Did you know ... | Search Documentation: |

Pack planner_api -- docs/kbcs00.ps.txt |

Foundations of an Object-centred Approach to Knowledge

Engineering for AI Planning Applications

D. E. Kitchin T. L. McCluskey R. M. Simpson

The School of Computing and Mathematics,

The University of Huddersfield,

Huddersfield, UK

June 28, 2001

1

Abstract Recent successful applications of AI planning technology have highlighted the knowledge engineering of planning domain models as an important research area. In this paper we describe an object-centred language for encoding planning domain models, and demonstrate that rigour is maintained by defining a truth criterion for the language. This truth criterion provides a sound basis for the development of object-centred planning algorithms. It can also be used to aid and simplify threat detection and resolution.

1 Introduction Due to the inherent complexity of planning it is well known that there is no one planning algorithm that can perform well on all problems in all domains. In practice, many planners still concentrate on simple, toy domains. However, recently there has been an increasing use of AI planning technology for real applications [15, 1]. It is also beginning to be recognised that, in order to plan for real domains, planners need to be able to utilise all the knowledge contained in a domain model to help improve planner efficiency [6, 5]. We propose that adopting an approach inspired by object oriented methods for the construction of domain models and planning engines can greatly facilitate this objective.

Approaches to modelling in planning have concentrated in the last 30 years on the dynamic, behavioural view. This is nowhere more true than in domain independent planning, where the aim seems to be to have a planner input a minimal description of dynamic behaviour. While this approach has, to some extent, facilitated the construction of and research into planning algorithms, it appears to have narrowed the concerns of the field. Developments in domain modelling in the last few years seem obsessed with extending the expressivity of these dynamic languages, and standardising the syntax (exemplified in PDDL [3]).

Our own work, inspired by object-oriented analysis and design practice, has concentrated on taking an object-centred approach to the engineering of planning domain models [13, 10]. This is facilitated by the recognition that in the traditional proposition-centred representation, propositions are clustered around objects, and actions change the state of objects [11]. Rather than a drastic change to the traditional dynamic domain model, we have sought to evolve the prevailing ideas so that, for example, it is not difficult to translate the standard conventional syntax (PDDL) to our representation language (OCL) [16].

This evolution resulted in an exploratory object-centred method specific to planning which involves building up a structural view of the domain in addition to the dynamic view. Objects in a planning domain (e.g. robots, packages, trucks, and so on) can be thought of as having a "state", and can naturally be arranged in object class hierarchies. The behavioural model can be further elaborated with transition diagrams for each object class.

In this paper we describe our object-centred approach to the engineering of planning domains together with the theoretical underpinnings of object-centred planning. We define a truth criterion for object-centred planning which provides a sound basis for the development of object-centred planning algorithms. Additionally we discuss how we can use this truth criterion in threat detection and resolution. Our approach produces a more structured domain

2

model, allowing the capture of more knowledge which we can make use of both before planning (via compilation tools) and during planning (via appropriate planning engines) to help reduce search.

2 Object Domain Model First we will outline the domain model, using as an example a multi-robot world, R3. R3 consists of seven rooms, connected by several doors and containing a variety of boxes and keys. Boxes can be moved from one location to another by robots, who may have to lock or unlock doors and switch lights on or off in order to carry out their allotted tasks. R3 domain model has:

- a set of primitive sorts Sorts denoting object classes: e.g. froom; door ; box ; robot ; key; :::g
- a set of object identifiers: e.g. fbox 1; box 2; r 1; r 2; key6; key7; door 12; door 23; :::g
- a set of static and dynamic predicate signatures, each with arguments identified by sort
names: e.g.
fbox

`in(box ; room)`

; robot next box (robot ; box ); unlocked (door ); connect (room; room; door )gThe first argument is reserved for an instance of the key sort, which we call the key object that is, the object whose state the predicate is describing.

- a set of atomic invariants. These are instances of static predicates Predicates that are always true in the domain: e.g. connect (room1; room2; door 12). Any possible instantiation of a static predicate that is not a member of atomic invariants is assumed false.
- a set of Substate Class definitions. Each sort in Sorts must be specified by describing the
possible states of the object instances of that sort. Each such state description is characterised
by a substate class expression. For example, objects of the sort robot can be in one of the
following four substates:
ffrobot

`in(Robot ; Room)`

; robot near door (Robot ; Door ; Room); connect (Room; Room1; Door )g; frobot`in(Robot ; Room)`

; robot next box (Robot ; Box )g; frobot`in(Robot ; Room)`

; robot next`key(Robot ; Key)`

g; frobot`in(Robot ; Room)`

gg Example 2:1All variables (shown with initial capitals) range over their sorts. Any instantiation of a substate class expression in which any static predicates are members of atomic invariants constitutes a valid substate. For example, this substate satisfies the first substate class above, and so is a valid substate for r 1:

3

frobot in(r 1; room1); robot near door (r 1; door 12; room1); connect (room1; room2; door 12)g

- a set of inconsistency constraints: these are invariants which place restrictions between
substates. Examples of state invariants are that two robot arms (A1 and A2) cannot be
holding the same key (K ).
inconsistent constraint(arm used (A1; K )&arm used (A2; K )&A1 6= A2) The domain model is completed by descriptions of the domain operators and problems as specified by their goal conditions and initial state. These are described in sections 2.2 and 2.3.

2.1 Substate Class Expressions and Substate Expressions In order to aid the subsequent discussion of operators, goals and truth criteria, we will give here some definitions, making clear the distinction between substate expressions and substate class expressions and define the relations of containment and generalisation that hold between substate expressions.

Substate Class Expressions A substate class expression is a description of a possible substate of a sort which, when fully instantiated under a legal substitution, describes a unique substate for any object of the sort. The substate class expressions for the sort Robot are given in Example 2.1 on page 3.

Substate Expressions Substate expressions describe conditions on substates. A substate expression for an object must contain a subset of predicates in any one of the object's sort substate class expressions. For example, robot

`in(Robot ; Room1)`

is a general condition on the state of the robot - the robot's substate may be any substate which contains this predicate. Predicates cannot be combined indiscriminately - for example, robot near door (Robot ; Door ; Room), robot next box (Robot ; Box ) would not constitute a valid substate expression, as there is no valid substate class expression containing both these predicates. A substate class expression is written under a local closed world assumption - if a predicate is not present, then it is false. This is not true of substate expressions.Containment We can say that one substate expression, z, contains another, z'. We write z ' z 0 if the predicates in one expression are a subset of the predicates in the other, and all terms are identical. This means that any variables will also be identical. Both the substate expressions below contain the substate expression frobot

`in(Robot ; Room)`

g:frobot

`in(Robot ; Room)`

g frobot`in(Robot ; Room)`

; robot next`key(Robot ; Key; Room)`

g4

To illustrate this, and the following points, we will use an abstract example, as shown in Figure 1. In Example 1, where x and y may or may not be variables, we can say that R3 contains L2.

Example 1 Example 2.f(x',y')&g(x',a)

`f(x,y)`

`f(x,y)`

R L21 3R`f(x,y)`

&`g(x,a)`

Note, we will use x and x' to show there is some legal substitution where the terms will unify

Figure 1: Abstract Example 1 Generalisation If we have two substate expressions, x and y, concerning an object of the same sort, then we can say that one expression (y) is a generalisation of the other (x) if there is some legal substitution s, such that xs contains ys . For example, we can say the substate expression

`f(x,y)`

is a generalisation of f(x',y'),g(x',a) where it is possible to unify x and x', and y and y', as shown in Example 2 in Figure 1. Note, the fact that two variables can be unified does not mean that they necessarily will be during planning. More concretely, robot`in(Robot ; RoomX )`

is a generalisation of robot`in(Robot ; RoomY )`

, robot next box (Robot ; Box ; RoomY ) since RoomX and RoomY can unify.2.2 Goal Conditions and Initial State A goal to be achieved by the planning process is defined in terms of a set of valid substate expressions for a subset of the objects in the problem domain. The initial state of the planning problem is defined by a set of ground substate class expressions fully specifying the start state of every object in the problem domain.

2.3 Operators Object-centred operators have 3 components: prevail conditions, necessary changes and conditional changes. Prevail conditions specify those conditions that must be true before the operator can be applied, and which will still be true after its application. The necessary changes are shown by condition-effect pairs, (1 pair for each object concerned), showing the conditions that must be true before the action and the effects that must result after its application. The conditional changes are also made up of condition-effect pairs, but the specified effects only occur if the conditions are true. Note that there is a difference between variable slots in the necessary effects and variable slots in the conditional effects. Those in the necessary effects are existentially quantified, whilst those in the key slots of the conditional effects are universally quantified.

5

Name:

`pushtodoor(Tom,Box,Door,Room)`

Prevail conditions:ffrobot

`in(Tom,Room)`

,robot next`box(Tom,Box)`

g gNecessary changes:

f fbox

`in(Box,Room)`

g ) fbox`in(Box,Room)`

,box near`door(Box,Door,Room)`

,`connect(Room,R,Door)`

ggConditional changes:

ffrobot

`in(Dick,Room)`

,robot next`box(Dick,Box)`

,`ne(Dick,Tom)`

g ) frobot`in(Dick,Room)`

gfbox

`in(B,Room)`

,box next`box(B,Box)`

,`ne(B,Box)`

g ) fbox`in(B,Room)`

gfkey on

`floor(K,Room)`

,key next`box(K,Box)`

g ) fkey on`floor(K,Room)`

ggFigure 2: An example of an operator from R3 The right-hand sides of the necessary and conditional effects are substate class expressions, describing the resulting unique substates of the objects affected by application of the operator. The prevail conditions, the left-hand side of the necessary changes and the left-hand side of the conditional changes describe the conditions that must (or may, for conditional changes) be true for the action to be applicable. They are all substate expressions. Figure 2 is an operator with conditional effects from R3.

2.3.1 Establishment in Partial Plans A partial plan is a partially developed plan containing the partially instantiated and partially ordered steps in the plan so far. A partial-order plan represents a set of total-order plans, as there may be many different ways of linearising the partial ordering.

A substate expression l is possibly established by a substate class expression r in the righthand side of an action in a partial plan PP , (or in an initial state), if there is some legal substitution t where lt ` rt and the constraints and static predicates in the partial plan, PPt , are consistent using the atomic domain invariants. We work this as:

p

`estab(rt ; lt ; PPt)`

= lt ` rt ^ consistent (PPt :static) ^`before(rt ; lt )`

By consistent (PPt :static), we mean that all the static predicates in the current partial plan, under the current substitution t , are true with respect to the atomic domain invariants and do not contradict each other.In the following discussion we will use Nec:lhs, Cond:lhs, Nec:rhs and Cond:rhs to refer to the left-hand sides and right-hand sides respectively of the necessary and conditional transitions.

6

We will use A:lhs to refer to all the substate expressions in the left-hand sides of the prevail conditions, necessary and conditional changes of a step A, and A:rhs to refer to all the substate class expressions in the right-hand sides of the necessary and conditional changes.

3 Truth Criteria In this section, we look at a literal-based definition of a necessary truth criterion, before defining our own version for object-centred planning. This truth criterion provides a sound basis for the design of object-centred planning algorithms.

Classical literal-based planning uses the literal as the basic level of representation, and operators representing actions as the basic knowledge structure. We define a literal as an atomic proposition or its negation. In object-centred planning we lift the level of domain representation from the literal to focus on objects and their behaviour. A literal-based truth criterion describes what conditions are necessary for ensuring that a literal is true at some point in a partial plan. An object-centred truth criterion describes what conditions are necessary for ensuring that a substate expression is true at some point in a partial plan.

t s

C W

p

r~q Chapman's necessary truth criterion

a clobberer a white knight

KEY

necessarily before possibly before

Figure 3: Chapman's Necessary Truth Criterion 3.1 Literal-based Truth Criteria We give here Kambhampati's definition of a truth criterion: "A truth criterion specifies a set of constraints on the orderings, bindings and steps of a plan, which, when satisfied, will ensure that a given precondition of a given step is necessarily true (in that it will be satisfied in all the completions of the plan) ......When all the preconditions of all the steps are necessarily true, then the plan itself is necessarily correct". [7]

Chapman's Modal Truth Criteria (MTC) defines necessary and possible truth for partial

7

plans. Necessary truth describes the truth of some literal for every completion of the plan, while possible truth refers to its truth in at least one completion. To paraphrase Chapman's MTC, a proposition p is true in a situation s, if it has previously been necessarily asserted (made true by some action), and if there is any action which clobbers (makes false) this proposition, then there must be another action ordered between the clobbering action and the situation s which asserts p. 1

Figure 3, illustrating Chapman's necessary truth criterion, is reproduced from [2]. The dashed box is a clobbering step and the dotted box what Chapman terms a White Knight, that is, a step that would make the dashed step safe at this point. White Knights are used to deal with plans like the one in Figure 4, which illustrates a part of a plan where p and q possibly codesignate. If they do codesignate, then the second step denies p, but it is reasserted by step 3.

p q Step 1 Step 2 Step 3

~q

Figure 4: Chapman's `Odd Plan' 3.2 Truth Criterion for Object-Centred Planning We can see that the example above is concerned with literals, and so is not applicable to object-based planning. (Chapman uses the term proposition rather than literal). Therefore we have developed an equivalent truth criterion for object-based planning. We will see that the structure of the object-centred truth criterion is similar to that of the literal-based truth criterion described in the previous section. This truth criterion depends on the following definitions, of achieves and n estab.

Achieves Earlier we discussed what it means for an action to be a possible achiever of a goal for a step in a partial plan in the object-centred formulation. We now give a more formal definition, for a step, A, under some substitution t (of objects to variables), achieving a goal, l , in a step, O , in a partial plan, PP, as:

`achieves(A; l ; O; PP )`

, 9 t : n`estab(At ; Ot ; lt ; PPt )`

^ consistent (PPt :static) l is a substate expression, which may be a prevail condition, or one of the left-hand sides of a necessary or conditional change, or a `top-level' goal.1We note that there have been subsequent corrections with regard to Chapman's modal truth criterion for possible truth by [8] and others. [8] point out that it is correct for both necessary and possible truth only if we "reinterpret it as a criterion for modal conditional truth (that is, modal truth conditional on plan executability)"

8

Necessary Establishment A substate expression l is necessarily 2 established by a substate class expression r , where there are steps At and O in the partial plan PP which have, respectively, a right-hand side r , and a left-hand side l, if A ! O is a member of the ordering constraints in PP, and there is no step C , possibly ordered after A and before O which has a prevail condition or substate change for the same key object as r and l, unless C also has a right-hand side u which contains l.

A; O 2 PP:steps ^ l 2 (O :lhs [ O :prevail ) ) n

`estab(A; O ; l; PP )`

,9 r 2 A:rhs^ : l ` r ^ A ! O ^ :(9 C 2 PP :steps :

C ! O ^ A ! C ^ 9 u 2 C :rhs : (l :object = u:object ^ :(l ` u))

In the following section we go on to look at the definition of a threat for an object-centred planner, and what techniques can be used for threat resolution. We show how the use of the object-centred truth criterion can simplify threat resolution.

4 Threats and their resolution For a plan to be sound, all goals, all prevail conditions and all the left-hand sides of all necessary conditions need to be necessarily achieved. Additionally, the left-hand sides of any conditional changes need to be necessarily achieved (for the relevant objects only) if their right-hand sides are used to achieve a substate expression for some step in the plan.

In the terminology of planners with `causal links', a causal link is usually defined [9, 17] as a data structure consisting of three fields, (A; p; O ), where p is a proposition, A is an action which has p as an effect, and O an action which has p as a precondition. Links for an objectcentred planner contain an identifier for an object rather than a proposition. We define a link as a structure which contains the ids of two steps, S effect and S need , where S effect is the step which brings about the substate change for an object, o, required by the step, S need . The link also contains o id , which is an identifier for o with the form (object ; sort ), where object can be identified by a constant or a variable, and sort is the object's sort. We refer to that substate expression in S need which is the required substate expression for o as OE. For example, o id could be (box 1; box ), and OE might be box in(box 1; Room). For any

2We follow Kambhampati in not requiring the executability condition for the partial plan.

9

object, each action will have only one transition or prevail condition, so the relevant substate expression in S effect can be easily identified.

A threatening step, C , is usually defined as a step with an effect that deletes the proposition p in the link, and which can possibly be ordered between steps A and O [17]. In some planners, the threatening effect may be conditional. In the object-centred formulation, a step, S threat , potentially threatens a link if it has either a prevail condition, a necessary or conditional change for the same object o as that in o id and if it can be ordered between S effect and S need . We say potentially threatens, as it may be the case that OE is also established by S threat .

Existing threat resolution techniques include demotion, promotion, separation and confrontation. Demotion and promotion are commonly used in non-linear planners without conditional effects [9], while separation and confrontation are additional techniques used in a planner with conditional effects and universal quantification [12, 17]. Demotion requires an ordering constraint to be added to make C before A, while promotion adds a constraint to order C after O. Where the threatening effect is conditional, confrontation adds the negation of the conditional effect's antecedent as a new goal. Separation adds new binding constraints on existentially quantified variables.

Threat resolution techniques for object-centred planners also include demotion and promotion, used in the conventional ways. However, the use of an OCL and an object-centred planner means we can use a version of separation which we call object separation. Object separation, when the threat is from a necessary effect, simply requires the addition of a constraint that the object in the link and the object in the threatening effect are not the same. It is not necessary to add constraints on all the subsidiary objects as well, as it is the key object's state changes which are important. If the threat is from a conditional effect, then we can add constraints on the existentially quantified variables in the conditional transition and their corresponding variables in the threatened substate. Thus, all types of threat can be resolved by one or more of these techniques. Confrontation is not required in object-centred planning.

4.1 Threats from Necessary Effects

A, C , O = Steps in a partial planSorts S1, S2 and S3 are the same

A

C L --> R L -->

O

1 21 S1 S2

L --> R3 3

S3

Figure 5: Step C is a potential threat

10

Let us assume a situation, illustrated in Figure 5, where a step, A, in a partial plan has a member of A:rhs which we will call R1. Another step, O , has a member of O:lhs which we will call L2. Step A provides the required substate transition for L2. A third step, C , has a Nec:rhs, R3. A is ordered before O in the partial plan and C can potentially be ordered between A and O. The substate transitions for R1, L2 and R3 all concern objects of the same sort.

Now let us consider all the possibilities regarding possible threats from necessary effects and how they may be resolved. To illustrate this, we will use again an abstract example, shown in Figure 6, where we can see that in all these examples R1 contains L2.

RR 21 3 L

`f(x,y)`

&`g(x,a)`

`f(x,y)`

&`g(x,a)`

`f(x,y)`

&`h(x,d)`

`f(x,y)`

`f(x,y)`

Example 5.Example 4.

Note, we will use x and x' to show there is some legal substitution where the terms will unify

`f(x,y)`

`f(x,y)`

Example 3.h(x',y')&g(x',a) f(x',z)&g(x,'a)Figure 6: Abstract Example Case 1 If L2 is not a generalisation of R3, then C is a threatening step, which would result in the object moving to some substate other than the one required by step O . This is shown in Example 3 of Figure 6), where we can see that step C would move x to the substate

`h(x ; y)`

&`g(x ; a)`

, rather than the substate f (x ; y) which we are necessarily achieving with step A. This type of threat must be resolved by promotion or demotion.Case 2 We now consider the case where L2 is a generalisation of R3. Given that L2 is f (x ; y) and R3 is f (x 0; y0)&k (x 0; c), then provided there is some legal substitution for x , and y where it is possible that x = x 0 and y = y0, then we can say that L2 is a generalisation of R3, as in Example 4. In this case, C may or may not be a threatening step. There is the possibility that R3 will become instantiated to an expression which contains L2, but also the possibility that it may not. This type of threat may be resolved by either promotion, demotion or object separation. Object separation in the case of threats from necessary effects requires only the addition of a constraint that x and x' must not codesignate.

Note, it is generally thought in literal-based planning that it is more efficient to postpone resolving threats like these until it is certain that a threat exists. It may be that conflict resolution is unnecessary if the threat doesn't actually occur in this partial plan. Postponement is also thought to be more efficient as using traditional separation techniques can increase the amount of search needed, by requiring the addition of several possible non-codesignation

11

constraints, resulting in several new partial plans. For example, given a link fA;

`a(w ; z )`

; Og and a threatening action with a delete effect`a(x ; y)`

, it would be necessary to add constraints that x is not equal to w or y is not equal to z , generating one new partial plan for each possibility. Clearly, a link of the type f (y; z )&f (g; h) threatened by a similar literal would generate many additional plans each time such a threat had to be resolved. In object-centred planning we do not have this problem, with object separation we only ever have to add one constraint for resolving threats from necessary effects, and so can choose to resolve threats eagerly or lazily, without adding to the search space.Case 3 If R3 contains L2, then C is not a threatening step, and no threat resolution is necessary. If C is eventually necessarily ordered between A and O , it can just be considered as another achiever for L2. Note the reason for C 's introduction into the partial plan may have been to bring about a transition for some other object, and its achievement of L2 may just be a side-effect of this other transition.

4.2 Threats from Conditional Effects The threatening effect may be conditional. In this case, the notion of a threat and threat resolution is slightly different, as the object variable in a conditional effect is universally quantified. For example, in a conditional substate transition from f(x',y'),g(x',j') to f(x',y'), then x' refers to all objects of the correct sort whose substate matches with f(x',y'),g(x',j'). If we have a step which establishes f (x ; y);

`g(x ; j )`

, then this conditional transition is a potential threat. One of these objects referred to by x' may or may not be the same object as x, and we cannot tell until the final plan is produced and the final ordering of steps decided on, or the partially ordered steps become sufficiently instantiated to make a threat certain. This is because the right-hand side of a conditional effect is only effected if the left-hand side is true when the action is applied. Let us consider a situation where we have a substate expression frobot`in(Tom,Room)`

,robot next`key(Tom,Key)`

g and a possibly threatening transition from frobot`in(R; Rm)`

; robot next`key(R; K )`

g to frobot`in(R; Rm)`

g. If Rm and K can possibly be unified with Room and Key then a potential threat exists. If they are identical then a threat will exist in the current ordering of steps.Let us now consider a situation, where a step, A, in a partial plan has a member of A:rhs which we will call R1. Another step, O , has a member of O :lhs which we will call L2. Step A provides the required substate transition for L2. A third step, C , has a conditional transition from Cond :lhs to Cond:rhs, which we refer to as CL3 and CR3 respectively. A is ordered before O in the partial plan and C can potentially be ordered between A and O . The substate transitions for R1, L2 and CR3 all concern objects of the same sort.

We will examine all the possibilities, again using an abstract example (see Figure 7). In the case of Examples 1-5, CR3 does not contain L2 - that is, C is a potential threatening step. If CR3 contained L2, we could simply view C as also achieving L2. In every case, R1 contains L2. The cases differ according to the presence or absence of constants in CR3 and L2.

12

`f(x, y)`

&`g(x, a)`

y)&

`g(x, a)`

f(x,`f(x,y)`

&`g( x,a)`

`f(x,y)`

&`g(x,a)`

2

`f(x,y)`

&`g(x,a)`

L

`f(x, y)`

&`g(x, a)`

`f(x,y)`

&`g( x,a)`

`f(x, y)`

&`g(x, a)`

Example 3. Example 1 Example 2.

Example 4. Example 5.

Note, we will use x and x' to show where there is some legal substitution where the terms will unify

to indicate constants We will use bold type

R1

`f(x,y)`

&`g(x,a)`

`f(x,y)`

&`g(x,a)`

f(x',y')&h(x',b) f(x',y')&k(x',c)

y')f(x', CR

f(x, )y

3 f(x',y') y')&g(x',

f(x',y')&g(x',a')

) CL

f(x', y)&g(x', a)

3 f(x',y')&g(x',a)

f(x', a' f(x',y')&g(x',a')

Figure 7: Abstract Conditional Example Case 1 If there are no constants in either L2 or CL3, as in Example 1, then there is a potential threat, because the existentially quantified variables in CL3 or L2 could become instantiated to the same objects. If this happens, then because the key object variable in the conditional effect is universally quantified, any other object in that substate will be affected, and a definite threat exists.

In this case, we can resolve the threat by either object separation, promotion or demotion, or we can leave resolution until instantiation of the existentially quantified variables makes it certain that there will be a threat with regard to the particular object the L2 transition concerns. The situation is equivalent when the object variable in L2 has been replaced by a constant, as in Example 3.

Case 2 If there is a set of constants in CL3 which possibly unify with a set of existentially quantified variables in L2, as in Example 2, then there is a potential threat which we can resolve by either object separation, promotion or demotion. Object separation would require the addition of constraints that each constant and its corresponding variable should not be equal. The situation is equivalent if there are constants in place of the existentially quantified variables in L2, but none in CL3, as in Example 4.

Case 3 If there is a set of constants in CL3 that are the same as a set of constants in L2, as in Example 5, then it must be possible for any universally quantified variable in CL3 to be unified with the corresponding variable in L2. In this case, we can only resolve the threat by demotion or promotion. Using confrontation would not be satisfactory in this case: because L2 must be necessarily achieved, adding the negation of CL3 as a new subgoal may still clobber L2. The object-centred equivalent of adding the negation as a new subgoal would be to say that the

13

key object in L2 which can unify with the universally quantified variable in CL3 must be in some substate other than those specified by CL3 - which may mean some substate other than those specified by L2, and which we want to necessarily achieve with R1.

Case 4 If CR3 contains L2, then C is not a threatening step, as it also achieves L2. It is possible that the achieving effect in A is conditional, but in this case any variables contained in the link become existentially quantified, and so threat resolution is the same as if the achieving effect is necessary. Another possibility is that we may have had a previous threat from a necessary effect and added a separation constraint on the key objects, and then later we have a threat for the same object from a conditional effect. Now we have the situation where:

R1 = f (x ; y)&

`g(x ; a)`

L2 = f (x ; y) CL3 = (1) f (x 0y0) or (2) f (x 0y)If there are no constants in the threatening conditional effect (1), then we can add further constraints that y and y' do not codesignate. If y is already a constant (2) in CL3 then we can only resolve the threat by promotion or demotion.

To summarise, threats from both necessary and conditional effects can be resolved by one of the following techniques: demotion, promotion and object separation. At its simplest, object separation requires the addition of a constraint that the object identified in the link and the object in the threatening effect are not the same.

5 Conclusion Traditionally in the literature of classical planning, domain models were expressed in a STRIPS-style [4] notation. More recently we have seen domains written in PDDL. Models expressed in either of these formats function at the level of the literal. They both have as their main feature a set of operator schema modelling domain actions, which transform states by adding and deleting literals. An object-centred operator transforms the entire state of an object, and doesn't just add or delete literals. Object-centred domain models are able to capture and express explicitly the legal behaviour of objects, and enable us to deal with behaviour at the object level.

The object-centred approach offers an expressive domain modelling language and associated domain encoding method that is well-founded and aids domain validation and maintenance. It provides a method that allows the non-AI planning expert to construct a precise, validated domain model. It provides opportunities for the additional structural information about the domain to be exploited by planning algorithms.

We have discussed truth criteria and described our own truth criterion for object-centred planning with conditional actions. This truth criterion describes what conditions are necessary for ensuring that a substate expression is true at some point in a partial plan. We can make

14

use of the truth criterion during planning in threat detection and resolution procedures. We have shown that the use of OCL and object-centred planning allows us to simplify existing threat resolution techniques whether the threatening effect is necessary or conditional. This object-centred truth criterion provides a sound basis for the development of object-centred planning algorithms.

References

[1] A. Tate (editor). Advanced Planning Technology: Technological Achievements of the ARPA/Rome Laboratory Planning Initiative. IOS Press, 1996.

[2] D. Chapman. Planning for Conjunctive Goals. Artificial Intelligence, 32, 1987. [3] AIPS-98 Planning Competition Committee. PDDL - The Planning Domain Definition Language. Technical Report CVC TR-98-003/DCS TR-1165, Yale Center for Computational Vision and Control, 1998.

[4] R. E. Fikes and N. J. Nilsson. STRIPS: A New Approach to the Application of Theorem Proving to

Problem Solving. Artificial Intelligence, 2, 1971.

[5] M. Fox and D. Long. The Automatic Inference of State Invariants in TIM. JAIR vol. 9, pages 367-421,

1997.

[6] A. Gerevini and L. Schubert. Computing Parameter Domains as an Aid to Planning. In Third International Conference on Artificial Intelligence Planning Systems, 1996.

[7] S. Kambhampati. Formalizing a Spectrum of Plan Generalizations Based on Modal Truth Criteria . In

Canadian AI Conference, 1994.

[8] S. Kambhampati and D. S. Nau. On the Nature of Modal Truth Criteria in Planning. In Proceedings of

the Twelfth National Conference on Artificial Intelligence, 1994.

[9] D. McAllester and D. Rosenblitt. Systematic Nonlinear Planning. In Proceedings of the Ninth National

Conference on Artificial Intelligence. AAAI Press, 1991.

[10] T. L. McCluskey and D. E. Kitchin. A Tool-Supported Approach to Engineering HTN Planning Models.

In Proceedings of 10th IEEE International Conference on Tools with Artificial Intelligence, 1998.

[11] T. L. McCluskey and J. M. Porteous. Engineering and Compiling Planning Domain Models to Promote

Validity and Efficiency. Artificial Intelligence, 95:1-65, 1997.

[12] J. S. Penberthy and D. S. Weld. UCPOP: A Sound, Complete, Partial Order Planner for ADL. In Proc.

KR, 1992.

[13] Planform. An Open Environment for Building Planners. http://helios.hud.ac.uk/planform. [14] J. Rintanen. A planning algorithm not based on directional search. In Proceedings of the 6th International

Conference on Principles of Knowledge Representation, 1998.

[15] S. A. Chien (editor). 1st NASA Workshop on Planning and Scheduling in Space Applications. NASA,

Oxnard, CA, 1997.

[16] R. M. Simpson, T. L. McCluskey, D. Liu, and D. E. Kitchin. Knowledge Representation in Planning: A

PDDL to OCLh Translation. In Proceedings of the 12th International Symposium on Methodologies for Intelligent Systems, 2000.

[17] D. S. Weld. An Introduction to Least Commitment Planning. AI Magazine, 1994.

15