Joint BCAM-UPV/EHU Data Science and Artificial Intelligence seminar: Neural Improvement Heuristics: from Alchemy to Science

Date: Fri, May 5 2023

Hour: 12:00

Location: UPV/EHU Donosti, Faculty of Computer Science, room 3.1 and Online

Speakers: Andoni Irazusta (UPV/EHU)

Link to the session here

This talk presents a deep reinforcement learning approach to combinatorial optimization, addressing a significant limitation of the widely used hill climbing method. While effective, hill climbing requires a time-consuming evaluation of a vast number of potential solutions before identifying the most improving move. To overcome this limitation, a novel neural network model is introduced, capable of identifying a promising move on the first try, without the need for any solution evaluation. In addition, a recent criticism of the deep learning community is also addressed. Some have accused deep learning researchers of behaving like alchemists, using a trial-and-error approach to develop models without a solid understanding. This criticism will be discussed, while trying to provide a justification for every module of the proposed neural network.