@online{Ansari-et-al_2023,
TITLE = {Large-batch, Iteration-efficient Neural Bayesian Design Optimization},
AUTHOR = {Ansari, Navid and Seidel, Hans-Peter and Babaei, Vahid},
LANGUAGE = {eng},
URL = {https://arxiv.org/abs/2306.01095},
DOI = {10.48550/arXiv.2306.01095},
EPRINT = {2306.01095},
EPRINTTYPE = {arXiv},
YEAR = {2023},
MARGINALMARK = {$\bullet$},
DATE = {2023},
ABSTRACT = {Bayesian optimization (BO) provides a powerful framework for optimizing<br>black-box, expensive-to-evaluate functions. It is therefore an attractive tool<br>for engineering design problems, typically involving multiple objectives.<br>Thanks to the rapid advances in fabrication and measurement methods as well as<br>parallel computing infrastructure, querying many design problems can be heavily<br>parallelized. This class of problems challenges BO with an unprecedented setup<br>where it has to deal with very large batches, shifting its focus from sample<br>efficiency to iteration efficiency. We present a novel Bayesian optimization<br>framework specifically tailored to address these limitations. Our key<br>contribution is a highly scalable, sample-based acquisition function that<br>performs a non-dominated sorting of not only the objectives but also their<br>associated uncertainty. We show that our acquisition function in combination<br>with different Bayesian neural network surrogates is effective in<br>data-intensive environments with a minimal number of iterations. We demonstrate<br>the superiority of our method by comparing it with state-of-the-art<br>multi-objective optimizations. We perform our evaluation on two real-world<br>problems -- airfoil design and 3D printing -- showcasing the applicability and<br>efficiency of our approach. Our code is available at:<br>https://github.com/an-on-ym-ous/lbn_mobo<br>},
}