Two preprints on Bayesian Optimization

We’re happy to see the first two preprints from our group online on arXiv! Both papers focus on Bayesian Optimization, tailored to the needs of an experimental lab.

In BoTier, we present tool for hierarchical multi-objective Bayesian optimization with a-priori defined objectives. BoTier was born out of necessity: we needed a BO tool for handling hierarchical multiobjective scenarios, where secondary objectives come from experimental inputs, e.g. reaction temperature or catalyst loading. BoTier is a composite objective that can flexibly encompass a hierarchy between input- and output-dependent objectives – and can be used as a plug-and-play solution with any BoTorch workflow.

In CurryBO, we ask the question: How can we find general optima (e.g. reaction conditions that work for multiple substrates) using Bayesian Optimization? What began as a hackathon project has turned into a comprehensive benchmark study of different algorithms – with some surprising twists!