Short-Term Scientific Mission

Reinforcement Learning for Dynamic Algorithm Configuration

Main theme: Algorithm Selection and Configuration
Grantee: Nguyen Dang, University of St Andrews, St Andrews, United Kingdom
Host: Carola Doerr, Sorbonne Université, Paris, France

Start date: 2026-09-14
End date: 2026-09-25
Awarded: 2026-07-30

Description

Reinforcement Learning (RL) offers a promising data-driven approach to Dynamic Algorithm Configuration (DAC), allowing metaheuristics to adapt their behaviour during search and potentially transfer learned strategies across related problem instances. However, relevant work is currently scattered across continuous black-box optimisation, hyper-heuristics, evolutionary computation, and related algorithm configuration settings. The STSM aims to consolidate this work into a structured review, identifying common design choices, empirical protocols, and open challenges. Building on this synthesis, a preliminary empirical research agenda will be outlined, focusing on benchmark requirements for RL-based DAC and on the ROAR-NET API as a candidate testbed for problem-independent, adaptive solver control.