Bonn 2020 – wissenschaftliches Programm

Die DPG-Frühjahrstagung in Bonn musste abgesagt werden! Lesen Sie mehr ...

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

GR: Fachverband Gravitation und Relativitätstheorie

GR 5: Gravitational Waves

GR 5.4: Vortrag

Dienstag, 31. März 2020, 12:15–12:30, H-HS IX

Combining Greedy Approaches and Gaussian Process Regression in Gravitational-Waveform Modelling — •Florian Wicke — Max Planck Institute for Gravitational Physics (Albert Einstein Institute), Hannover, Germany — Leibniz Universität Hannover, Hannover, Germany

One of the goals of gravitational wave analysis is to estimate the source parameters of compact binary coalescence signals. This is done by comparing the measured signal to gravitational wave templates generated by a model. The most accurate templates are provided by numerical simulations. However, these are computationally expensive to produce. Varma et al. [Phys. Rev. D99, 2019] use a machine learning method called Gaussian Process Regression (GPR) to build a significantly less expensive surrogate model. GPR is an interpolation method that uses a limited number of numerically generated waveforms as a training set and predicts waveform templates for the surrounding parameter space. The training set is selected in a greedy manner using computationally inexpensive post-Newtonian waveforms. Here we evaluate whether the selection of the training set can be achieved solely by the GPR. To accomplish this, we use that GPR returns an uncertainty at each position in parameter space in addition to a predicted waveform. This uncertainty is then used as an error measure to select the next best training point position. We find that we achieve similar, if not better, convergence behaviour, but at the expense of worse computational efficiency.