Explorační analýza účelových funkcí

Abstract

Exploratory Landscape Analysis (ELA) is a method that extracts features describing the properties of an optimization function's landscape from a sample of points and their function values. In practice, these features are used for selecting a suitable optimization algorithm and predicting its performance on a given task. This bachelor's thesis investigates whether the 24 BBOB test functions can be grouped based on ELA features so that they match the five theoretical groups. For this purpose, a tool with graphical interface was developed in Python, which enables the computation of ELA features from the pflacco library, their preprocessing, clustering using K-Means, K-Medoids and DBSCAN method and visualization of the results. Experiments tested individual feature families, manually selected subsets and an optimized selection using a genetic algorithm. Clustering based on various feature subsets showed that for the BBOB benchmark functions it is possible to find better-separated clusters, which however do not capture the expected theoretical division.