Comparing Solvability Patterns of Algorithms across Diverse Problem Landscapes
Ana Nikolikj, Tome Eftimov · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
In the field of continuous single-objective black-box optimization, understanding the varying performances of algorithms across different problem instances is crucial. A recent approach based on the concept of "algorithm footprint" identifies both easy and challenging problem instances for an algorithm. However, a major challenge persists --- the lack of comparability among different algorithm footprints for effective benchmarking. This study introduces a solution through a multi-target regression model (MTR), which predicts the performance of multiple algorithms simultaneously, using a common set of problem landscape features. By establishing a common landscape feature set and using a single performance prediction model, not only can algorithm footprints be compared, but the explanations for the predicted algorithm performance derived with Explainable Artificial Intelligence (XAI) techniques can also be analyzed systematically. The methodology is applied to a set of three distinct algorithms, revealing their respective strengths and weaknesses on the Black-Box Optimization Benchmarking (BBOB) suite.