Learning robust archetypes with Rashomon Partition Sets

This proposal offers a new approach to decision-making in the presence of uncertainty using the idea of Rashomon Sets--to enumerate and explore a small number of high (posterior) probability models, which is referred to as the Rashomon Partition Set (RPS) because each item in the RPS partitions the factorial space of covariates using a tree-like geometry. This proposal will expand the nascent literature on RPS development and evaluation in three key ways. First, it will develop a framework for robust causal discovery in high-dimensional settings using the RPS. Next, the proposal turns to design and uses the RPS to develop robust experiments using limited pilot data. Finally, the proposal develops a framework for rapidly incorporating new information into the RPS using a dynamic updating framework. These three innovations will facilitate rapidly and robustly learning from large volumes of high-frequency data.