On Generalization of ELA Feature Groups
Gašper Petelin, Gjorgjina Cenikj · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
Algorithm selection, i.e., selecting the most suitable algorithm for a specific problem, is a vital task in continuous black-box optimization. A popular strategy used to address this task is to characterize optimization functions using Exploratory Landscape Analysis (ELA) features, which are then utilized to train a machine learning meta-model to select the appropriate algorithm for that function. A significant challenge with meta-models trained on current benchmarks is their often restricted ability to effectively generalize to new functions, limiting their practical application. In this study, we investigate which ELA feature groups are the best at generalizing to previously unseen functions when performing algorithm selection. Using the Comparing Continuous Optimizers functions, novel functions are generated through affine recombinations of existing functions. For each ELA feature group, a meta-model is developed on these functions, enabling it to rank various optimization algorithms. Subsequently, these trained meta-models are assessed using functions that are increasingly out-of-distribution to what was observed during training. We show that most ELA feature groups do not generalize well to out-of-distribution functions, implying reduced effectiveness of selecting algorithms for unfamiliar functions. In such situations, meta-models using different ELA features for algorithm ranking often do not outperform basic predictions based on average ranks.