Abstract:
Gravitational microlensing is one of the methods of the detection of exoplanets.
The NASA Roman telescope is expected to be launched in 2025, one of the
missions of which will be the Roman Galactic Exoplanet Survey. The number
of observations is expected to increase, which will require more automated and
faster methods of the analysis of the microlensing events. Recently, some of the
machine learning applications to classify the microlensing events and determine
the microlensing events' parameters have been presented. In this work, we make
use of and extend the feature-based methodology of estimating the microlensing
planetary parameters. The cases of predicting the mass ratio parameter and
multiple targets using this methodology are presented for the first time. The
sets of the most relevant features were determined for each of the cases. Two
prediction algorithms were used for the prediction of the parameters using these
sets. The Random Forest models had an out-of-sample explained variances of
more than 99.5% in each of the cases. The Bayesian models produced the
probability distributions for the prediction of the parameters in each of the
cases. The Bayesian Linear Regression implementation is presented for the case
of multiple targets.