Speaker: 黄思源, 北京工业大学

Inviter: 熊世峰

Title: High-dimensional Bayesian optimization via transformation and additivity

Language: Chinese

Time & Venue: 2026.02.03 09:30-10:30 思源楼 803

Abstract: We propose TABO (Transformed Additive Bayesian Optimization), a novel algorithm specifically designed for high-dimensional Bayesian optimization. TABO is built upon a flexible transformed additive surrogate model, which effectively captures the main structural properties of high-dimensional functions by learning a nonparametric monotonic transformation. This approach substantially relaxes the restrictive assumption of direct additivity in prior work, while retaining the computational efficiency of optimizing decomposable acquisition functions. We provide the first theoretical regret analysis for this framework, proving that the cumulative regret remains sublinear and scales linearly with dimension, even when the nonparametric transformation is learned online. This result formally validates TABO’s ability to mitigate the curse of dimensionality. Empirically, TABO outperforms state-of-the-art high-dimensional baselines even when the core assumption is violated, demonstrating the practical value of our more flexible model on challenging problems.