Automatic Configuration of AILS-II for the Capacitated Vehicle Routing Problem

Irace is a racing-based automatic configurator that supports parameter tuning by evaluating candidate configurations, reducing manual trial-and-error, and focusing the search on promising regions of the parameter space. It has been extensively used for parameter tuning of metaheuristic parameters, achieving outstanding results for a number of combinatorial optimization problems. In this study, we employ irace for the automatic configuration of Adaptive Iterated Local Search (AILS-II) for the Capacitated Vehicle Routing Problem (CVRP). AILS-II is a state-of-the-art metaheuristic with strong performance, especially on large-scale instances. Its default configuration was obtained through a one-factor-at-a-time analysis focused on the AGS benchmark instances. We investigate whether configurations obtained by irace can improve or match the default AILS-II setting across different benchmark families. For this, we evaluate irace elite configurations on AGS and X. While AGS represents the large-scale instances, X provides a broader and structurally diverse benchmark.  Our results show that the default configuration remains a robust baseline. On AGS, the best elite configuration matched the default aggregate mean gap, with both reaching 0.0629%. The elite reached its best solutions faster in 9 of the 10 instances, but the default remained slightly stronger in the instance-wise comparison of mean gaps and best solution values. On the set of X instances, the default obtained the best aggregate mean gap, with 0.0722%, while the best elite reached 0.0757%. However, the best elite was the most frequent winner in the Dolan-Moré performance profile and reached its best solutions faster in all 100 instances. The size-based experiments did not show a consistent benefit from specialized tuning, and the global tuning campaign produced the strongest elite configuration among the tuned alternatives. These results indicate that automatic configuration did not yield a dominant new setting, but it revealed competitive alternatives and a clear trade-off between average solution quality and time to best solution.