Leveraging Expert Knowledge in Bayesian Optimization for Engineering Design

Wouter Koppen; Tom Dhaene; Ivo Couckuyt; Sebastian Rojas Gonzalez

doi:10.5281/zenodo.17435324

Leveraging Expert Knowledge in Bayesian Optimization for Engineering Design

Wouter Koppen; Tom Dhaene; Ivo Couckuyt; Sebastian Rojas Gonzalez

Publisher: Zenodo
DOI: 10.5281/zenodo.17435324
Repository: zenodo.org
Language: en

Abstract

Engineering design often requires evaluating expensive simulations or experiments, making exhaustive search impractical. Bayesian Optimization (BO) addresses this by building a surrogate model—such as a Gaussian Process—to approximate the objective function and guide the search efficiently. Instead of brute force, BO uses an acquisition function to balance exploration and exploitation, focusing evaluations where they matter most. Many real-world problems are multi-objective, involving trade-offs rather than a single optimum. Traditional methods often exclude human expertise, which can lead to poorly defined objectives. Our approach integrates expert knowledge and grey-box techniques to improve relevance and efficiency. Experts can also help select a preferred solution from the Pareto front. To further accelerate optimization, we combine advanced strategies: - Function Networks to capture complex dependencies. - Multi-Fidelity BO to mix cheap approximations with costly evaluations. - Preferential BO to learn from comparisons (e.g., "design A is better than B") instead of exact scores. This expert-in-the-loop framework reduces costs, speeds up design, and ensures solutions align with engineering priorities.

Full text

Leveraging expert knowledge in Bayesian optimization for engineering design IDLAB –SURROGATE MODELING (SUMO) Wouter Koppen, Tom Dhaene, Ivo Couckuyt & Sebastian Rojas Gonzalez Bayesian optimization: smarter search for the best solution Imagine you have a black-box function that’s expensive to evaluate —like running a complex simulation or performing real-world experiments. You want to find the best design, but you can't afford to try everything. Bayesian optimization helps by building a model of your function and using it to smartly decide where to look next. We’re solving: Multi-objective optimization: when one objective isn’t enough Engineering design often involves huge search spaces and conflicting objectives, so there is rarely a single “best” design. Instead, we aim for a set of trade-offs. Traditional optimization often removes the human from the loop, which can lead to poorly defined objectives. Rather than removing the human from the loop, we incorporate expert input and grey-box insights to guide the search. This can be done a priori, interactively or a posteriori. But instead of brute force or random search, we use a surrogate model (e.g., a Gaussian Process) to predict outcomes and guide the search. Each step: •Use past data to update the surrogate model. •Pick the next design using an acquisition function (e.g., Expected Improvement). •Evaluate the real function at that point. •Repeat until we find something great —or run out of budget! We use: Function networks capture complex dependencies to improve model accuracy. Multi-fidelity Bayesian optimization combines cheap approximations with expensive evaluations. Preferential Bayesian optimization learns from comparisons (e.g., “design A is better than B”) instead of exact scores. Contact IDLab - SUMO [email protected] Universiteit Gent @ugent Ghent University