Actually, if you are following the method described in the article, if the best candidate is in the learning set, you end up with the last one. So it's effecti

Actually, if you are following the method described in the article, if the best candidate is in the learning set, you end up with the last one. So it's effectively a 36.8% chance of picking a random one from the "not-best" population.