Bayesian optimization (BO) is a family of methods for global optimization of black-box expensive-to-evaluate objective functions. Such objective functions arise in a wide range of settings, including hyperparameter tuning machine learning algorithms, drug discovery, and A/B-testing-based design of mobile apps and online marketplaces. BO methods use a statistical model over the objective function, typically a Gaussian process, and Bayesian decision theory to sequentially select points at which to evaluate. In this talk, I will provide an introduction to classical BO methods, their applications, and some theoretical results. I will also discuss some modern extensions of these methods to the grey-box setting, i.e., where additional structure is known. In particular, I will describe how objective functions with a nested structure can be leveraged within the BO framework to improve efficiency.
Meeting ID: 993 5025 7333
To get the passcode please see the email or contact the organizers (Roman Aranda or Anup Poudel)