Planning Initiative SPAR - Sentences

This is a more detailed discussion of the SPAR issue 32, "Modification to KRSL-Plans 20-Sep-96 Definition in 15 Sentences".

Current working version

• General Comments
• SPAR Core Sentences
• SPAR Extension 1 - Additional Terminology
• SPAR Extension 2 - World State
• SPAR Extension 3 - Timelines
• SPAR Extension 4 - Values, Uncertainty and Imprecision
• SPAR Extension 5 - Conditions and Effects
• SPAR Extension 6 - Evaluations
• SPAR Extension 7 - Libraries

AT stated: Good idea. As in PIF, NIST PSL, etc., the primitives are meant to be
• entity
  • activity
  • activity-relatable-object
  • time-point
  • relationship

AP would prefer: C.1. A PLAN relates ACTIVITY(s) to OBJECTIVE(s)
KM stated: Should this include a reference to a WORLD-STATE? Something along the lines of: A PLAN relates an ACTIVITY-SPECIFICATION to one or more OBJECTIVES for a designated WORLD-STATE.
AT stated: ... objectives may relate to more than the world state. An objective could be to have completed an activity, or to have used some resource, or to have respected some (say spatial) constraint over some time period. WORLD STATE based objectives are common in our AI world but not so common beyond. In SPAR, world state descriptions are just one of the many possible ways. I also pointed out that an objective may define some function over sets of world states or times that is to be satisfied. e.g. the specification of completing a time trial course where milestone times are part of the spec. So we should not refer to any designated world state in the main definition of plan or objective.

C.2 An ACTIVITY-SPECIFICATION denotes or describes zero or more ACTIVITIES.

AT stated: "I agree with AP that ACTIVITY, SUB-ACTIVITY, PRIMITIVE ACTIVITY, PROCESS and SUB-PROCESS (and even EVENT) can all be the same entity. Each of these would be ACTIVITY in SPAR. ACTIVITY has a specification associated with it which may or may not include more detailed (sub-activities)."
SP: Should we consider keeping SUB-ACTIVITY as additional terminology via this sentence, "SUB-ACTIVITIES may be specified as required or prohibited by a set of ACTIVITY-CONSTRAINTS in an ACTIVITY-SPECIFICATION.", or is it unnecessary?
KM stated: I'm not convinced that ACTIVITY-SPECIFICATION is really necessary for the kinds of abstract characterizations that they are intended to provide.
AT stated: We have ended up using the term "activity" in 2 senses in SPAR. Some of the earlier KRSL-PLans sentences I felt were clearer in this respect, but others at that time wanted to be more concrete rather than referring to the "space of behaviour". I.e. in SPAR we use activity for both "an activity" and for the more general notion of "activity". Can we reword to make this clearer? An activity specification describes some sub-set of the whole space of legitimate behaviour in the domain.
SP stated: With an Activity-spec. we can represent various activity-constraint configurations independent of a plan (reason about various configurations) This also separates constraints related to the activity specification from constraints for a specific plan.
AP stated: I agree with the need for an activity-spec.

C.3 An ACTIVITY-SPECIFICATION may include relationships between those ACTIVITIES and ACTIVITY-RELATABLE-OBJECTS (AROs).

AP stated: I would prefer to delete this sentence.
AT stated: Just to be fully accurate the "and" in this sentence should be "and/or", as there can be required relationships just between activities, or relationships just between AROs (say equality of some objects referred to by variables), as well as the more usual relationships that mention both activities and AROs.
KM stated: ACTIVITY-RELATABLE-OBJECT is never defined -- frankly, I don't see the value in introducing it. I would suggest replacing it with the more general concept ENTITY.

C.4. EXECUTION of an ACTIVITY may change the WORLD-STATE.

KM stated: CHANGE TO: EXECUTION of an ACTIVITY can change the WORLD.

C.5. An AGENT is an ACTIVITY-RELATABLE-OBJECT which can PERFORM ACTIVITIES and/or HOLD OBJECTIVES.

AP stated: I would prefer to have everything in SPAR other than CONSTRAINT and ANNOTATION be an ACTIVITY-RELATABLE-OBJECT

C.6. An ACTIVITY takes place over a TIME-INTERVAL identified by its two ends, the BEGIN-TIME-POINT and the END-TIME-POINT.

KM: Suggests the addition of the sentence: "The TIME-INTERVAL, BEGIN-TIME-POINT and END-TIME-POINT need not be specified however."
AT: This clarification is inaccurate as the relationship of actions to timepoints and their relationship to one another IS a specification.
AP stated: KM's comment sounds right to me but maybe is unnecessary. We might state that none of the information specified in SPAR is required in an instantiated plan. We could have a plan with just activites, or just objectives, or none of the activities could have time points etc.
AT stated: I agree we could have some overall statement that things may be unspecified. But I disagree with all three parts which give details [to AP's] statement above.
1. An activity does have two time points (conceptually) as its begin and end and these can always be referred to (as activity.begin-time-point and activity.end-time-point in an object model). But these need not be actually allocated as "real objects" in an object model implementation unless they need to be for some other reason.
2. A plan may have a specification of activity that does not include any specific activity at all (zero sub-activities). One legitimate instance of a plan is literally to do no activity. Constraint may be stated to avoid some things, etc but that is an orthogonal issue.
3. A plan will always contain at least one objective. This is a defining characteristic of a plan. A specification of activity without an associated objective is not a plan. Its just a specification of activity.
AT: Suggests the addition of the sentence: "The BEGIN-TIME-POINT is temporally before the END-TIME-POINT".

E1.1. An ACTION is an ACTIVITY which has or could have (in the domain model) an ACTOR.

AT: Alter E1.1 to be ACTION is a SYNONYM for ACTIVITY, as there is no dispute over that term.

E1.2. An EVENT is an ACTIVITY which does not have and could not have (in the domain model) an ACTOR.

AT: Delete E1.2.

E1.3. An ACTOR is an AGENT which is the motive force behind an ACTIVITY.

AT: Could add that we call such an ACTOR a PERFORMER.

AP stated: I don't think we need both.
KM stated: What does `motive force' in E1.3 mean? Is the intent here to express a concept that is more general than `performance' of the activity?
AT stated: No, as you thought, its meant to just meant to refer to an actor that is in the "perform" relationship to an activity. This relationship may or may not be specified in the model, and the performers of some activities may not be expressed in the model at all - e.g. in the case of some "natural" activities.
SP: It sounds like it would be simplest to to replace motive force with the notion of performer.

E1.4. A RESOURCE is an ACTIVITY-RELATABLE-OBJECT which is USED, MODIFIED, CONSUMED or DESTROYED during the EXECUTION of an ACTIVITY.

AP stated: I'd like to distinguish between concepts which cover "all information in the plan" and "physical things". I'd like RESOURCE and ACTOR to be roles which physical things play. If we stick with the sentences as given, then we're allowing an ACTIVITY (which is an ACTIVITY-RELATABLE-OBJECT) to play the role of RESOURCE and that doesn't make sense to me.

AT stated: Can I suggest the addition of a sentence after E1.4, meaning that the ones that follow need to be renumbered.

E1.5 A PRODUCT is an ACTIVITY-RELATABLE-OBJECT which is CREATED during the EXECUTION of an ACTIVITY.

E1.5 A SUB-ACTIVITY of an ACTIVITY is an ACTIVITY included in an ACTIVITY-SPECIFICATION of the parent ACTIVITY.

E1.6 A PROCESS is an ACTIVITY whose ACTIVITY-SPECIFICATION includes more than one SUB-ACTIVITY.

E1.7 A PRIMITIVE-ACTIVITY is an ACTIVITY whose ACTIVITY-SPECIFICATION includes no SUB-ACTIVITY to the level detailed in the model.

KM stated: I don't see how an ACTIVITY-SPECIFICATION can possibly determine whether an ACTIVITY is primitive or not. The notion of `primitivity' seems like it's in inherent property of a given ACTIVITY type/instance, which will of course vary with different domain models.
AT stated: Thats why we relate "primitivity" to whether the model includes a specification of activity that further decomposes the activity. It it has sub-activities, then it is not primitive (wrt the given model).
SP stated: Note that if primitivity is simply determined by the number of sub-activities, then all unexpanded activities in a plan would be considered primitive by definition. Intuitively, we should be able to have abstract activities (make-part, fly-to-paris) that are non-primitive irrespective of their expansion state.

AT stated: I like this description. It fits with my own intuition on what we are trying to capture. Do others agree? I also prefer the term environment and its the term we have used in our own contributions to plan ontology work.

KM Stated: Here is an attempt to compile a set of sentences that capture the relevant details (perhaps (1) and (2) should simply be `background material' rather than sentences, since they really contribute to defining the background framework for SPAR).

1. A PLAN is designed for and (possibly) executed within a specified World (Environment).
2. A particular snaphsot of the world at a given point in time is called a World-State.
3. A WORLD-MODEL provides a description of the World (possibly incomplete and/or inaccurate).
4. A WORLD-STATE-DECRIPTION describes one or more World-States (actual, expected, or hypothetical).

E2.1. A WORLD-STATE-DESCRIPTION partially denotes or describes the WORLD-STATE.

KM: Suggests adding "at some actual or hypothetical point in time".

AT: Too limiting. May describe a difference between WORLD-STATEs.
AT: Whole sets of world states could be described by some constructs, and globally (always) true statements could be made.
SP: What is the difference between a specification and a description? Are they roughly the same thing just used for different purposes? Perhaps, to be uniform, we could call this a WORLD-STATE-SPECIFICATION...
AT: I suggest that HOLDS [from E4.3] be added as the second sentence in the world state extension (E2) so that we have a "hook" there on which to hang the uncertain values.

E3.2. A TIMELINE [or CALENDAR?] has a nominated begin point and a time unit and may have a nominated end point.

AT stated: I recommend that we adopt the more generic term "timeline" for this extsnsion. A calendar has a narrower scope with some idea of DATES on it. We have used it in a broader way where we just meant any time line with associated "tokens" such as "Day 32 of the project", but that's not very natural in English.

AT: Alternatively we could say: Each property of each entity has a VALUE, i.e. PROPERTY(ENTITY)=VALUE
AP stated: I like AT's rewording.
DLD: In order to set up uncertainties (or imprecision, for that matter), what we really want to express here is that a property has a "domain", or set of possible values.  The domain might not be (completely) known or specified, it might change over time, etc., but at some conceptual level at least, there is a range of possible values that each property could take on.

E4.2. VALUEs may be imprecise.

AP: Suggests that: "An imprecise VALUE may have an associated FUZZY-SET."
AT: This is an overcommitment for SPAR. This should be left to specific plug-ins that can coexist with other approaches. I am pretty sure that there will be no agreement to this level - and it could be undesirable to try to standardise things that may be better left to alternatives that can interwork.
AP: I guess the question is how to specify imprecision. I still run into occasional confusion about the difference between imprecision (fuzziness) and uncertainty so I'd like to make it clear that this sentence is addressing the former. Are there other ways to relate values to imprecise labels other than fuzzy logic?
DLD: I agree that it would be helpful to distinguish imprecision from uncertainty (and I think that most researchers in either area would agree).  Treatments of this are, of course, a religious issue, which raises certain difficulties.  It is my belief that one can completely and soundly represent imprecision by changing the domain of the attribute to be either a set of values, or more generally, a mapping from values to fuzzy degrees.

So: one way to approach the issue would be minimalist:  an attribute can take a value from a domain, and it is up to the user to specify if that domain consists of "fuzzy values" (subsets or fuzzy measures) or not.  The advantage of this approach is that it does not require you to add anything to your specification (not even statement E4.2).  The disadvantage is that it does not provide any support to those who would wish to make use of fuzziness.   The other approach, of course, is to attempt to explicitly support fuzziness, in which case I like AP's suggested statement better (though to be general, it should specify it in terms of the mapping to fuzzy degrees).

E4.3. VALUEs may have a PROBABILITY.

AT: Suggests it should read: "VALUES may have alternatives with differing PROBABILITIES."
AP: Either is fine by me.
DLD:  As per above, I agree with AT, though note that it is not technically the value which has a probability, but the association, attribute=value, which has a probability.
AT: I agree, we will rephrase suitably.
However, there are other entities which have probabilities, most particularly, logical sentences over attribute assignments (i.e. the sentence "vehicle X is black and it is moving quickly" may have a probability).  It is also useful to directly support probabilities for subsets ("the speed of vehicle X is between 50 and 62 mph"), particularly when it may not be possible to enumerate the subset.
AT: I saw such statements (we would call them world state descriptions in SPAR now) as being of form holds(proposition)=value
holds("vehicle X is black and it is moving quickly")=true
(with some probability)
Note the string would of course most likely not be in such an unstuctured form and would be represented in whatever plug in ontology/syntax was adopted for "world state description".
Additionally, we absolutely have to have the ability to represent conditional probabilities, which is say, we must be able to attach probabilities to statements of the form "if X is true, then the probability of Y is p", or "P(Y|X) = p".  In this case, X and Y can each be logical sentences over attribute assignments, as above.

Basically, I guess what I am implying is that there needs to be a "database" of sentences and attached probabilities.

AT: Yes that sounds plausible. ... We could define HOLDS(world state description)=truth-value in [another] Extension for world state and take it on from there ... The idea of the HOLDS statements being a hook for the probability sentence data base that DLD suggests is something we can explore. Its pretty much the way some planners already work with such "fact data bases"...

E4.4. PROBABILITYs may be subclassed into PROBABILITY-PREDICTION and PROBABILITY-SENSED.

AP: I think this distinction is pretty fundamental if we're to address both planning and execution monitoring. If we lump them together we'll be missing an important bit of information for decision support. Uncertainty due to sensing is very different I think than uncertainty due to prediction. It's not that the measures of uncertainty themselves would have a different form, just that they would be labelled differently as to their source.
DLD: I don't really agree --- see E4.6.

E4.5. A PROBABILITY-PREDICTION is the likelihood of a WORLD-STATE-DESCRIPTION being valid in the future.

E4.6. A PROBABILITY-SENSED is a likelihood that a WORLD-STATE-DESCRIPTION did in fact have the specified value in the past or at the current time.

DLD: I'm not sure exactly what the right distinction to make is here.  Note that you can still do "predictive modeling" about past events---you can predict how a game, recorded on television, will come out (and in some places, you can appearantly even bet on it).  Also, if you are doing Bayesian reasoning, you combine both predictive and sensory models to come to a common probability assessment.

It certainly is important to make clear if an attribute is (possibly implicitly) time-varying, and it is also important to make clear if a time-varying attribute's value (or probability distribution) is changed, does that change reflect a change to the attribute value, or a change to our knowledge about the attribute value.

It is will also generally be important to understand where probabilities come from (i.e. from what models, what inference), and I think that this need subsumes the predicted/sensed distinction.

E4.7. An INFLUENCE-NETWORK is a structure which relates PROBABILITYs and specifies their dependency structure.

AP stated: I thought this would actually be one level up from making the sentence too specific like requiring Bayes nets. I guess any determination of what is too specific for our model is a judgement call.
AT stated : I am not happy about including this one. I feel it is an over commitment for SPAR and thus should not be left in. This should be left to be something that is a plug-in to the more generic 6 sentences that come before it. If we can get agreement on the 6 sentences we will have a suitabvle framework I trust.
AP stated: I'd be happy if we then put this under [another] extension.
DLD: I do not like this statement either.  It is either too specific, or too general.
AT: We [can] create an "E4a: Values, Uncertainty and Imprecision - Details" extension and move E4.7 across into that as E4a.1 - as it does not directly refine any others sentence in E4.

AP: I think all constraints on actions are preconditions.
AP suggests:
E5.1.a The CONDITIONS of an ACTIVITY are called ACTIVITY-CONSTRAINTs.
E5.1.b ACTIVITY-CONSTRAINTs are CONSTRAINTs on ACTIVITES.
E5.1.c CONSTRAINTs restrict the VALUEs of PROPERTIES (unary constraints).
CONSTRAINTS may also restrict the VALUEs of several PROPERTIES relative to
one another (binary or higher arity constraints).
KM stated: I'm not sure why CONDITIONS need to be introduced (they are never defined anywhere). How about rewording as follows: ACTIVITY-SPECIFICATIONS include ACTIVITY-CONSTRAINTS that impose restrictions on its constituent ACTIVITIES and ENTITIES.

E5.2. The EFFECTS of an ACTIVITY are the changes made to the WORLD-STATE by EXECUTION of the ACTIVITY.

KM stated: CHANGE TO: The EFFECTS of an ACTIVITY describe expected changes to the WORLD that would be occasioned by EXECUTION of the ACTIVITY.
AT stated: I guess we have to be clear if effects mean "effects in the real world" or a "specification of modelled effects". KM is using effects to mean the latter.
KM stated: Actually, my intent was to cover both -- this comes for free if one adopts the model of WORLD including the internal WORLD-MODEL.
AT stated: So maybe we have to talk about EFFECTS (in the real world) and EFFECTS-MODEL (in the world model)?
AP suggests: E5.5 The EFFECTS-RECORD of an ACTIVITY is the record of changes made to the WORLD-STATE by EXECUTION of the ACTIVITY as sensed by some ARO.

E5.3 CONDITIONS which are mutually exclusive may be associated with a PROBABILITY.

AT: Believes this may redundant as all VALUES have already been defined as having a PROBABILITY.
AP: I agree with AT. Should be deleted.
KM stated: I don't understand the need for E5.3 or E5.4 --- it should be possible to associate probabilities (more generaly, uncertainty information) with all constraints/sentences within SPAR.

E5.4. WORLD-STATE-DESCRIPTIONS which describe changes to the WORLD-STATE may have a PROBABILITY.

AT: Believes this may redundant as all VALUES have already been defined as having a PROBABILITY.
AP: Maybe should be deleted.
SP: This has been changed from "EFFECTS may have a PROBABILITY." because it didn't seem ontologically consistent with the notion of EFFECTS described above.

E6.2. An EVALUATION-CRITERION may be applied to WORLD-STATE to create an EVALUATION.

E6.3. An EVALUATION may be a predicate (holds/does not hold) or a partial order on the results of [one or more] EVALUATION-CRITERIA.

AT stated: "The 'or' in this needs to be related correctly. We mean that you can say this criteria holds or does not (yes/no) or that there is a preference ranking (better or worse) on the results of applying the criteria to a range of alternatives or outcomes. We dropped the phrase including the word "ranking" in favour of partial order, but I thinkwe may have moved away from what we are trying to communicate. We need BS to adjudicate on this. It was AP and SP who felt that partial order was a better description of what was neing defined I think."
AP: I think just having a partial order would be find. A degenerate case of a partial order would be 1 and 0 or holds/does not hold.
BS stated: With the "or" there are really two types of evaluations: those that are categorical---go/no-go and those that are satisfied to some degree. Perhaps we should distinguish these as fundamentally different things. In that case, an evaluation criterion would consist of two parts: categorical criteria and preference(?) criteria. It might make sense to make this split even at this high level in the spec, since I think the processing of the two types of criteria will be quite different.
AT: Lets add an "6a: Evaluation - Details" extension and put in two sentences compatible with the current E6.3. These would be say E6.3.1 and E6.3.2. This would address BS's suggestions.

ARPI Home Page | SPAR Home Page

Page maintained by Steve Polyak (Steve_Polyak@ed.ac.uk), Last updated: Fri Apr 28 13:29:23 2000