Date of Award
7-2-2024
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Experimental design offers an elegant model of many problems where one navigates within a vast search space seeking data points with certain characteristics. A multitude of applications in science and engineering fall under this umbrella, with drug and materials discovery being prime examples. The experimental design approach maintains a probabilistic model of the search space, and uses Bayesian decision theory accounting for this model to guide the accumulation of observed data to maximize an experimentation objective of interest. This dissertation explores Bayesian optimization and active search, two realizations of the experimental design framework that model discovery tasks. While existing solutions are available for these two problems under conventional settings, there are important scenarios to which these solutions cannot be readily applied, namely those of high dimensions or with multiple data sources, objectives that favor diversity in the collected data, and settings where efficient policy computation is crucial such as real-time systems and large-scale databases. We address these gaps, putting forward optimization and search policies with competitive empirical performance under their respective settings. The algorithmic solutions in our works provide practitioners with the tools to tackle a broad range of experimental design tasks, and ultimately advance machine learning-aided scientific discovery efforts.
Language
English (en)
Chair
Roman Garnett