I am writing a little test app that tries
to connect Neural Nets and a GA (that tries to optimize the net).

So the Neural Nets are actually not trained by a learning procedure
(like Backpropagation)
But evolved using a GA. (the evolution searches in the field of possible combinations
for the best sollution, and
the nets only act to their evolved properties without some magical
"backpropagation god")

This might not be the fastes choice, but the GA can go further and manipulate
the actual properties of the net too (connections, amount of neurons)