Seminár z teoretickej informatiky - Martin Takáč (7.3.2025)
v piatok 7.3.2025 o 11:00 hod. v posluchárni C
Prednášajúci: Martin Takáč (Zayed University of Artificial Intelligence)
Názov: Solving Combinatorial Problems Using Reinforcement Learning and Large Language Models
Termín: 7.3.2025, 11:00 hod., poslucháreň C
Abstrakt:
The Vehicle Routing Problem (VRP) remains a critical challenge in logistics and resource management due to its NP-hard nature and the complexity of stochastic demands. This talk explores solutions to VRP through the application of Reinforcement Learning (RL) and Large Language Models (LLMs). Firstly, we introduce a novel end-to-end framework for solving the Vehicle Routing Problem using RL. Our approach leverages the correlation between stochastic demands and other observable variables, demonstrating that non-i.i.d. stochastic demands can enhance routing solutions. In parallel, we discuss the application of Large Language Models (LLMs) to VRP through a novel prompting strategy called Self-Guiding Exploration (SGE). This method enhances LLMs’ ability to solve combinatorial problems by autonomously generating multiple thought trajectories for each VRP task. These trajectories are decomposed into actionable subtasks, executed sequentially, and refined to ensure optimal outcomes. Our findings indicate that SGE significantly outperforms existing prompting strategies, demonstrating its effectiveness in optimizing VRP solutions.
Martin Takac is an Associate Professor and the Deputy Department Chair of Machine Learning at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in the UAE. Prior to joining MBZUAI, he was an Associate Professor in the Department of Industrial and Systems Engineering at Lehigh University, where he had been on faculty since 2014. He earned his B.S. (2008) and M.S. (2010) degrees in Mathematics from Comenius University, Slovakia, and his Ph.D. (2014) in Mathematics from the University of Edinburgh, United Kingdom. Martin’s research focuses on the development and analysis of algorithms for machine learning, AI-driven scientific discovery, protein-DNA interaction studies, and the application of machine learning to energy systems. His work has earned him several prestigious awards, including the Best Ph.D. Dissertation Award by the OR Society (2014), the Leslie Fox Prize (2nd Prize; 2013) by the Institute for Mathematics and its Applications, the INFORMS Computing Society Best Student Paper Award (runner-up; 2012), and the Charles Broyden Prize for the best paper published in the 2022 volume of Optimization Methods and Software. He has received funding from multiple U.S. National Science Foundation programs, including a TRIPODS Institute grant in collaboration with Lehigh, Northwestern, and Boston University. Recently, he was also awarded three grants in partnership with the Weizmann Institute of Science. Martin has served as an Associate Editor for journals such as Mathematical Programming Computation, Journal of Optimization Theory and Applications, and Optimization Methods and Software. Additionally, he has taken on the role of Area Chair for top conferences such as AISTATS, NeurIPS, and ICLR, and served as a Senior Area Chair for NeurIPS 2024.