FunBO - Discovering Acquisition Functions for Bayesian Optimization with FunSearch

Watch a 34-minute AutoML Seminar presentation where Virginia Aglietti explores the development of FunBO, a novel method for discovering acquisition functions in Bayesian optimization using FunSearch technology. Learn how sample efficiency in Bayesian optimization algorithms relies on carefully designed acquisition functions (AFs) and the challenges of selecting optimal AFs across different optimization problems. Discover how FunBO leverages large language models to explore and generate computer code-based AFs, demonstrating superior generalization capabilities both within and outside training distribution functions. Understand how this innovative approach outperforms traditional general-purpose AFs and competes effectively with function-specific AFs developed through transfer-learning algorithms. Gain insights into solving the significant challenge of automatically identifying new acquisition functions that can enhance optimization performance across diverse problem domains.