Summary: The derivation of stellar parameters by automated methods is only meaningful if some measure of confidence for the predicted value can be stated. In this work we introduce one possible method for obtaining standard errors and confidence intervals for stellar parameters Teff , log g, [Fe/H] and extinction Av as predicted by neural networks (NN). We applied the bootstrapping method to a feedforward NN for Blind Testing Cycle 2 medium band 1X and 2F photometry for end of mission magnitudes G=15 and 19 mag. We further tested whether the parametrization results can be improved if multiple noisy versions of a filter flux vector (for a given astrophysical parameter) are used in the training set. The obtained results show that the bootstrap standard errors only change significantly if the overall signal to noise ratio is high.

 

Bibtex entry for this abstract:

@UNPUBLISHED{LL:ICAP-PW-004,
author = {P.G.~Willemsen and C.A.L.~Bailer-Jones and T.A.~Kaempf},
title={{A}nalysis of stellar parameter uncertainty estimates from bootstrapping neural networks},
institution={},
year={2004},
month={September},
url={http://www.rssd.esa.int/doc_fetch.php?id=390246},
note={ICAP-PW-004},
type={Technical note}
}