| This study introduced algorithm selection for hyper‐heuristics in extended HyFlex (Hyper‐heuristics Flexible framework). Hyper‐heuristics (HHs) are high‐level, black‐ box search and optimization strategies operating in a problem‐independent manner that can improve the capability of solving different problems under varying experimental environment. To some extent, hyper‐heuristics can be identified as cross‐ domain approaches that can be applied to a list of problems. Though there are effective hyper‐heuristic designs providing a certain level of generality in problem solving, no single hyper‐heuristic can always perform the best in different problem domains. Algorithm selection has been investigated essentially to address this issue, mainly for the problem‐specific algorithms, by automatically identifying the best algorithm for each given problem instance. This paper performed algorithm selection on selection hyper‐heuristics, delivering cross‐domain algorithm selection. For this purpose, this study mainly bore three research contributions. First, it extended the HyFlex benchmark for both number of problem domains and number of instances for each problem domain. There are in total 9 problem domains and 50 instances for each problem domain in the setting of this study. Second, this study introduced a suite of problem‐dependent features on an instance basis to characterize the individual instance for all problem domains. Another set of problem‐independent features were generated and utilized by the teammate of this research. Third, algorithm selection methods, namely ASAP and SUNNY, were performed to select the most suitable Hyper‐heuristic for a given instance based on the performance data and the problem‐dependent and independent features of the instance. Results revealed that algorithm selection methods on average outperformed all Hyper‐heuristics, and SUNNY performed slightly better than ASAP. |
