Seminar: Luigi Nardi - Lund University, Stanford University, DBtune

Harnessing new information in Bayesian optimization

Abstract

This seminar delves into two key advancements that contribute to the capabilities of Bayesian optimization methods. We first focus on a novel information-theoretical acquisition function called joint entropy search (JES) which introduces a new quantity to improve statistical efficiency. We then explore the integration of active learning techniques and propose self-correcting Bayesian optimization (SCoreBO) which consciously learns the model hyperparameters while optimizing the black-box function. We emphasize the benefits of adaptively selecting informative data points and address the challenges related to model misspecification.

Notes

• References:
  • "Joint Entropy Search For Maximally-Informed Bayesian Optimization", C. Hvarfner, F. Hutter, and L. Nardi, NeurIPS (2022).
  • "Self-Correcting Bayesian Optimization through Bayesian Active Learning”, Carl Hvarfner, Erik Hellsten, Frank Hutter, Luigi Nardi arXiv (2023).
• Personal website can be found here .