Vasily A. Belokurov
Keble College
University of Oxford

Trinity Term 2003
Thesis submitted for the degree of Doctor of Philosophy at the
University of Oxford


To cope with the avalanche of data from future variability surveys
it is essential to make object classification automatic. Conventional
methods utilized by earlier optical transient searches such as MACHO
lack flexibility and require vast computer time. I show how simple
supervised learning algorithms can enhance greatly the lightcurve
classification routine. I illustrate the advantages of the method
with a number of examples and compare my results with those obtained
using classical algorithms.

In the first two chapters of the thesis, I present a new method for
lightcurve classification based on neural networks. I envisage that
neural classifiers offer a superior way of detecting rare objects such
as lightcurves of microlensed stars in big datasets (with a number of
entries greater than a few hundreds of thousands). The power of the
method is in the distributed processing of all available
information. It is designed to reproduce complicated decision
boundaries in many dimensions and so naturally the technique
out-performs conventional selection procedures that approximate a
decision boundary with a series of straight line cuts. Moreover, a
trained network needs only a fraction of a second to yield the class
membership for a large number of data patterns.

One example of such a survey is the Gaia space mission.  Gaia will
scan the entire sky, with the goal to compile the largest ever
catalogue of stars in our Galaxy. Although directions at which Gaia
will point its telescopes are pre-programmed, during the mission the
light from numerous transient objects will be recorded and send to the
ground. The concluding chapters of the thesis are concerned with the
data-stream simulation of the Gaia satellite. The main contribution of
this work is a detailed study of the microlensing and supernova signal
available for Gaia. I generate both photometric and astrometric
measurements. From the analysis of the dataset, I predict the
detection rates and outline possible follow-up strategies.