Investigating the landscape of a hybrid local search approach for a timetabling problem

Abstract

Curriculum-Based Course Timetabling is an NP-hard problem that can be efficiently solved by metaheuristics. The International Time-tabling Competition (ITC) 2007 was won by a hybrid local search (HLS) combining Hill Climbing, Great Deluge and Simulated Annealing. HLS remains one of the best local search algorithms to solve this problem. In this paper, we investigate the search landscape of 21 instances to analyze the behavior of the HLS components. We also propose a new distance metric that aims to be more robust and be less influenced by symmetry. Experiments show that the HLS and the embedded simulated annealing have the same general behavior but HLS leads to better robustness. This analysis strongly suggests that the HLS components and/or parameter values should be automatically configured to further improve performance.