying its strengths and limitations. Key contributions include a comprehensive and critical analysis of NCO methods, novel neural constructive and improvement methods, and the integration of memory modules for improved exploration."