Compositional Instance-Based Acquisition of Preference Predicates

Learning

Knowledge to guide search can be represented as a preference predicate PQ(x, y) expressing that state x is preferable to state y. Interactions by a learning apprentice with a human expert provide an opportunity to acquire exemplars consisting of pairs that satisfy PQ. CIBL (compositional instancebased learning) is a strategy for learning preference predicates that permits multiple exemplars to be composed, directly exploiting the transitivity of PQ. An empirical evaluation with artificial data showed that CIBL is consistently more accurate than an instance-based learning strategy unable to compose exemplars. Moreover, CIBL outperforms decision tree induction when the evaluation function Q underlying PQ contains one or more extrema or a severe discontinuity.