Speaker
Holger Schulz
(Fermilab)
Description
We present a range of improvements to the model parameter tuning program Professor. A severe limitation of the so far applied polynomial surrogates, namely the capture of $\frac{1}{x}$ behaviour of functional forms, is overcome
by a new algorithm that calculates multivariate rational approximations thus allowing for more truthful approximations and wider applicability of the method.
We also revisit the topic of tuning uncertainties and present a statistically
robust method to produce representative tuning-errors. Finally, we present our solutions to address the challenges faced when deploying the tool-chain on HPC resources.
Primary authors
Andy Buckley
(University of Glasgow (GB))
Dr
Sven Leyffer
(Argonne National Laboratory)
Dr
Juliane Müller
(Lawrence Berkeley National Laboratory)
Holger Schulz
(Fermilab)
