An Interactive Method for Inducing Operator Descriptions

Specifying operator descriptions for planning domain models, especially using standard preand post condition symbolism, is a slow and painstaking process. This is because one is trying to capture what is essentially procedural knowledge in a declarative way in a language whose design is influenced by the construction of planning engines. The problem is acute if nonplanning experts are undertaking this task, and/or the operators are complex or hierarchical. In this paper we describe opmaker, a method in which the domain expert specifies the declarative structure of the domain (in terms of an object hierarchy, object descriptions etc) and provides training operator sequences. This input is made in the context of a tools environment supporting planner domain acquisition and modelling. opmaker then induces a set of parameterised operator descriptions from these examples, removing the need for the user to become involved in complex parameter manipulation within the underlying symbolic, logicbased language. We discuss the empirical evaluation of the implemented induction algorithm with the help of a range of domains, and draw conclusions for future work.