In this paper, we describe the problems inherent in designing algorithms
to reliably detect rare signatures.
The choices and modifications to be made - of fuzzy membership functions,
aggregation operator, skill score and the classifier - are described.
A simple genetic algorithm with minor modifications
is used for achieving intermediate goals, with different fitness functions
at each stage. The resulting
detection approach performs very well even in a distribution of
disproportionate classes.
We demonstrate the results on a severe updraft
detection scheme.
