Two Novel Instance Selection Methods Combining Algorithm Performance and Landscape Analysis: A Comparative Study in Continuous Optimization

Two Novel Instance Selection Methods Combining Algorithm Performance and Landscape Analysis: A Comparative Study in Continuous Optimization

Článek WoS

A reliable benchmark library is essential for advancing research in global optimization by enabling fair comparisons and rigorous testing of optimization algorithms across diverse problem landscapes. In this article, we focus on instance selection methods, which aim to choose representative problems for evaluating algorithm performance. We present a comprehensive review of existing instance selection methods, highlighting their strengths and limitations, particularly in balancing the consideration of algorithm performance and the analysis of problem characteristics using exploratory landscape analysis. Building on these insights, we introduce two novel instance selection methods that leverage both algorithm performance data and landscape analysis information to construct diverse and informative benchmark sets. For evaluation, we benchmark our approaches against four existing instance selection methods on the recently expanded DIRECTGOLib v2.0 library. Our results demonstrate that the proposed methods effectively identify representative instances that capture a wide range of problem characteristics, enabling a more comprehensive evaluation of algorithm performance. These findings have significant implications for the development and assessment of new optimization algorithms, ultimately contributing to more reliable and robust solutions for real-world optimization problems.

A reliable benchmark library is essential for advancing research in global optimization by enabling fair comparisons and rigorous testing of optimization algorithms across diverse problem landscapes. In this article, we focus on instance selection methods, which aim to choose representative problems for evaluating algorithm performance. We present a comprehensive review of existing instance selection methods, highlighting their strengths and limitations, particularly in balancing the consideration of algorithm performance and the analysis of problem characteristics using exploratory landscape analysis. Building on these insights, we introduce two novel instance selection methods that leverage both algorithm performance data and landscape analysis information to construct diverse and informative benchmark sets. For evaluation, we benchmark our approaches against four existing instance selection methods on the recently expanded DIRECTGOLib v2.0 library. Our results demonstrate that the proposed methods effectively identify representative instances that capture a wide range of problem characteristics, enabling a more comprehensive evaluation of algorithm performance. These findings have significant implications for the development and assessment of new optimization algorithms, ultimately contributing to more reliable and robust solutions for real-world optimization problems.

Benchmark testing, Libraries, Optimization, Vectors, Runtime, Reviews, Cybernetics, Training, Reliability, Euclidean distance, Algorithm performance, black-box global optimization, exploratory landscape analysis, instance selection methods, numerical benchmarking

Benchmark testing, Libraries, Optimization, Vectors, Runtime, Reviews, Cybernetics, Training, Reliability, Euclidean distance, Algorithm performance, black-box global optimization, exploratory landscape analysis, instance selection methods, numerical benchmarking

STRIPINIS, L.; KŮDELA, J.; PAULAVICIUS, R.

2026

19.01.2026

IEEE

IEEE Transactions on Cybernetics

56

3

Spojené státy americké

1202

1215

14

https://ieeexplore.ieee.org/abstract/document/11224009