
Probabilistic
Neural Networks
207
There are some general guidelines that should be followed if we
expect to find values of σ that result in good performance. First, it
almost goes without saying that the individual components of a
multivariate sample vector should be comparably scaled
to
some "unit"
variation. This equalizes the contributions of each variable. It also
eliminates interaction between choice of
σ
and scaling due to measure-
ment, which has no effect on classification ability, but can help the
experimental design by keeping σ in reasonable ranges. For example,
if the measured variables are the height and weight of persons, it
woul ...