In Proceedings of the First International Conference on Artificial Intelligence Planning Systems, pages 243-251, 1992.
The use of abstraction in problem solving is an effective approach to reducing search, but finding good abstractions is a difficult problem. The first algorithm that completely automates the generation of abstraction hierarchies is Knoblock's ALPINE, but this algorithm is only able to automatically abstract the preconditions of operators. In this paper we present an algorithm that automatically abstracts not only the preconditions but also the effects of operators, and produces finer-grained abstraction hierarchies than ALPINE. The same algorithm also formalizes and selects the primary effects of operators, which is thus useful even for planning without abstraction. We present a theorem that describes the necessary and sufficient conditions for a planner to be complete, when guided by primary effects.