BayBE: a Bayesian Back End for experimental planning in the low-to-no-data regime

Abstract

Due to its potential for high-dimensional black-box optimization and automation, Bayesian optimization(BO) is an excellent match for the iterative low-to-no-data regime many experimentalist practice in. It can be cumbersome to make BO work for real-world problems, as the application of code frameworks focusing only on implementing the core loop often requires substantial adaptation. Furthermore, with an extremely active research community, it can be challenging to find, select and learn the right components and code frameworks that best match the specific problem at hand. This is striking, as the BO framework in principle is highly modular, and such fragmentation is a headwind for the adoption of BO in industry. In this work, we present the Bayesian Back End (BayBE), an open-source framework for BO in real-world industrial contexts. Besides core BO, BayBE provides a wide range of additions relevant for practitioners, four of which we highlight in case studies in the domains of chemical reactions and housing prices: The impact of (i) chemical and (ii) custom categorical encodings; (iii) transfer learning BO; and (iv) automatic stopping of unpromising campaigns. These features can reduce the average number of experiments by at least 50%, cost and time requirements being reduced by the same factor compared to default implementations such as one-hot encoding. With this, we engage interested users and researchers from either industrial or academic backgrounds, and actively invite them to evaluate and contribute to the framework.