Teaching Classes with High Teaching Dimension Using Few Examples

摘要

We consider the Boolean concept classes of 2-term DNF and 1-decision lists which both have a teaching dimension exponential in the number n of variables. It is shown that both classes have an average teaching dimension linear in n. We also consider learners that always choose a simplest consistent hypothesis instead of an arbitrary consistent one. Both classes can be taught to these learners by efficient teaching algorithms using only a linear number of examples.