Bayesian target-vector optimization for efficient parameter reconstruction
Abstract:
Bayesian optimization (BO) is a powerful method for the efficient optimization of expensive black-box functions. BO uses previous evaluations of the objective in order to train a stochastic model. The model, a Gaussian process, is subsequently used to derive promising parameter values by means of Bayesian inference. Compared with other global and local optimization methods, BO requires typically significantely fewer iterations to find the global minimum.
Recently, the method has been extended in order to match a vectorial function output to a target vector. We discuss the application of this approach for the efficient parameter reconstruction from a vector of measurement data. A comparison with standard Bayesian optimization and standard least square optimization shows that the new method can significantly speed up the parameter reconstruction.
Finally, we discuss the application of the method to a grazing-incidence small-angle X-Ray scattering (GISAXS) setup, for which parameter reconstruction is challenging when relying on standard BO; as well as an example of EUV scattering which contains more than 300 degrees of freedom.
This project is funded by the German Federal Ministry of Education and Research (BMBF, project number 05M20ZAA, siMLopt). Further, we acknowledge financial support from the Federal Ministry of Economics and Energy (BMWi) under the Central Innovation Programme (ZIM, Qupus).
QUNOM_10:00-14:00_Soltwisch_Heidenreich
Uhrzeit: 6.Nov.2020 10:00 AM Amsterdam, Berlin, Rom, Stockholm, WienZoom-Meeting beitreten
Meeting-ID: 814 8734 3502
Schnelleinwahl mobil
+12532158782,,81487343502# Vereinigte Staaten von Amerika (Tacoma)
+13017158592,,81487343502# Vereinigte Staaten von Amerika (Germantown)