What is it about?
Many engineering and scientific design problems rely on simulations or physical tests that are slow and expensive, so an optimizer may be allowed only a small number of evaluations. A common idea is to combine several prediction models, known as surrogate models, in the hope that an ensemble will make better decisions. We tested whether this added complexity actually helps when only 40 to 80 evaluations are available. Across two standard multi-objective problems, a single Kriging model achieved the best average result in five of six settings. Simple model averaging never performed best, while adaptive fusion led only once by a very small margin. Although the statistical differences were not conclusive, the results suggest that when data are extremely limited, starting with one strong surrogate may be safer and simpler than automatically combining several models.
Featured Image
Photo by Giorgio Trovato on Unsplash
Why is it important?
This question matters because many real-world optimization tasks, such as engineering design and scientific simulation, can afford only a small number of costly evaluations. Researchers often assume that combining several surrogate models will improve reliability, but building and tuning an ensemble also adds complexity and requires enough data to estimate which models should be trusted. Our study challenges the idea that more models are automatically better. It shows that under severe data scarcity, a strong single surrogate can be a competitive and often safer starting point. This practical insight can help researchers avoid unnecessary model complexity, use limited evaluation budgets more efficiently, and decide when surrogate fusion is genuinely justified rather than applying it by default.
Perspectives
This is the first paper I completed largely independently as an undergraduate. I developed the research question and took primary responsibility for the experimental design, implementation, analysis, and writing. The process was challenging but rewarding, and it taught me how to turn an initial idea into a complete research study. One small regret is that, because the computational resources available to me were limited, I was unable to test a wider range of surrogate models, benchmark problems, and larger experimental settings. This limits how broadly the conclusions can be applied, but it also provides a clear direction for future work. I hope to continue learning from other researchers, expand these experiments when more resources become available, and gradually improve the depth and rigor of my work. I would be very grateful for any feedback or suggestions from readers.
Danqi Deng
Read the Original
This page is a summary of: DoSurrogate Ensembles Help Under Tight Budgets? AStudent Study in Expensive Multi-Objective Optimization, July 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3795101.3814712.
You can read the full text:
Contributors
The following have contributed to this page
