Short-Term Scientific Mission

Measuring Algorithm Components' Contribution

Main theme: Algorithm Selection and Configuration
Grantee: Dindar Öz, Yasar Universitesi, Izmir, Türkiye
Host: Alexandros Tzanetos, Jönköping University, Jönköping, Sweden

Start date: 2025-07-21
End date: 2025-08-08
Awarded: 2025-06-05
Report approved: 2025-09-08

Description

To provide a solid foundation for the extensibility and usability of Randomised Optimisation Algorithm (ROA) components, we need to systematically analyse and compare ROAs in terms of their defining components and quantify how each algorithm component contributes to an algorithm. The STSM aims to introduce (preliminary) measures that provide insights into the algorithmic components’ contribution to a ROA and foster collaborations that will illuminate how algorithmic components affect ROA performance. During the STSM, systematic experiments will be conducted using a modular ROA framework to evaluate the impact of individual algorithm components on performance and behaviour, using benchmark problems from the BBOB suite. The analysis will apply behavioural methods such as recurrence plots, search trajectory networks, and machine learning to propose preliminary measures that quantify each component’s contribution to exploration, exploitation, and convergence.

Achievements

The STSM established a modular experimental framework enabling component-level analysis of Randomised Optimisation Algorithms (ROAs). Rich behavioural data structures (STN and PHM) were utilised to capture search dynamics, and novel interpretable metrics were designed to quantify exploration and exploitation via search space coverage and stagnation patterns. Preliminary results demonstrated these metrics’ ability to distinguish algorithmic behaviour and component effects on convergence and coverage. The methodology showed adaptability to multi-objective problems, offering insights into Pareto front behaviour. The work fostered ongoing collaboration, planned publications, and future studies, providing significant methodological advances toward improved ROA evaluation and evidence-based algorithm design.