Short-Term Scientific Mission
Towards a Constrained Multiobjective Suite of Test Problems with Realistic Difficulties and Known Pareto Sets
Main theme: Single- and Multiobjective Optimisation
Grantee: Luka Opravš, Jožef Stefan Institute, Ljubljana, Slovenia
Host: Anne Auger, Inria Saclay Centre, Palaiseau, France
Start date: 2026-05-11
End date: 2026-05-22
Awarded: 2026-03-10
Report approved: 2026-07-02
Description
A wide range of algorithms exists for solving constrained multiobjective problems. To benchmark them efficiently, we generally need test problems with known optimal solutions (Pareto sets). However, most published test problems either originate from real-world applications, where the optimal solutions are unknown, or are artificially constructed and overly simplistic, often failing to encapsulate the difficulties found in real-world problems. The goal of this STSM is to investigate how realistic, yet understandable single-objective functions can be combined into a concrete suite of constrained multiobjective problems with known optima. In addition, we aim to release a Python test function generator designed to produce problems with properties encountered in real-world applications.
Achievements
The COBI (COnstrained BI-objective) problem generator was published on GitHub, and its description and theoretical background were made available on arXiv. The generator can produce problem instances with difficulties often observed in real-world problems and compute accurate approximations of their Pareto sets.
Different approaches for constructing a concrete constrained bi-objective benchmark suite were explored. Multiple methods for constructing problem instances were considered, and their properties were analysed both theoretically and empirically using prototype Python implementations. A method generating instances exhibiting a range of properties frequently encountered in real-world problems was selected, with the goal of ensuring that the performance of optimisation algorithms on these instances is indicative of their performance on real-world problems.
