Short-Term Scientific Mission

Extensive Benchmarking of the PlatEMO Algorithms on the Bi-Objective Test Suite from the COCO Platform

Main theme: Benchmarking
Grantee: Martin Juříček, Brno University of Technology, Czechia
Host: Tea Tušar, Jožef Stefan Institute, Ljubljana, Slovenia

Start date: 2026-03-02
End date: 2026-03-31
Awarded: 2026-02-19
Report approved: 2026-05-04

Description

The objective of this STSM is to benchmark randomized optimization algorithms by connecting the PlatEMO framework to the COCO platform bi-objective test suite. The initial phase involves the technical preparation of a reproducible MLOps pipeline and the selection of suitable algorithms. Extensive benchmarking experiments will be executed using high-performance computing resources, followed by a rigorous statistical analysis of performance data. In parallel, a qualitative analysis with large language models will be conducted to categorize algorithms based on source code structures and evolutionary logic. The benchmarking results will be made available to the research community through the COCO platform.

Achievements

The mission successfully achieved the benchmarking of 129 algorithms from the PlatEMO library using the COCO bi-objective suite. A soft clustering methodology was implemented to classify algorithms into performance archetypes, allowing for a weighted understanding of shared behavioral traits across problem classes. In parallel, two complementary LLM-based frameworks were developed: one utilizing a vector database for architectural classification and another employing a structured questionnaire to map code similarities via LLM agents. These results established a robust foundation for a joint publication targeted at PPSN 2026 or GECCO 2027. All generated datasets were prepared for public release on Zenodo (https://doi.org/10.5281/zenodo.19071647).