PLANET's alignment with the IKB Jim Blythe, ISI, Cleo Condoravdi, Teknowledge and Ben Rode, Cycorp Updated February 2000 This document summarizes the alignment of ISI's PLANET ontology with Cyc and the IKB. The alignment process, which took place under the HPKB project, resulted in the addition of several microtheories the the IKB. In these, we have identified some of the entities in PLANET with existing entities in the IKB, and added entities to Cyc that correspond to many of the others. Some of these were straightforward to model in Cyc while others presented technical challenges whose solutions we describe below. There remain some entities that are not yet mapped in Cyc but are still under consideration by Cycorp. The ontology was used in Cycorp's representation of plans for the HPKB year 2 challenge problem as well as in ISI's plan critiquer for the challenge problem. Of the 48 constants that capture aspects of PLANET in Cyc, 24 have been added to the IKB. In addition, 8 of the relations that are specific to COAs in the IKB are defined in terms of PLANET relations. The remaining PLANET constants were defined in a Meld file that can be loaded into the IKB, but are not included in the latest release of the core IKB. Before going into the details of the alignment strategy we describe three interesting features. First, several entities that are modelled as concepts in Loom are more naturally represented using microtheories in Cyc. Although this may make the surface representation look different, it captures the same information in each system in the most natural way, and automatic mappings are possible between the two representations without loss of information. Second, when entities in PLANET were added to Cyc the name was often changed in order to conform with Cyc's naming conventions. For example, PLANET's "problem-refinement" is called "planSelectionRefinement". Third, sometimes entities that appear to need to be directly aligned because they are denoted with similar symbols turn out to be very different in nature. A lot of careful analysis was done in this alignment in order to address this. An example is the relation "before". In PLANET there is a relation type that allows temporal relations between tasks to be expressed as commitments, for example that taskA should finish before taskB in some plan. Cyc already had temporal relations such as "before", but they refer to an actual occurrence of plan actions and do not have the (deontic) aspect of intention that PLANET captures. A good solution for Cyc is to provide a predicate that takes such a relation, eg taskA finishes before taskB and expresses the deontic version. This is the route we have taken. Now on to the details. First, these are the one-to-one mappings and subclasses that are made to existing Cyc concepts or relations. Where the concept is equivalent it has been marked with "(equiv)", otherwise the PLANET entity is marked as a sub-entity of the Cyc one. The Cyc constants are found in the microtheory PLANETAlignmentMt in the IKB, linked to the PLANET names with one of the two predicates synonymousExternalConcept (those marked "equiv") or overlappingExternalConcept. PLANET Loom concept or relation Cyc class or relation plan Plan (equiv) plan-task PurposefulAction These subconcepts will also be mapped by extension: primitive-plan-task, plan-step goal-specification Goal (equiv) (was state-based-goal-spec, but Cleo pointed out that although #$Goal is explicitly state-based, we can use propositions to capture objective-based-goal-specifications. This is done in the Cyc version of PLANET.) These subconcepts will also be mapped by extension: state-based-goal-spec, objective-based-goal-spec sub-task (relation) subEvents (equiv) There are more detailed relations in Cyc that may be more appropriate, but they are not as yet released in the IKB. preconditions preconditionFor-PropSit (equiv) plan-task-template scriptType (equiv) plan-restriction planContextRestrictionPredicate (equiv) temporal-restriction timeConstraintOfAction required-resource resourceInPlan, resourceInEvent accomplishes intentInPlan plan-goal purposeOfPlan planning-problem-context PlanBackgroundContext (equiv) The remaining constants defining PLANET in Cyc are contained in a separate file, which is available from the PLANET web page, http://www.isi.edu/isd/HPKB/planet, or by contacting Jim Blythe. We now describe some of the issues that arose in aligning PLANET with Cyc, listing the remaining constants as we discuss the relevant topics. temporal-restriction, relation, mapped to timeConstraintOfAction in the IKB: Originally this was modelled as a ComplexTemporalRelation, but this was inappropriate because subrelations of this do not have deontic force. A new predicate was to be added to Cyc to express the deontic equivalent of any ComplexTemporalRelation appearing in its scope. For instance, to express that task A should start after the end of task B, the predicate would make use of the ComplexTemporalRelation StartsAfterEndingOf to say something like (plan-requires StartsAfterEndingOf taskA taskB) (this is not the actual predicate name, which Ben will choose.) This would avoid the problem of re-defining all the existing temporal predicates in deontic form. The distinction in PLANET between constraints (imposed by forces external to the planner, such as the commander) and commitments (made during the planning process to specify a plan) could probably be capture by the kind of microtheory in which the assertions are made. However for reasons of limited resources, the explicit predicate timeConstraintOfAction was used. planning-problem-context (mapped to PlanBackgroundContext in the IKB) and planning-problem: Some of the concepts used in PLANET to organise planning problems and groups of plans are modelled in Cyc as microtheories. Thus where in Loom a planning-problem-context may have a number of planning-problems defined from it, each with a number of plans, each in turn with refinement plans, Cyc has a hierarchy of microtheories to capture the same structure. During the challenge problem we were able to make an automatic translation between the Loom and Cyc versions despite this difference in structure. This approach is used to model the PLANET concepts planning-problem, planning-problem-context and plan. Some of the relations between these concepts are be modelled as relations between microtheories, including problem-refinement and refinement-plan. Others, such as sub-plan, are be captured in a different way. PLANET (Loom) PLANET (Cyc) planning-problem planSelectionContext planning-problem-context planSelectionBackgroundContext problem-refinement planSelectionRefinement As a result of our discussions some changes were also made to PLANET. The concepts feasible-plan, rejected-plan etc that were subconcepts of plan were replaced by subrelations of the candidate relations feasible-candidate, rejected-candidate and so on. The following relations, which capture the process of making and examining choices, are defined in the Meld file: candidate candidatePlanningContext candidate-plan candidatePlanInSelectionContext feasible-candidate feasiblePlanInSelectionContext unfeasible-candidate inFeasiblePlanInSelectionContext untried-candidate untriedPlanInSelectionContext selected-candidate selectPlanInSelectionContext rejected-candidate rejectPlanInSelectionContext refinesPlan Action consequences and effects: Cleo and Ben built a richer representation for the effects of actions, which was aligned with the following relations in PLANET: effects, primary-effects and accomplishes (named "intentInPlan" in the Cyc version). Jim also contributed to the representation for action effects. The following relations group together a set of assertions about the plan and its components that are either made externally (constraints) or internally (commitments). In Cyc, some of the restrictions themselves are implicitly made in microtheories. PLANET (Loom) PLANET (Cyc) plan-restrictions planContextRestrictionPredicate constraints planContextConstraintPredicate commitments planContextCommitmentPredicate plan-commitments commitmentOfPlan plan-constraints task-of eventOfPlan task-commitments commitmentOfPlanAction task-constraints constraintForTaskType (related) ordering-restriction orderingRestrictionInPlanContext ordering-commitment ordering-constraint The following concepts and relations do not describe resources in detail, but capture the way the planning ontology should link to an ontology of resources. At ISI, we use the resource ontology that is part of the plan critiquing problem-solving method, while at Cyc there is still some work to be done on a taxonomy of resources. These entities are modelled in Cyc: required-resource resourceInPlan, resourceInEvent resource-requirement resource planning-resource In addition the following Cyc constants were used which overlap with the ISI resources ontology: actionResourceQuantForTime, planResourceQuantForTime, agentResourceInEvent, agentResourceInPlan. The remaining entities complete the PLANET ontology. Where a Cyc name is specified, an equivalent constant was defined in the separate Meld file. Loom name Cyc name problem-goal planSelectionGoal plan-goal purposeOfPlan intermediate-goal plan-task-description primitive-plan-task-description planning-level-specifier planning-level The following group of concepts and relations correspond to the lexical annotations on constants used in Cyc to support natural language: human-readable-description human-readable-task-description human-readable-description-of human-readable-goal-specification-description human-readable-goal-description The following group link PLANET with the evaluation ontology. Yolanda Gil and Ben Rode did initial work on the alignment of this ontology with Cyc. evaluation-structure evaluation comparison-structure comparison The following Cyc constants used in the representation of HPKB COAs are defined in terms of the PLANET ontology: intentOfCommanderInCOA intentOfG2InCOA intentOfG3InCOA intentOfCommanderForAction depthConstraintOfTask traversalConstraintOfTask attritionConstraintOfAction specialtyConstraintOfTask