Description
Date depot: 25 mars 2024
Titre: Enhancing Dynamic Algorithm Configuration via Theory-guided Benchmarks
Directrice de thèse:
Carola DOERR (LIP6)
Encadrante :
Nguyen DANG (University of St Andrews)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: This PhD project is aimed at the development of automated Machine Learning (AutoML) techniques for learning dynamic control policies for black-box optimization heuristics. Building on recent works of the two supervisors, Carola Doerr (CNRS, LIP6, Sorbonne University) and Nguyen Dang (School of Computer Science, St Andrews), we will develop a diverse set of benchmark problems with rigorously proven optimal control policies. We will use these benchmarks to challenge state-of-the-art algorithm configuration techniques, such as deep reinforcement learning (deep-RL) approaches, as well as racing-based and surrogate-based configuration techniques.
The two main goals of this PhD project are hence
(1) the development of benchmarks with proven “ground truth” optimal policies and scalable characteristics such as the size of action and state space, degree of noise, and similar.
(2) the adaptation of deep-RL and configuration techniques to the black-box optimization setting.
This PhD thesis is a co-tutelle between Sorbonne University and St Andrews.