Bayesian inference of expensive simulators: survey and comparison of design choices
Statistics and Computing · 2026-12 · 36 (6)
DOI: 10.1007/s11222-026-11015-3 ↗Kurz gefasst
Fachartikel in der Zeitschrift Statistics and Computing, online erschienen am 05.11.2026. Verfasst von Šimon Soldát und 1 weiteren Autor:innen. Die Arbeit nennt 75 Literaturangaben.
Abstract (Originalsprache)
We study Bayesian inference of expensive simulators: settings where the likelihood can be directly evaluated via forward simulation, but only a small number of evaluations (tens to hundreds) are feasible due to the simulator’s high computational cost. Under such tight budgets, traditional sampling methods such as MCMC are infeasible, making Bayesian optimization-based methods the state of the art. However, existing work is fragmented into functionally similar methods with ad hoc, uncompared design choices. We unify these approaches under a common taxonomy, where previous methods become particular design choices along three axes. Our second contribution is a systematic empirical comparison across these axes on 37 benchmark problems, spanning analytical toy problems, physics-based simulators, and higher-dimensional benchmarks. The surrogate model’s response proves to be the most important design choice. Notably, modeling the log-likelihood performs poorly despite its common use, while a new adaptive approach proves robust across all benchmarks. Based on the results, we provide actionable guidance on each design choice, along with a robust implementation in the form of a stand-alone Julia package.
Zitieren
APA
Soldát, Š., & Šmídl, V. (2026). Bayesian inference of expensive simulators: survey and comparison of design choices. Statistics and Computing, 36(6). https://doi.org/10.1007/s11222-026-11015-3
Vancouver
Soldát Š, Šmídl V. Bayesian inference of expensive simulators: survey and comparison of design choices. Statistics and Computing. 2026;36(6). doi:10.1007/s11222-026-11015-3
BibTeX
@article{Soldt2026,
author = {Soldát, Šimon and Šmídl, Václav},
title = {Bayesian inference of expensive simulators: survey and comparison of design choices},
journal = {Statistics and Computing},
year = {2026},
volume = {36},
number = {6},
doi = {10.1007/s11222-026-11015-3}
}
EUVIMED