Scalable Bayesian Optimization via Online Gaussian Processes

Speaker: Marcel Neugebauer, University of Cologne

Abstract: Bayesian optimization is a state-of-the-art method for optimizing black box functions. It typically assumes that the unknown function is a sample path of a Gaussian process, which serves as surrogate model. As observations are collected, the belief is getting updated, enabling both prediction and uncertainty quantification. An acquisition function then guides the selection of new evaluation points by leveraging the posterior belief to balance exploration of uncertain regions and exploitation of promising areas. However, standard algorithms recompute the entire Gaussian process with each new observation or hyperparameter update, limiting scalability for large datasets.
To overcome this drawback, we employ a low-rank approximation of the Gaussian process kernel matrix that enables both the incorporation of new observations and online hyperparameter learning. This leads to online Gaussian processes and scalable optimization.

Biography: Marcel Neugebauer is a PhD student in the Department of Mathematics and Computer Science (Division of Mathematics) at the University of Cologne, supervised by Angela Kunoth. He previously studied mathematics at the same university and completed his master's degree in November 2024, with Angela Kunoth and Michael Multerer as his thesis advisors. His primary research interests focus on the numerical challenges of Gaussian processes, particularly in the context of Bayesian optimization.

Host: Prof. Michael Multerer