Learning Hierarchical Problem Networks for Knowledge-Based Planning

In this paper, we reconsider the representation and use of expertise about sequential goal-directed activities. We discuss previous research on this topic, identify its limitations, and present a new theoretical framework – hierarchical problem networks – that addresses them. The core idea is that procedural knowledge consists of conditional methods that decompose problems – sets of goals – into ordered subproblems. Another innovation is that methods incorporate tests that forbid their use when certain goals are unsatisfied. We state the theory’s key postulates about representation and processing, after which we describe HPD, a problem-solving architecture that makes these assumptions operational. Next we report empirical demonstrations of HPD’s behavior in three planning domains, including studies of the relative importance of different types of conditions on constraining search. In closing, we review the theory’s main ideas and their intellectual precursors, along with our plans for future research.