Towards an Improved Understanding of Features for More Interpretable Landscape Analysis
Abstract: Landscape analysis has received increasing attention in the literature, with a major focus on producing feature data to be used in a machine learning pipeline for automatic algorithm selection and configuration. In contrast, the original purpose of landscape analysis (i.e. to develop theory and techniques to understand the structure of fitness landscapes) has been somewhat overshadowed. This paper aims to clarify the purpose of landscape features, emphasizing the need for clear definitions of distance or similarity between optimization problem instances. It argues that current methodologies' shortcomings can be better understood from this perspective and advocates for more interpretable features and distances to gain explainable insights into problem instances, algorithm-problem mappings, and improved benchmarking of black-box algorithms.
Loading