Abstract

Google Vizier has performed millions of optimizations and acceleratednumerous research and production systems at Google, demonstrating the successof Bayesian optimization as a large-scale service. Over multiple years, itsalgorithm has been improved considerably, through the collective experiences ofnumerous research efforts and user feedback. In this technical report, wediscuss the implementation details and design choices of the current defaultalgorithm provided by Open Source Vizier. Our experiments on standardizedbenchmarks reveal its robustness and versatility against well-establishedindustry baselines on multiple practical modes.

 

Quick Read (beta)