B0441
Title: A python package for Bayesian experiment design and is application in physics experiments
Authors: Robert D McMichael - National Institute of Standards and Technology (United States) [presenting]
Abstract: The python package optbayesexpt implements optimal Bayesian experiment design for parameter estimation. A \textit{runs good} philosophy emphasizes ease of programming, minimal demand for statistical know-how, and fast-enough execution for automation of laboratory experiments. In the code, a particle filter with typ. $10^4$ particles represents the distribution of a handful of model parameters, and design setting values are chosen from a few hundred possibilities. The challenging computation of the Kullback-Liebler utility $U(d)$ is avoided using a pseudo-utility $U^*(d)$ that requires only 1D variances of forecast result distributions. Further, sampling noise is eliminated by using a single set of parameter samples to compute pseudo-utility for all candidate settings. In each measurement epoch, calculations of setting selection, likelihood of measured data, and Bayesian inference together typically require $\approx$ 10 ms of computation time. The optbayesexpt package is provided with eight example scripts and an interface module for communication with popular instrument control languages. In a laboratory demonstration, automated design with optbayesexpt reduced measurement time by a factor of 60 in magnetic resonance experiments. In simulations of Ramsey measurements (the canonical quantum measurement of energy differences), optbayesexpt outperformed published protocols.
