Nullifying the Inherent Bias of Non-invariant Exploratory Landscape Analysis Features

Exploratory landscape analysis (ELA) in single-objective black-box optimization relies on a comprehensive and large set of numerical features characterizing problem instances. Those foster problem understanding and serve as basis for constructing automated algorithm selection models choosing the best suited algorithm for a problem at hand based on the aforementioned features computed prior to optimization. This work specifically points to the sensitivity of a substantial proportion of these features to absolute objective values, i.e., we observe a lack of shift and scale invariance. We show that this unfortunately induces bias within automated algorithm selection models, an overfitting to specific benchmark problem sets used for training and thereby hinders generalization capabilities to unseen problems. We tackle these issues by presenting an appropriate objective normalization to be used prior to ELA feature computation and empirically illustrate the respective effectiveness focusing on the BBOB benchmark set.