Learning Human Acquisition Preferences Through Inverse Bayesian Optimization

• Activity Number: 335 - Spatial Smoothing and Bayesian Uncertainty Quantification
• Type: Contributed
• Date/Time: Wednesday, August 5, 2020 : 10:00 AM to 2:00 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Nathan Sandholtz* and Luke Bornn and Maurice Smith and Yohsuke Miyamoto
• Companies: Simon Fraser University and Simon Fraser University and Harvard University and Waymo
• Keywords: Bayesian optimization; inverse optimization; exploration vs. exploitation; human decision making

Abstract

We explore how well standard Bayesian optimization (BO) acquisition functions can explain human exploration-exploitation decision behavior. Data was collected on human decisions in a search task designed to force subjects to balance exploration and exploitation in their strategy. Using acquisition functions commonly employed in the BO literature, we solve the inverse problem for this task, estimating each subject's latent acquisition preferences based on the noisy optimization data we observe. We examine the fits to diagnose the differences and guide the construction of a better model. We find that subjects exhibit a wide array of acquisition preferences and that some subjects do not map well to any of the candidate acquisitions functions. We propose alterations to the set of acquisition functions to better explain human behavior in this search task.