Münster 2017 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

T: Fachverband Teilchenphysik

T 23: Experimentelle Techniken der Astroteilchenphysik 2

T 23.9: Talk

Monday, March 27, 2017, 18:45–19:00, S 055

Dealing with Data/Simulation Mismatches in Machine Learning based Analyses — •Mathis Börner, Jens Buß, and Thorben Menne for the IceCube collaboration — Technische Universität Dortmund, Dortmund, Deutschland

The widespread use of machine learning algorithms in physical analyses require an intensive check for the compatibility between measured data and simulations. Since all frequently used algorithms use us more than one observable as the input. Therefore, the typical univariate comparison might not be sufficient. Furthermore, simulations always have finite mismatches, so it is necessary to decide whether they can be neglected or not. In this talk an approach utilizing machine learning algorithms to tackle both challenges is presented. The approach can be used exploratory to discover observables and areas in the observable space with significant mismatches. In a different application the approach is applicable to select observables with the lowest mismatch from a large set. Moreover, a way to show that no significant mismatches are presented in the simulation is shown. All presented methods are illustrated with results based on IceCube data.