Abstract

Combining multiple models has a long history in statistics, such as Bayesian Model Averaging, and more recently Stacking. However, there are few approaches available for combining models in a local sense and in a flexible and adaptive manner. We introduce Bayesian Model Mixing using a Bayesian Additive Regression Tree (BART) modeling approach for this problem. BART is a popular Bayesian regression tree model widely used in the modern regression setting. Its popularity stems from the ability to model complex responses depending on high-dimensional inputs while simultaneously being able to quantify uncertainties. In this talk, we first outline the proposed BART-based model mixing framework and then demonstrate it on a motivating climate application.