On the sample complexity of pac-learning using random and chosen examples