[stefs]

with the training/test data sets, wouldn't it be possible to find the best parameters with a genetic algorithm? i mean, sure, it'd take really long ... well, probably too long.