Files
Abstract
Bayesian Optimization (BO) has emerged as a powerful framework for optimizing black-box functions where evaluations are expensive. However, deploying BO in complex real-world scenarios presents significant challenges, including high-dimensional search spaces, the presence of unknown constraints, the need to balance multiple objectives, and the demand for efficient end-to-end modeling and decision-making. This proposal outlines a research agenda focused on developing novel BO methodologies that are efficient, scalable, and capable of handling constraints to address these challenges. The proposed work is structured around three main thrusts: (1) Efficient Bayesian Optimization via Regions of Interest (ROI): We explore methods to learn ROIs to make BO more efficient, particularly in high-dimensional or heterogeneous applications. This involves adaptive level-set estimation to identify promising sub-regions of the search space. (2) Efficient End-to-End Modeling and Decision Making (DRO): We investigate an end-to-end learning framework that moves beyond hand-crafted acquisition functions and myopic decision-making. This involves leveraging Decision Transformers for direct regret optimization, trained with a combination of simulated and real-world data. (3) Bayesian Optimization with Unknown Constraints: We develop principled approaches for BO problems where constraints are unknown and must be learned concurrently with the objective function. This includes COBAR for single-objective constrained BO, focusing on the principled treatment of feasibility, and CMOBO for constrained multi-objective BO, enabling a principled tradeoff among multiple objectives under learned constraints. These research thrusts aim to significantly advance the capabilities of BO, enabling its application to a wider range of challenging real-world problems such as protein design, material discovery, and drug development.