Exploiting landscape features for fitness prediction in university timetabling

Thomas Feutrier (1) , Nadarajen Veerapen (1) , Marie-Éléonore Kessaci (1)

1 ORKAD - Operational Research, Knowledge And Data

A small but growing number of papers have shown that landscape metrics can be useful for performance prediction, usually on classic unconstrained problems. In this paper, we consider the Curriculum-Based Course Timetabling problem, a heavily constrained problem known to have very neutral landscapes, and extract over 100 instance and landscape features to construct prediction models. An Iterated Local Search is used to sample the landscape, and the performance of both Simulated Annealing and a Hybrid Local Search algorithm are predicted using linear regression. Using as few as 4 features obtained via feature selection, our simple models are able to accurately predict the final fitness for either approach with an R-squared of approximately 0.95.

Mots clés: University timetabling; performance prediction; landscape analysis; local search

Domaines: Apprentissage [cs.LG]

Thomas Feutrier, Nadarajen Veerapen, Marie-Éléonore Kessaci. Exploiting landscape features for fitness prediction in university timetabling. GECCO ’22 Companion: Companion Conference on Genetic and Evolutionary Computation, Jul 2022, Boston, MA, United States. pp.192-195, ⟨10.1145/3520304.3528910⟩. ⟨hal-03791808⟩