Fitness Landscape k-Nearest Neighbors Classification Based on Fitness Values Distribution

Metaheuristic algorithms prove efficient in addressing complex optimization problems. Selecting a suitable algorithm for a specific problem type can improve optimization per-formance that can vary across different problem types. A problem can be characterized by its Fitness Landscape (FL). Exploratory Landscape Analysis approximates the problem's FL by features computed from only a small number of random samples. In this work, we investigate FL representation by a normalized histogram of samples' fitness values. The histograms are used as simple problem representations for a k-Nearest Neighbors classification to examine their ability to represent the problems. We provide a comprehensive study of classification performance on 24 single-objective benchmark problems. Especially, the impact of different sampling strategies, distance measures, and numbers of histogram bins on classification accuracy for different problems is examined. The results support the usefulness of this representation and overall approach and reveal some interesting trends.