AUTONOMOUS EXPERIMENTATION SYSTEMS AND BENEFIT OF SURPRISE-BASED BAYESIAN OPTIMIZATION

抄録

Five years ago the US Air Force Office of Scientific Research (AFOSR) issued a call for developing an autonomous material experimentation platform, where the system is able to execute, evaluate, and plan new experiments iteratively with minimal human intervention. Before and since then, there has been active research to accomplish this goal set forth in the AFOSR call. In this paper, we discuss the landscape of the existing autonomous experimentation systems with a focus on their intelligent control capability (i.e., the “planner”). A commonly used mechanism for planners is based on Bayesian optimization (BO) and, within the BO frame, a widely used acquisition function is the expected improvement criterion. To address autonomous systems' ability to handle surprising observations, we present our ongoing work in exploring a surprise-based BO method. We apply the method retrospectively to an autonomous experimental 3D printing system at the Air Force Research Laboratory and discuss some interesting observations.