Nov 5, 2003 ------------- - Recap EBG - worked out Housemates problem - what is a good choice of operationality here? - notice that the training data is truly superfluous, since domain theory is enough to entail the rule! - we still need the training data of Sue and Joe, to "point us in the right direction" - Evaluation of EBG - empirical utility - one person's criterion of operationality might be different from another's - too many specific rules versus few general rules - should take into account the distribution of situations (problems) that will occur in future - Type B learning - general schema (new knowledge ^ descriptions) entails classifications (old knowledge ^ descriptions ^ classifications) entails new knowledge - example "In India, elevators are called lifts." - Instantiation of this form of learning - old knowledge: "in a country, people call some things by the same names" - description: "Harsha, an Indian, ..." - classification: "..., referred to an elevator as a lift." - new knowledge: "In India, people call elevators as lifts." - Need both old knowledge and {descriptions, classifications} - cannot infer new knowledge from old knowledge alone => does not logically follow - cannot infer new knowledge from {descriptions, classifications} alone => too much generalization - Basic idea: determinations or functional dependencies - nationality "determines" language - think "functional dependency" from databases - if two tuples (instances) agree on nationality, they agree on language - How to find determinations - search through space of all possibilities - e.g., "In India, people are called Harsha." - e.g., "In India, people call elevators as lifts." - Other forms of determinations - multivalued dependencies - e.g., "All Indians speak the same languages." - More examples of Type B learning - "In Ohio State, they have a quarter system." (University determines Term Periodicity - "Law and Order airs at 10pm on Thursdays in NBC" (TVShow determines time, and maybe even the reverse)