Representation of Actions in Cyc and KM

1

• Representation of Actions in Cyc and KM
• RKF PI MeetingThursday, October 18, 2001

Aarati Parmar FRG Stanford
Pierluigi Miraglia Cycorp
2
Representation of Actions in Cyc KM

• Cyc and KM (Component Library v1.0) are both
  logic-based ontologies with inheritance, and some
  non-monotonic reasoning.
• Compare on
• Basic Temporal Formalism
• Action Ontology
• Reasoning about Change

3
Basic Temporal Formalism States

• Basic unit of state is implemented as a
  microtheory/context in both
• States are nested, so that facts from all
  super-states are visible
• One unique super-situation (BaseKB/Global)
  housingtimeless facts, visible to all
  situations.
• Cyc's holdsIn and ist-Asserted corresponds to
  KM's holds-in, in-situation

F
4
Basic Temporal Formalism States

• KM uses situations of sitcalc (state space)
• Cyc has two temporal formalisms
• 1. Davidsonian framework
• Action sentences are implicit existential
  assertions
• Instances of Events (subclass of
  Situation-Temporal) have spatio-temporal extent
• Slots and temporal relations (ActorSlots
  startsAfterEndingOf) relate properties of actions
• 2. In development
• Assertions modified by temporal and modal
  operators
• (possible-Historical (eats Fritz Caviar))

5
Basic Temporal Formalism Actions

• In both, actions are defined as events with a
  protagonist.
• KM actions connect situations.
• Has next-situation result of sitcalc (can
  represent possible futures
• Actions can be composed of subactions, etc.
• Future support for situation during the action.
• Cyc actions more process-like, (instances of
  Event have temporal extent.)

6
Action Ontology

• Action properties inherited in both Cyc and KM
  through hierarchy
• KM
• uses slots and values for arbitrary properties
• more powerful than most frame-based languages as
  values can be evaluable expressions containing
  quantification and implication
• precondition list for Move
• (if (has-value (the source of Self))
• then (forall (the object of Self))
• (triple It location (the source of Self)))

7
Action Ontology

• Cyc
• properties formalized through Roles, ActorSlots
  other temporal relations
• employs "skolem functions" to relate objects to
  actions, e.g. (relationAllExists buyer Buying
  IntelligentAgent)
• how an action is done formalized
  throughperformanceLevel, rateofEvent
• also categorizes different temporal objects
  (AccomplishmentType (actions that have a
  completion point), etc.)

8
Reasoning about change Preconditions

• To do progression (regression), preconditions, as
  well as the result of an action need to be
  formalized.
• KM uses STRIPS prec, add, delete lists to
  compute effects of actions.
• Cyc has an expressively rich set of
  preconditions, but they are not uniformly used
  (what predicate do we query to see if action a
  executable?).

9
Reasoning about change Preconditions

• Cyc preconditions represented through a multitude
  of predicates
• some ActorSlots are specific preconditions
  (inputs)
• preconditionFor-PropSit, Events, Props,
  SitProp
• (preSituation Event1 StaticSit2) a very weak kind
  of precondition
• necessary conditions necConditionFor-Event and
  necConditionFor-Scene

10
Reasoning about change Results

• KM
• STRIPS lists compute direct effect of actions
• a simple form of non-monotonic reasoning used to
  compute the inertial effects
• extra support for ramifications (non-inertial
  effects)

11
Reasoning about change Results

• Cyc
• (postSituation Event1 StaticSituation2)
  closest thing to result
• causation between other/more general
  classeseventOutcomes, causes-EventEvent,
  causes-SitProp, causes-ThingProp,causes-PropProp
  
• looser notion of salience postEvents and
  inReactionTo, and functions STIB, STIF.
• Once again, a plethora of different levels of
  result used in Cyc, but no one used generally.

12
Conclusions

• Cyc has a rich ontology, but current formalism
  does not go the route of talking about the set of
  facts which change, like KM.
• KM is better qualified to infer the results of
  actions, for this reason, as well as the
  non-monotonicity built into the system.
• While Cyc can teach us much about actions and
  properties of them, KM can actually simulate
  these actions.

13
Bibliography

• Clark, P. and Porter, B. (1998). KM (v1.3) Users
  Manual. Knowledge-Based Systems Group, Univ. of
  Texas at Austin, Austin, Texas.
• Cyc. http//www.cyc.com.
• McCarthy, J. and Hayes, P. J. (1969). Some
  Philosophical Problems from the Standpoint of
  Artificial Intelligence. In Meltzer, B. and
  Michie, D., editors, Machine Intelligence
  4,pages 463--502. Edinburgh University Press.