| Summary: | How an agent classifies the world reflects its goals; an Eskimo has many words for "snow" because sometimes the distinctions are important. Some recent work in conceptual clustering has involved the use of the classifier's domain-dependent goals to determine relevant attributes of the objects being classified. We continue in this direction by describing techniques from planning research that more precisely, more powerfully, and more uniformly represent goals in an integrated planning and learning system. In such a system preconditions of action descriptions are used to build lambda-expressions that characterize classifications. The scope is then briefly enlarged to discuss more general interactions between planning and learning work.
|
|---|