A1404
Title: Hierarchical Bayesian calibration with Bayesian committee machine
Authors: Sebastian Heinekamp - Paul Scherrer Institut (Switzerland) [presenting]
David Higdon - Virginia Tech (United States)
Abstract: Calibrating computer simulation code to experimental observations is an inherent task in any field using simulations to guide and inform experiments. Motivated by the goal to understand parameter uncertainties in particle accelerator experiments, hierarchical Bayesian calibration is used. Certain inputs (such as the beam injection amplitude) need to be calibrated separately for each experiment. The approach follows the calibration model suggested by a prior study, while a hierarchical description of the collection of parameters inferred for the individual experiments is added. The implementation makes use of a large number of simulations, leading to computational challenges for posterior exploration via Markov chain Monte Carlo (MCMC). To overcome this, the Bayesian committee machine (BCM) approximation is used for large GPs, allowing for leveraging high-performance computing infrastructure to speed up the resulting MCMC. The BCM offers problem and dataset-agnostic speed up and parallelization. Using the No-U-turn sampling algorithm allows us to leverage Julia's automatic differentiation capabilities. The improvements in runtime are demonstrated on analytic examples and simulation data for the Argonne Wakefield Accelerator.
