AUTOMATED MACHINE LEARNING: NEW ADVANCES ON BAYESIAN OPTIMIZATION

Abstract

Recent advances in Bayesian optimization (BO) have delivered a promising suite of tools for optimizing an unknown expensive to evaluate black-box objective function with a finite budget of evaluations. A significant advantage of BO is its general formulation: BO can be utilized to optimize any black-box objective function. As a result, BO has been applied in a wide range of applications such as automated machine learning, robotics or environmental monitoring, among others. Furthermore, its general formulation makes BO attractive for deployment in new applications. However, potential new applications can have additional requirements not satisfied by the classical BO setting. In this thesis, we aim to address some of these requirements in order to scale up BO technology for the practical use in new real-world applications.