Meta-Learning Sampling Method for Quantifying Extreme-Event Statistics

In this work, we propose a novel meta-learning sampling method aimed at estimating extreme-event statistics. Unlike traditional approaches, which rely heavily on expert knowledge and iterative hand-engineering, our algorithm seeks to minimize manual derivations while enhancing performance. The core principle of meta-learning is to leverage information from previous tasks (source tasks) to expedite and improve outcomes in the current task (target task) through transfer learning. Although such a transfer-learning approach has been extensively explored in optimization tasks, it remains underutilized in uncertainty quantification, particularly for extreme events. To address this gap, we developed a meta-learning framework specifically tailored for extreme-event quantification. Our new sampling technique exhibits significant improvements in synthetic cases with diverse response functions, as well as in a real-world application for quantifying extreme ship statistics.