Extending Problem Speci cations for Plan-Space Planners

Problem speciications for classical planners based on a STRIPS-like representation typically consist of an initial situation and a partially deened goal state. Hierarchical planning approaches, e.g., Hierarchical Task Network (HTN) Planning, have not only richer representations for actions but also for the representation of planning problems. The latter are deened by giving an initial state and an initial task network in which the goals can be ordered with respect to each other. However, studies with a speciication of the domain of process planning for the plan-space planner CAPlan (an extension of SNLP) have shown that even without hierarchical domain representation typical properties called goal orderings can be identiied in this domain that allow more eecient and correct case retrieval strategies for the case-based planner CAPlan/CbC. Motivated by that, this report describes an extension of the classical problem spec-iications for plan-space planners like SNLP and descendants. These extended problem speciications allow to deene a partial order on the planning goals which can interpreted as an order in which the solution plan should achieve the goals. These goal ordering can theoretically and empirically be shown to improve planning performance not only for case-based but also for generative planning. As a second but diierent way we show how goal orderings can be used to address the control problem of partial order planners. These improvements can be best understood with a reenement of Barrett's and Weld's extended taxonomy of subgoal collections.