Temporal Description Logics

Journal of Arti cial Intelligence Research 9 (1998) 463-506 Submitted 4/98; published 12/98
A Temporal Description Logic

for Reasoning about Actions and Plans
Alessandro Artale artale@irst.itc.it
ITC-IRST, Cognitive and Communication Technologies Division
I-38050 Povo TN, Italy
Enrico Franconi franconi@cs.man.ac.uk
Department of Computer Science, University of Manchester
Manchester M13 9PL, UK
Abstract
A class of interval-based temporal languages for uniformly representing and reasoning
about actions and plans is presented. Actions are represented by describing what is true
while the action itself is occurring, and plans are constructed by temporally relating actions
and world states. The temporal languages are members of the family of Description Logics,
which are characterized by high expressivity combined with good computational properties.
The subsumption problem for a class of temporal Description Logics is investigated and
sound and complete decision procedures are given. The basic language TL-F is considered
rst: it is the composition of a temporal logic TL { able to express interval temporal
networks { together with the non-temporal logic F { a Feature Description Logic. It is
proven that subsumption in this language is an NP-complete problem. Then it is shown
how to reason with the more expressive languages TLU -FU and TL-ALCF . The former
adds disjunction both at the temporal and non-temporal sides of the language, the latter
extends the non-temporal side with set-valued features (i.e., roles) and a propositionally
complete language.
1. Introduction
The representation of temporal knowledge has received considerable attention in the Ar-
ti cial Intelligence community in an attempt to extend existing knowledge representation
systems to deal with actions and change. At the same time, many logic-based formalisms
were developed and analyzed by logicians and philosophers for the same purposes. In this
class of logical formalisms, properties such as expressive power and computability have been
studied as regards typical problems involving events and actions.
This paper analyzes from a theoretical point of view the logical and computational
properties of a knowledge representation system that allows us to deal with time, actions
and plans in a uniform way. The most common approaches to model actions are based
on the notion of state change { e.g., the formal models based on the original situation
calculus (McCarthy & Hayes, 1969; Sandewall & Shoham, 1994) or the Strips-like planning
systems (Fikes & Nilsson, 1971; Lifschitz, 1987) { in which actions are generally considered
instantaneous and de ned as functions from one state to another by means of pre- and
post-conditions. Here, an explicit notion of time is introduced in the modeling language
and actions are de ned as occurring over time intervals, following the Allen proposal (Allen,
c 1998 AI Access Foundation and Morgan Kaufmann Publishers. All rights reserved.
Artale & Franconi
1991). In this formalism an action is represented by describing the time course of the
world while the action occurs. Concurrent or overlapping actions are allowed: e ects of
overlapping actions can be di erent from the sum of their individual e ects; e ects may
not directly follow the action but more complex temporal relations may hold. For instance,
consider the motion of a pointer on a screen driven by a mouse: the pointer moves because
of the movement of the device on the pad { there is a cause-e ect relation { but the two
events are contemporary, in the common-sense notion of the word.
A class of interval temporal logics is studied based on Description Logics and inspired by
the works of Schmiedel (1990) and of Weida and Litman (1992). In this class of formalisms
a state describes a collection of properties of the world holding at a certain time. Actions are
represented through temporal constraints on world states, which pertain to the action itself.
Plans are built by temporally relating actions and states. To represent the temporal dimen-
sion classical Description Logics are extended with temporal constructors; thus a uniform
representation for states, actions and plans is provided. Furthermore, the distinction made
by Description Logics between the terminological and assertional aspects of the knowledge
allows us to describe actions and plans both at an abstract level (action/plan types) and
at an instance level (individual actions and plans). In this environment, the subsumption
calculus is the main inference tool for managing collections of action and plan types. Action
and plan types can be organized in a subsumption-based taxonomy, which plays the role
of an action/plan library to be used for the tasks known in the literature as plan retrieval
and individual plan recognition (Kautz, 1991). A re nement of the plan recognition no-
tion is proposed, by splitting it into the di erent tasks of plan description classi cation {
involving a plan type { and speci c plan recognition with respect to a plan description {
involving an individual plan. According to the latter reasoning task, the system is able to
recognize which type of action/plan has taken place at a certain time interval, given a set
of observations of the world.
Advantages of using Description Logics are their high expressivity combined with de-
sirable computational properties { such as decidability, soundness and completeness of de-
duction procedures (Buchheit, Donini, & Schaerf, 1993; Schaerf, 1994; Donini, Lenzerini,
Nardi, & Schaerf, 1994; Donini, Lenzerini, Nardi, & Nutt, 1995). The main purpose of this
work is to investigate a class of decidable temporal Description Logics, and to provide com-
plete algorithms for computing subsumption. To this aim, we start with TL-F , a language
being the composition of a temporal logic TL { able to express interval temporal networks {
together with the non-temporal Description Logic F { a Feature Description Logic (Smolka,
1992). It turns out that subsumption for TL-F is an NP-complete problem. Then, we show
how to reason with more expressive languages: TLU -FU , which adds disjunction both at
the temporal and non-temporal sides of the language, and TL-ALCF , which extends the
non-temporal side with set-valued features (i.e., roles) and a propositionally complete De-
scription Logic (Hollunder & Nutt, 1990). In both cases we show that reasoning is decidable
and we supply sound and complete procedures for computing subsumption.
The paper is organized as follows. After introducing the main features of Description
Logics in Section 2, Section 3 organizes the intuitions underlying our proposal. The technical
bases are introduced by brie y overviewing the temporal extensions of Description Logics
relevant for this approach { together with the inter-relationships with the interval temporal
modal logic { speci cally intended for time and action representation and reasoning. The
464
A Temporal Description Logic for Reasoning about Actions and Plans
basic feature temporal language (TL-F ) is introduced in Section 4. The language syntax is
rst described in Section 4.1, together with a worked out example illustrating the informal
meaning of temporal expressions. Section 4.2 reveals the model theoretic semantics of TL-F ,
together with a formal de nition of the subsumption and instance recognition problems.
Section 5 shows that the temporal language is suitable for action and plan representation
and reasoning: the well known cooking domain and blocks world domain are taken into
consideration. The sound and complete calculus for the feature temporal language TL-F
is presented in details in Section 6. A proof that subsumption for TL-F is an NP-complete
problem is included. The calculus for TL-F forms the basic reasoning procedure that
can be adapted to deal with logics having an extended propositional part. An algorithm
for checking subsumption in presence of disjunction (TLU -FU ) is devised in Section 7.1;
while in Section 7.2 the non-temporal part of the language is extended with roles and
full propositional calculus (TL-ALCF ). In both cases, the subsumption problem is still
decidable. Operators for homogeneity and persistence are presented in Section 8 for an
adequate representation of world states. In particular, a possible solution to the frame
problem, i.e., the problem to compute what remains unchanged by an action, is suggested.
Section 9 surveys the whole spectrum of extensions of Description Logics for representing
and reasoning with time and action. This Section is concluded by a comparison with State
Change based approaches by brie y illustrating the e ort made in the situation calculus
area to temporally extend this class of formalisms. Section 10 concludes the paper.
2. Description Logics
Description Logics1 are formalisms designed for a logical reconstruction of representa-
tion tools such as frames, semantic networks, object-oriented and semantic data models
{ see (Calvanese, Lenzerini, & Nardi, 1994) for a survey. Nowadays, Description Logics
are also considered the most important unifying formalism for the many object-centered
representation languages used in areas other than Knowledge Representation. Important
characteristics of Description Logics are high expressivity together with decidability, which
guarantee the existence of reasoning algorithms that always terminate with the correct
answers.
This Section gives a brief introduction to a basic Description Logic, which will serve as
the basic representation language for our proposal. As for the formal apparatus, the formal-
ism introduced by (Schmidt-Schau & Smolka, 1991) and further elaborated by (Donini,
Hollunder, Lenzerini, Spaccamela, Nardi, & Nutt, 1992; Donini et al., 1994, 1995; Buchheit
et al., 1993; De Giacomo & Lenzerini, 1995, 1996) is followed: in this way, Description
Logics are considered as a structured fragment of predicate logic. ALC (Schmidt-Schau &
Smolka, 1991) is the minimal Description Logic including full negation and disjunction {
i.e., propositional calculus, and it is a notational variant of the propositional modal logic
K(m) (Halpern & Moses, 1985; Schild, 1991).
The basic types of a Description Logic are concepts, roles, features, and individuals. A
concept is a description gathering the common properties among a collection of individuals;
from a logical point of view it is a unary predicate ranging over the domain of individu-
1. Description Logics have been also called Frame-Based Description Languages, Term Subsumption Lan-
guages, Terminological Logics, Taxonomic Logics, Concept Languages or KL-One-like languages.
465
Artale & Franconi
C; D ! A j (atomic concept)
>j (top)
?j (bottom)
:C j (complement)
C uD j (conjunction)
C tD j (disjunction)
8P .C j (universal quanti er)
9P .C j (existential quanti er)
p:Cj (selection)
p#qj (agreement)
p"qj (disagreement)
p" (unde nedness)
p; q ! f j (atomic feature)
pq (path)
Figure 1: Syntax rules for the ALCF Description Logic.
als. Properties are represented either by means of roles { which are interpreted as binary
relations associating to individuals of a given class values for that property { or by means
of features { which are interpreted as functions associating to individuals of a given class
a single value for that property. The language ALCF , extending ALC with features (i.e.,
functional roles) is considered. By the syntax rules of Figure 1, ALCF concepts (denoted by
the letters C and D) are built out of atomic concepts (denoted by the letter A), atomic roles
(denoted by the letter P ), and atomic features (denoted by the letter f ). The syntax rules
are expressed following the tradition of Description Logics (Baader, Burckert, Heinsohn,
Hollunder, Muller, Nebel, Nutt, & Pro tlich, 1990).
The meaning of concept expressions is de ned as sets of individuals, as for unary pred-
icates, and the meaning of roles as sets of pairs of individuals, as for binary predicates.
Formally, an interpretation is a pair I = (I ; I ) consisting of a set I of individuals (the
domain of I ) and a function I (the interpretation function of I ) mapping every concept
to a subset of I , every role to a subset of I I , every feature to a partial function
from I to I , and every individual into a di erent element of I { i.e., aI 6= bI if a 6= b
(Unique Name Assumption) { such that the equations of the left column in Figure 2 are
satis ed.
The ALCF semantics identi es concept expressions as fragments of rst-order predicate
logic. Since the interpretation I assigns to every atomic concept, role or feature a unary or
binary (functional) relation over I , respectively, one can think of atomic concepts, roles
and features as unary and binary (functional) predicates. This can be seen as follows: an
atomic concept A, an atomic role P , and an atomic feature f , are mapped respectively to
the open formulas FA ( ), P ( ; ), and Ff ( ; ) with Ff satisfying the functionality axiom
8y; z.Ff (x; y) ^ Ff (x; z) ! y = z { i.e., Ff is a functional relation.
The rightmost column of Figure 2 gives the transformational semantics of ALCF ex-
pressions in terms of FOL well-formed formul, while the left column gives the standard
extensional semantics. As far as the transformational semantics is concerned, a concept C ,
a role P and a path p correspond to the FOL open formulae FC ( ), FP ( ; ), and Fp ( ; ),
466
>I = I true
?I = ; false
(:C )I = I n C I :FC ( )
(C u D)I = C I \ DI FC ( ) ^ FD ( )
(C t D)I = C [D
I I FC ( ) _ FD ( )
(9P .C )I = fa 2 j 9b.(a; b) 2 P ^ b 2 C g
I I I 9x.FP ( ; x) ^ FC (x)
(8P .C )I = fa 2 I j 8b.(a; b) 2 P I ) b 2 C I g 8x.FP ( ; x) ) FC (x)
(p : C )I =fa 2 dom pI j pI (a) 2 C I g 9x.Fp ( ; x) ^ FC (x)
p # qI =fa 2 dom pI \ dom qI j pI (a) = qI (a)g (9x.Fp ( ; x) ^ Fq ( ; x))
p " qI =fa 2 dom pI \ dom qI j pI (a) 6= qI (a)g (9x; y.Fp ( ; x) ^ Fq ( ; y))^
(8x; y .Fp ( ; x) ^ Fq ( ; y ) ! x 6= y )
(p ")I = I n dom pI :9x.Fp ( ; x)
(p q) = p q
I I I 9x.Fp ( ; x) ^ Fq (x; )
Figure 2: The extensional and transformational semantics in ALCF .
respectively. It is worth noting that the extensional semantics of the left column gives also
an interpretation for the formulas of the right column so that the following proposition
holds.
Proposition 2.1 (Concepts vs. fol formul) Let C be an ALCF concept expression.
Then the transformational semantics of Figure 2 maps C into a logically equivalent rst
order formula.
A terminology or TBox is a nite set of terminological axioms. For an atomic concept A,
called de: ned concept, and a (possibly complex) concept C , a terminological axiom is of the
form A = C . An atomic concept not appearing on the left-hand side of any terminological
axiom is called a primitive concept. Acyclic simple TBoxes only are considered: a de ned
concept may appear at most once as the left-hand side of an axiom, and no terminological
cycles are allowed, i.e., no de ned concept may occur { neither directly: nor indirectly {
within its own de nition (Nebel, 1991). An interpretation I satis es A = C if and only if
AI = C I .
As an example, consider the unary relation (i.e., a concept) denoting the class of happy
fathers, de ned using the atomic predicates Man, Doctor, Rich, Famous (concepts) and
CHILD, FRIEND (roles):
:
HappyFather = Man u (9CHILD.>) u 8CHILD.(Doctor u 9FRIEND.(Rich t Famous))
i.e., the men whose children are doctors having some rich or famous friend.
An ABox is a nite set of assertional axioms, i.e. predications on individual objects. Let
O be the alphabet of symbols denoting individuals; an assertion is an axiom of the form
C (a), R(a; b) or p(a; b), where a and b denote individuals in O. C (a) is satis ed by an
interpretation I i aI 2 C I , P (a; b) is satis ed by I i (aI ; bI ) 2 P I , and p(a; b) is satis ed
by I i pI (aI ) = bI .
467
Artale & Franconi
A knowledge base is a nite set of terminological and assertional axioms. An interpre-

tation I is a model of a knowledge base i every axiom of is satis ed by I . logically
implies A v C (written j= A v C ) if AI C I for every model of : we say that A is
subsumed by C in . The reasoning problem of checking whether A is subsumed by C in
is called subsumption checking. logically implies C (a) (written j= C (a)) if aI 2 C I
for every model of : we say that a is an instance of C in . The reasoning problem of
checking whether a is an instance of C in is called instance recognition.
An acyclic simple TBox can be transformed into an expanded TBox having the same
models, where no de ned concept makes use in its de nition of any other de ned concept.
In this way, the interpretation of a de ned concept in an expanded TBox does not depend
on any other de ned concept. It is easy to see that A is subsumed by C in an acyclic simple
TBox if and only if the expansion of A with respect to is subsumed by the expansion of
C with respect to in the empty TBox. The expansion procedure recursively substitutes
every de ned concept occurring in a de nition with its de ning expression; such a procedure
may generate a TBox exponential in size, but it has been proved (Nebel, 1990) that it works
in polynomial time under reasonable restrictions. The following interchangeably refers either
to reasoning with respect to a TBox or to reasoning involving expanded concepts with an
empty TBox. In particular, while devising the subsumption calculus for the logics considered
here, it is always assumed that all de ned concepts have been expanded.
3. Towards a Temporal Description Logics

Schmiedel (1990) proposed to extend Description Logics with an interval{based temporal
logic. The temporal variant of the Description Logic is equipped with a model-theoretic
semantics. The underlying Description Logic is FLENR (Donini et al., 1995): it di ers
from ALCF in that it does not contain the > and ? concepts, it does not have neither
negation nor disjunction, and it has cardinality restrictions and conjunction over roles.
The new temporal term-forming operators are the temporal quali er at, the existential and
universal temporal quanti ers sometime and alltime. The quali er operator speci es the
time at which a concept holds. The temporal quanti ers introduce the temporal variables
constrained by means of temporal relationships based on Allen's interval algebra extended
with metric constraints to deal with durations, absolute times, and granularities of intervals.
To give an example of this temporal Description Logic, the concept of Mortal can be de ned
by:
:
Mortal = LivingBeing u (sometime(x) (after x NOW) (at x (:LivingBeing)))
with the meaning of a LivingBeing at the reference interval NOW, who will not be alive
at an interval x sometime after the reference interval NOW. Schmiedel does not propose
any algorithm for computing subsumption, but gives some preliminary hints. Actually,
Schmiedel's logic is argued to be undecidable (Bettini, 1997), sacri cing the main bene t
of Description Logics: the possibility of having decidable inference techniques.
Schmiedel's temporal Description Logic, when closed under complementation, contains
as a proper fragment the temporal logic HS proposed by Halpern and Shoham (1991).
The logic HS is a propositional modal logic which extends propositional logic with modal
formul of the kind hRi. and [R]. { where R is a basic Allen's temporal relation and hi
468
and [] are the possibility and necessity modal operators. For example, the modal formula
LivingBeing ^ hafteri.:LivingBeing corresponds to the abovementioned Mortal concept.
Unfortunately, the HS logic is shown to be undecidable, at least for most interesting classes
of temporal structures: \One gets decidability only in very restricted cases, such as when
the set of temporal models considered is a nite collection of structures, each consisting of
a nite set of natural numbers." (Halpern & Shoham, 1991)
Weida and Litman (1992, 1994) propose T-Rex, a loose hybrid integration between
Description Logics and constraint networks. Plans are de ned as collections of steps together
with temporal constraints between their duration. Each step is associated with an action
type, represented by a generic concept in K-Rep { a non-temporal Description Logic. Thus
a plan is seen as a plan network, a temporal constraint network whose nodes, corresponding
to time intervals, are labeled with action types and are associated with the steps of the plan
itself. As an example of plan in T-Rex they show the plan of preparing spaghetti marinara:
( defplan Assemble-Spaghetti-Marinara
((step1 Boil-Spaghetti)
(step2 Make-Marinara)
(step3 Put-Together-SM))
((step1 (before meets) step3)
(step2 (before meets) step3)))
This is a plan composed by three actions, i.e., boiling spaghetti, preparing marinara sauce,
and assembling all things at the end. Temporal constraints between the steps establish
the temporal order in doing the corresponding actions. A structural plan subsumption
algorithm is de ned, characterized in terms of graph matching, and based on two separate
notions of subsumption: pure terminological subsumption between action types labeling
the nodes, and pure temporal subsumption between interval relationships labeling the arcs.
The plan library is used to guide plan recognition (Weida, 1996) in a way similar to that
proposed by Kautz (1991). Even if this work has strong motivations, no formal semantics
is provided for the language and the reasoning problems.
Starting from the assumption that an action has a duration in time, our proposal con-
siders an interval-based modal temporal logic { in the spirit of Halpern and Shoham (1991)
{ and reduces the expressivity of (Schmiedel, 1990) in the direction of (Weida & Litman,
1992). While Schmiedel's work lacks computational machinery, and Halpern and Shoham's
logic is undecidable, here an expressive decidable logic is obtained, providing sound and
complete reasoning algorithms. Di erently from T-Rex which uses two di erent languages
to represent actions and plans { a non temporal Description Logic for describing actions
and a second language to compose plans by adding temporal information { here an exten-
sion of a Description Logic is chosen in which time operators are available directly as term
constructors. This view implies an integration of a temporal domain in the semantic struc-
ture where terms themselves are interpreted, giving the formal way both for a well-founded
notion of subsumption and for proving soundness and completeness of the corresponding
procedure. As an example of the formalism, the plan for preparing spaghetti marinara
introduced above is represented as follows:
469
Artale & Franconi
Assemble-Spaghetti-Marinara =
: 3(y z w) (y (before; meets) w)(z (before; meets) w).
(Boil-Spaghetti@y u
Make-Marinara@z u
Put-Together-SM@w)
Moreover, it is possible to build temporal structured actions { as opposed to the atomic

actions proposed in T-Rex { describing how the world state changes because of the occur-
rence of an action: in fact, our language allows for feature representation in order to relate
actions to states of the world (see Section 5.2). This kind of expressivity is not captured
by T-Rex, since it uses a non-temporal Description Logic to represent actions. The main
application of T-Rex is plan recognition; according to the ideas of Kautz (1991) a Closed
World Assumption (CWA) (Weida, 1996) is made, assuming that the plan library is com-
plete and an observed plan will be fully accounted for by a single plan. CWA is not relied
on here, following the Open World Semantics characterizing Description Logics. Weaker,
but monotonic, deductions are allowed in the plan recognition process. However, their pro-
cedures for recognizing a necessary, optional or impossible individual plan with respect to
a plan description is still applicable, if the plan library is given a closed world semantics.
4. The Feature Temporal Language TL-F

The feature temporal language TL-F is the basic logic considered here. This language is
composed of the temporal Logic TL { able to express interval temporal networks { and the
non-temporal Feature Description Logic F . Note that, each logic of the family of Temporal
Description Logics introduced in this paper is identi ed by a composed string in which
the rst part refers to the temporal part of the language while the other one refers to the
non-temporal part.
4.1 Syntax
Basic types of the language are concepts , individuals, temporal variables and intervals.
Concepts can describe entities of the world, states and events. Temporal variables denote
intervals bound by temporal constraints, by means of which abstract temporal patterns in
the form of constraint networks are expressed. Concepts (resp. individuals) can be speci ed
to hold at a certain temporal variable (resp. interval). In this way, action types (resp.
individual actions) can be represented in a uniform way by temporally related concepts
(resp. individuals).
For the basic temporal interval relations the Allen notation (Allen, 1991) (Figure 3) is
used: before (b), meets (m), during (d), overlaps (o), starts (s), nishes (f), equal (=), after
(a), met-by (mi), contains (di), overlapped-by (oi), started-by (si), nished-by ( ). Concept
expressions (denoted by C; D) are syntactically built out of atomic concepts (denoted by A),
atomic features (denoted by f ), atomic parametric features (denoted by ?g) and temporal
variables (denoted by X; Y ). Temporal concepts (C; D) are distinguished from non-temporal
concepts (E; F ), following the syntax rules of Figure 4. Names for atomic features and
atomic parametric features are from the same alphabet of symbols; the ? symbol is not
intended as operator, but only as di erentiating the two syntactic types.
470
Relation Abbr. Inverse i j

before(i; j ) b a
meets(i; j ) m mi
overlaps(i; j ) o oi
starts(i; j ) s si
during(i; j ) d di
finishes(i; j ) f
Figure 3: The Allen's interval relationships.
Temporal variables are introduced by the temporal existential quanti er \3" { excluding
the special temporal variable ], usually called NOW, and intended as the reference interval.
Variables appearing in temporal constraints (Tc) must be declared within the same temporal
quanti er, with the exception of the special variable ]. Temporal variables appearing in the
right hand side of an \@" operator are called bindable. Concepts must not include unbound
(a.k.a. free) bindable variables. Informally, a bindable variable is said to be bound in a
concept if it is declared at the nearest temporal quanti er in the body of which it occurs;
this avoid the usual formal inductive de nition of a bound variable. Moreover, in chained
constructs of the form ((C [Y1 ]@X1 )[Y2 ]@X2 : : :) non bindable variables { i.e., the ones on
the left hand side of an \@" operator { cannot appear more than once. Note that, since
Description Logics are a fragment of FOL with one free variable, the above mentioned
restrictions force the temporal side of the language to have only one free temporal variable,
i.e., the reference time ].
As usual, terminological axioms for building simple acyclic TL-F TBoxes are allowed.
While using in a concept expression a name referring to a de ned concept, it is possible to
use the substitutive quali er construct, to impose a coreference with a variable appearing
in the de nition associated to the de ned concept. The statement C [Y ]@X constrains the
variable Y , which should appear in the de nition of the de ned concept C , to corefer with X
(see Section 5.2 for an example). A drawback in the use of this operator is the requirement
to know the internal syntactical form of the de ned concept, namely, the names of its
temporal variables.
Let O and OT be two alphabets of symbols denoting individuals and temporal intervals,
respectively. An assertion { i.e., a predication on temporally quali ed individual entities {
is a statement of one of the forms C (i; a); p(i; a; b); ?g(a; b); R(i; j ), where C is a concept, p
is a feature, ?g is a parametric feature, R is a temporal relation, a and b denote individuals
in O, i and j denote temporal intervals in OT .
471
Artale & Franconi
TL C; D ! E j (non-temporal concept)
C uD j (conjunction)
C @X j (quali er)
C [Y ]@X j (substitutive quali er)
3(X ) Tc.C (existential quanti er)
Tc ! (X (R) Y ) j (temporal constraint)
(X (R) ]) j
(] (R) Y )
Tc ! Tc j Tc Tc
R; S ! R , S j (disjunction)
s j mi j f j : : : (Allen's relations)
X; Y ! x j y j z j : : : (temporal variables)
X ! XjX X
F E; F ! A j (atomic concept)
>j (top)
EuF j (conjunction)
p#qj (agreement)
p:E (selection)
p; q ! f j (atomic feature)
?g j (atomic parametric feature)
pq (path)
Figure 4: Syntax rules for the interval Description Logic TL-F
4.1.1 A clarifying Example

Let us now informally see the intended meaning of the terms of the language TL-F (for the
formal details see Section 4.2). Concept expressions are interpreted over pairs of temporal
intervals and individuals hi; ai, meaning that the individual a is in the extension of the con-
cept at the interval i. If a concept is intended to describe an action, then its interpretation
can be seen as the set of individual actions of that type occurring at some interval.
Within a concept expression, the special \]" variable denotes the current interval of
evaluation; in the case of actions, it is thought that it refers to the temporal interval
at which the action itself occurs. The temporal existential quanti er introduces interval
variables, related to each other and possibly to the ] variable in a way de ned by the set of
temporal constraints. To evaluate a concept at an interval X , di erent from the current one,
it is necessary to temporally qualify it at X { written C @X ; in this way, every occurrence of
472
] -
Basic-Stack(BLOCK)
x -
OnTable(BLOCK)
y
OnBlock(BLOCK)
-
Figure 5: Temporal dependencies in the de nition of the Basic-Stack action.
] embedded within the concept expression C is interpreted as the X variable2 . The informal
meaning of a concept with a temporal existential quanti cation can be understood with the
following examples in the action domain.
:
Basic-Stack = 3(x y ) (x m ])(] m y ). ((?BLOCK : OnTable)@x u (?BLOCK : OnBlock)@y )
Figure 5 shows the temporal dependencies of the intervals in which the concept Basic-Stack
holds. Basic-Stack denotes, according to the de nition (a terminological axiom), any
action occurring at some interval involving a ?BLOCK that was once OnTable and then
OnBlock. The ] interval could be understood as the occurring time of the action type being
de ned: referring to it within the de nition is an explicit way to temporally relate states
and actions occurring in the world with respect to the occurrence of the action itself. The
temporal constraints (x m ]) and (] m y) state that the interval denoted by x should meet
the interval denoted by ] { the occurrence interval of the action type Basic-Stack { and
that ] should meet y. The parametric feature ?BLOCK plays the role of formal parameter of
the action, mapping any individual action of type Basic-Stack to the block to be stacked,
independently from time. Please note that, whereas the existence and identity of the ?BLOCK
of the action is time invariant, it can be quali ed di erently in di erent intervals of time,
e.g., the ?BLOCK is necessarily OnTable only during the interval denoted by x.
Let us comment now on the introduction of explicit temporal variables. The absence of
explicit temporal variables would weaken the temporal structure of a concept since arbitrary
relationships between more than two intervals could not be represented anymore. For
example, having only implicit intervals it is not possible to describe the situation in which
two concept expressions, say C and D, hold at two meeting intervals (say x, y) with the rst
interval starting and the second nishing the reference interval (i.e., the temporal pattern
(x meets y)(x starts ])(y nishes ]) cannot be represented). More precisely, it is not possible
to represent temporal relations between more than two intervals if they are not derivable by
the temporal propagation of the constraints imposed on pairs of variables. While explicit
variables go against the general thrust of Description Logics, the gained expressive power
together with the observation that the variables are limited only to the temporal part of
the language are the main motivations for using them. However, it is easy to drop them by
limiting the temporal expressiveness as proposed by Bettini (1997) (see also Section 9).
An assertion of the type Basic-Stack(i; a) states that the individual a is an action of
the type Basic-Stack occurred at the interval i. Moreover, the same assertion implies that
a is related to a ?BLOCK , say b, which is of type OnTable at some interval j , meeting i, and
of type OnBlock at another interval l, met by i.
2. Since any concept is implicitly temporally quali ed at the special ] variable, it is not necessary to
explicitly qualify concepts at ].
473
Artale & Franconi
(s)E = fh[u; v]; [u1 ; v1 ]i 2 T<? T<? j u = u1 ^ v < v1 g

(f )E = fh[u; v]; [u1 ; v1 ]i 2 T<? T<? j v = v1 ^ u1 < ug
(mi)E = fh[u; v]; [u1 ; v1 ]i 2 T<? T<? j u = v1 g
: : : (meaning of the other Allen temporal relations)
(R , S )E = RE [ S E
hX; TciE = fV : X 7! T<? j 8(X (R) Y ) 2 Tc. hV (X ); V (Y )i 2 RE g:
Figure 6: The temporal interpretation function.
Basic-Stack( i; a) =) 9b. ?BLOCK(a; b) ^ 9j; l. (OnTable(j; b) ^ OnBlock(l; b) ^

m(j; i) ^ m(i; l))
An individual action is an object in the conceptual domain associated with the relevant
properties { or states { of the world a ected by the individual action itself via a bunch of
features; moreover, temporal relations constrain time intervals imposing an ordering in the
change of the states of the world.
4.2 Semantics
In this Section, a Tarski-style extensional semantics for the TL-F language is given, and a
formal de nition of the subsumption and recognition reasoning tasks is devised.
Assume a linear, unbounded, and dense temporal structure T = (P ; <), where P is
a set of time points and < is a strict partial order on P . In such a structure, given an
interval X and a temporal relation R, it is always possible to nd an interval Y such that
(X (R) Y ). The assumption of linear time { which means that for any two points t1 and
t2 such that t1 t2 the set of points ft j t1 t t2 g is totally ordered { ts the intuition
about the nature of time, so that the pair [t1 ; t2 ] can be thought as the closed interval of
points between t1 and: t2 . The interval set of a structure T is de ned as the set T<? of all
closed intervals [u; v] = fx 2 P j u x v; u 6= vg in T .
A primitive interpretation I =: hT<? ; I ; I i consists of a set T<? (the interval set of
the selected temporal structure T ), a set I (the domain of I ), and a function I (the
primitive interpretation function of I ) which gives a meaning to atomic concepts, features
and parametric features:
AI T<? I ; f I : (T<? I ) partial
7 ! I ; ?gI : I partial
7 ! I
Atomic parametric features are interpreted as partial functions; they di er from atomic
features for being independent from time.
In order to give a meaning to temporal expressions present in generic concept expres-
sions, Figure 6 de nes the temporal interpretation function. The temporal interpretation
function E depends only on the temporal structure T . The labeled directed graph hX; Tci
{ where X is the set of variables representing the nodes, and Tc is the set of temporal con-
straints representing the arcs { is called temporal constraint network. The interpretation
474
AIV ;t;H = fa 2 I j ht; ai 2 AI g = AIt

>IV ;t;H = I = >I
(C u D)IV ;t;H = CVI ;t;H \ DVI ;t;H
(p # q)IV ;t;H = fa 2 dom pIt \ dom qtI j pIt (a) = qtI (a)g = (p # q)It
(p : C )IV ;t;H = fa 2 dom pIt j pIt (a) 2 CVI ;t;Hg
(C @X )IV ;t;H = CVI ;V (X );H
(C [Y ]@X )IV ;t;H = CVI ;t;H[fY 7!V (X )g
(3(X ) Tc. C )IV ;t;H = fa 2 I j 9W . W 2 hX; TciEH[f]7!tg ^ a 2 CWI ;t;; g
ftI = f^t : I partial
7 ! I j 8a. (a 2 dom f^t $ ht; ai 2 dom f I ) ^
f^t (a) = f I (t; a)
(p q)It = pIt qtI
?gtI = ?gI
Figure 7: The interpretation function.
of a temporal constraint network is a set of variable assignments that satisfy the temporal
constraints. A variable assignment is a function V : X 7! T<? associating an interval value to
a temporal variable. A temporal constraint network is consistent if it admits a non empty
interpretation. The notation, hX; TciEfx1 7!t1 ;x2 7!t2 ;:::g , used to interpret concept expressions,
denotes the subset of hX; TciE where the variable xi is mapped to the interval value ti .
It is now possible to interpret generic concept expressions. Consider the equations
introduced in Figure 7. An interpretation function IV ;t;H, based on a variable assignment
V , an interval t and a set of constraints H = fx1 7! t1; : : :g over the assignments of inner
variables, extends the primitive interpretation function in such a way that the equations
of Figure 7 are satis ed. Intuitively, the interpretation of a concept CVI ;t;H is the set of
entities of the domain that are of type C at the time interval t, with the assignment for the
free temporal variables in C given by V { see (C @X )IV ;t;H { and with the constraints for
the assignment of variables in the scope of the outermost temporal quanti ers given by H.
Note that, H interprets the variable renaming due to the temporal substitutive quali er {
see (C [Y ]@X )IV ;t;H { and it takes e ect during the choice of a variable assignment, as the
equation for (3(X ) Tc. C )IV ;t;H shows.
In absence of free variables in the concept expression { with the exception of ]{ for
notational simpli cation the natural interpretation function CtI ; being equivalent to the
interpretation function CVI ;t;H with any V such that V (]) = t and H = ; is introduced. The
set of interpretations fCVI ;t;Hg obtained by varying I ; V ; t with a xed H is maximal wrt set
inclusion if H = ;, i.e., the set of natural interpretations includes any set of interpretations
with a xed H. In fact, since H represents a constraint in the assignment of variables, the
unconstrained set is the larger one. Note that, for feature interpretation only the natural
one is used since it is not admitted to temporally qualify them.
475
Artale & Franconi
]
Boil-Spaghetti
-
x
Make-Spaghetti
- ]
Boil
-
Figure 8: Temporal dependencies in the de nition of the Boil-Spaghetti plan.
An interpretation I satis es the terminological axiom A =: C i AIt = CtI , for every t.

A concept C is subsumed by a concept D (C v D) if CtI DtI for every interpretation I
and every interval t. An interpretation I is a model for a concept C if CtI 6= ; for some t.
If a concept has a model, then it is satis able, otherwise it is unsatis able.
Each TL-F concept expression is always satis able, with the proviso that the temporal
constraints introduced by the existential quanti ers are consistent. This latter condition
can be easily checked during the reduction of the concept into a normal form when the
minimal temporal network (see Section 11, de nition 6.5) is computed.
It is interesting to note that only the relations s, f, mi are really necessary, because it is
possible to express any temporal relationship between two distinct intervals using only these
three relations and their transpositions si, , m (Halpern & Shoham, 1991). The following
equivalences hold:
3x (x a ]). C @x 3xy (y mi ])(x mi y). C @x
3x (x d ]). C @x 3xy (y s ])(x f y). C @x
3x (x o ]). C @x 3xy (y s ])(x y). C @x
To assign a meaning to ABox axioms, the temporal interpretation function E is extended
to temporal intervals so that iE is an element of T<? for each i 2 OT . The semantics of
assertions is the following: C (i; a) is satis ed by an interpretation I i aI 2 CiIE ; p(i; a; b)
is satis ed by I i pIiE (aI ) = bI ; ?g(a; b) is satis ed by I i ?gI (aI ) = bI ; and R(i; j ) is
satis ed by I i hiE ; j E i 2 RE . Given a knowledge base , an individual a in O is said to
be an instance of a concept C at the interval i if j= C (i; a).
Now we are able to give a semantic de nition for the reasoning task we already called
speci c plan recognition with respect to a plan description. This is an inference service that
computes if an individual action/plan is an instance of an action/plan type at a certain
interval, i.e., the task known as instance recognition in the Description Logic community.
Given a knowledge base , an interval i, an individual a and a concept C , the instance
recognition problem is to test whether j= C (i; a).
5. Action and plan representation: two examples

An action description represents how the world state may evolve in relation with the possible
occurrence of the action itself. A plan is a complex action: it is described by means of
temporally related world states and simpler actions. The following introduces examples of
action and plan representations from two well known domains, the cooking domain (Kautz,
476
z
Make-Marinara
- w -
Put-Together-SM
y -
Boil-Spaghetti
x - y
Make-Spaghetti Boil
-
Figure 9: Temporal dependencies in the de nition of Assemble-Spaghetti-Marinara.
1991; Weida & Litman, 1992) and the block world (Allen, 1991), with the aim of showing
the applicability of our framework.
5.1 The Cooking Domain

Let us introduce the plan Boil-Spaghetti:
:
Boil-Spaghetti = 3x (x b ]). (Make-Spaghetti@x u Boil)
Figure 8 shows the temporal dependencies of the intervals in which the concept Boil-Spa-
ghetti holds. The de nition employs the ] interval to denote the occurrence time of the
plan itself; in this way, it is possible to describe how di erent actions or states of the world
concurring to the de nition of the plan are related to it. This is why the variable ] is
explicitly present in the de nition of Boil-Spaghetti, instead of a generic variable: the
Boil action should take place at the same time of the plan itself, while Make-Spaghetti
occurs before it.
The de nition of a plan can be reused within the de nition of other plans by exploiting
the full compositionality of the language. The plan de ned above Boil-Spaghetti is used
in the de nition of Assemble-Spaghetti-Marinara:
:
Assemble-Spaghetti-Marinara = 3(y z w) (y b w)(z b w).
(Boil-Spaghetti@y u
Make-Marinara@z u
Put-Together-SM@w)
In this case, precise temporal relations between the nodes of two corresponding temporal
constraint networks are asserted: e.g., the action Put-Together-SM takes place strictly after
the Boil action (Figure 9). Observe that the occurrence interval of the plan Assemble-Spa-
ghetti-Marinara does not appear in the Figure because it is not temporally related with
any other interval.
A plan subsuming Assemble-Spaghetti-Marinara is the more general plan de ned be-
low, Prepare-Spaghetti, supposing that the action Make-Sauce subsumes Make-Marinara.
This means that among all the individual actions of the type Prepare-Spaghetti there are
all the individual actions of type Assemble-Spaghetti-Marinara:
:
Prepare-Spaghetti = 3 (y z ) (). (Boil-Spaghetti@y u Make-Sauce@z )
477
Artale & Franconi
]
Stack(OBJ1, OBJ2) -
- w
Clear-Block(OBJ1)
v
- z
Holding-Block(OBJ1) Clear-Block(OBJ1)
Clear-Block(OBJ2)
x
- y
ON(OBJ1, OBJ2)
Figure 10: Temporal dependencies in the de nition of the Stack action.
However, note that Boil-Spaghetti does not subsume Prepare-Spaghetti, even if it is a

conjunct in the de nition of the latter. This could be better explained observing how the
de nition of Prepare-Spaghetti plan is expanded:
:
Prepare-Spaghetti = 3 (x y z ) (x b y ). (Make-Spaghetti@x u Boil@y u
Make-Sauce@z )
Then, the Boil action occurs at the interval y { which can be di erent from the occurring
time of Prepare-Spaghetti { as the e ect of binding Boil-Spaghetti to the temporal vari-
able y. On the contrary, in the de nition of Boil-Spaghetti the Boil action takes place nec-
essarily at the same time. Subsumption between Prepare-Spaghetti and Boil-Spaghetti
fails since di erent temporal relations between the actions describing the two plans and the
plans themselves are speci ed. In particular, observe that the Boil-Spaghetti plan denotes
a narrower class than the plan expression
3(x y) (x b y). (Make-Spaghetti@x u Boil@y)
which subsumes both Prepare-Spaghetti and Boil-Spaghetti itself.
5.2 The Blocks World Domain
As a further example of the expressive power of the temporal language, it is now shown
how to represent the Stack action in the blocks world, in a more detailed way than the
previous simple Basic-Stack action used as a clarifying example. Thus a stacking action
involves two blocks, which should be both clear at the beginning; the central part of the
action consists of grasping one block; at the end, the blocks are one on top of another, and
the bottom one is no longer clear (Figure 10).
Our representation borrows from the Rat Description Logic (Heinsohn, Kudenko, Nebel,
& Pro tlich, 1992) the intuition of representing action parameters by means of partial
functions mapping from the action itself to the involved action parameter (see Section 9). In
the language, these functions are called parametric features. For example, the action Stack
has the parameters ?OBJECT1 and ?OBJECT2, representing in some sense the objects that are
involved in the action independently from time. So, in the assertion \?OBJECT1(a; block-a)",
block-a denotes the rst object involved in the action a at any interval. On the other hand,
an assertion involving a (non-parametric) feature, e.g., \ON(i; block-a; block-b)", does not
imply anything about the truth value at intervals other than i.
The concept expression, which de nes the Stack action, makes use of temporal quali ed
concept expressions, including feature selections and agreements: the expression (?OBJECT2 :
Clear-Block)@x means that the second parameter of the action should be a Clear-Block
478
at the interval denoted by x; while (?OBJECT1 ON # ?OBJECT2)@y indicates that at the

interval y the object on which ?OBJECT1 is placed is ?OBJECT2. The formal de nition of the
action Stack is:
:
Stack = 3(x y z v w) (x ])(y mi ])(z mi ])(v o ])(w f ])(w mi v ).
((?OBJECT2 : Clear-Block)@x u (?OBJECT1 ON # ?OBJECT2)@y u
(?OBJECT1 : Clear-Block)@v u (?OBJECT1 : Holding-Block)@w u
(?OBJECT1 : Clear-Block)@z )
The above de ned concept does not state which properties are the prerequisites for the
stacking action or which properties must be true whenever the action succeeds. What this
action intuitively states is that ?OBJECT1 will be on ?OBJECT2 in a situation where both
objects are clear at the start of the action. Note that the world states described at the
intervals denoted by v; w; z are the result of an action of grasping a previously clear block:
:
Grasp = 3(x w z ) (x o ])(w f ])(w mi x)(z mi ]).
((?OBJECT1 : Clear-Block)@x u (?OBJECT1 : Holding-Block)@w u
(?OBJECT1 : Clear-Block)@z )
The Stack action can be rede ned by making use of the Grasp action:
:
Stack = 3(x y u v ) (x ])(y mi ])(u f ])(v o ]).
((?OBJECT2 : Clear-Block)@x u (?OBJECT1 ON # ?OBJECT2)@y u
(Grasp[x]@v)@u)
The temporal substitutive quali er (Grasp[x]@v) renames within the de ned Grasp action
the variable x to v and it is a way of making coreference between two temporal variables,
while the temporal constraints peculiar to the renamed variable x are inherited by the
substituting interval v. Furthermore, the e ect of temporally qualifying the grasping action
at u is that the ] variable associated to the grasping action { referring to the occurrence
time of the action itself { is bound to the interval denoted by u. Because of this binding on
the occurrence time of the grasping action, the ] variable in the grasping action and the ]
variable in the stacking action denote di erent time intervals, so that the grasping action
occurs at an interval nishing the occurrence time of the stacking action.
Now it is shown how from a series of outside observations action recognition can be
performed { i.e., the task called speci c plan recognition with respect to a plan description.
The following ABox describes a situation in which blocks can be clear, grasped and/or on
each other, and in which a generic individual action a is taking place at time interval ia
having the blocks block-a and block-b as its parameters:
?OBJECT1(a; block-a); ?OBJECT2(a; block-b);
o(i1 ; ia ); Clear-Block(i1 ; block-a); (i2 ; ia ); Clear-Block(i2 ; block-b);
mi(i3 ; i1 ); f (i3 ; ia ); Holding-Block(i3 ; block-a);
mi(i4 ; ia ); Clear-Block(i4 ; block-a); mi(i5 ; ia ); ON(i5 ; block-a; block-b)
The system deduces that, in the context of a knowledge base composed by the above
ABox and the de nition of the Stack concept in the TBox, the individual action a is of
type Stack at the time interval ia , i.e., j= Stack(ia ; a).
479
Artale & Franconi
C @X u D@X ! (C u D)@X
(C @X1 )@X2 ! C @X1
(C @X1 u D)@X2 ! C @X1 u D@X2
C u 3(X ) Tc. D ! 3(X ) Tc. (C u D)
if C doesn't contain free variables
3(X )Tc1 .(C u

3(Y ) Tc2 . D [Y1 ]@X1 : : : [Yp ]@Xq @X ) ! 3(X ][Y1 =X1 ]:::[Yp=Xq ] Y )Tc1 [ Tc2+[]=X] .(C u D+ @X )
if D doesn't contain existential temporal quanti ers
p : (q : C ) ! (p q) : C
p : (C u D) ! p : C u p : D
p : (q1 # q2 ) ! p q1 # p q2
Prescriptions: X ][Y1 =X1 ]:::[Yp=Xq ] Y returns the union of the two sets of variables X and Y , where each
occurrence of Y1 ; : : : ; Yp is substituted by X1 ; : : : ; Xq , respectively, while all the other elements of Y occurring
in X are renamed with fresh new identi ers. Z+ is intended to be the expression Z where the same
substitution or renaming has taken place. The condition on the last rule forces application to start from the
last nested existential temporal quali ed concept.
Figure 11: Rewrite rules to transform an arbitrary concept into an existential concept.
6. The Calculus for TL-F

This Section presents a calculus for deciding subsumption between temporal concepts in the
Description Logic TL-F . The calculus is based on the idea of separating the inference on
the temporal part from the inference on the Description Logic part. This is achieved by rst
looking for a normal form of concepts. Concept subsumption in the temporal language is
then reduced to concept subsumption between non-temporal concepts and to subsumption
between temporal constraint networks.
6.1 Normal Form
Every TL-F concept expression can be reduced to an equivalent existential concept of the
form: 3(X ) Tc. (Q0 u Q1 @X1 u : : : u Qn@Xn ), where each Q is a non-temporal concept, i.e., it
is an element of the language F . A concept in existential form can be seen as a conceptual
temporal constraint network, i.e., a labeled directed graph hX; Tc; Q@X i where arcs are
labeled with a set of arbitrary temporal relationships { representing their disjunction {
nodes are labeled with non-temporal concepts and, for each node X , the temporal relation
(X = X ) is implicitly true. Moreover, since the normalized concepts do not contain free
variables or substitutive quali ers, in the following the natural interpretation function (see
Section 4.2) is used.
Proposition 6.1 (Equivalence of EF) Every concept C can be reduced in linear time
into an equivalent existential concept (ef C ), by exhaustively applying the set of rewrite
rules of Figure 11.
480
Procedure Covering( hX; Tci; y):

mid ;;
result ;;
Z = fz 2 X j (z (=; : : :) y) 2 Tcg;
8s 2 }(Z ) do
if j s j 2 and the graph hX; Tc i obtained by deleting the \=" temporal relation between

the node y and each of the nodes in s is inconsistent

then mid mid [ fsg;
8s 2 mid do
if :9t 2 mid. t s
then result result [ fsg;
return result.
Figure 12: Procedure which computes a covering.

Note that (ef C ) makes explicit all the possible chains of features by reducing each non-
temporal concept Q to a conjunction of atomic concepts, feature selections restricted to
atomic concepts and feature agreements { i.e., each Q is a feature term expression (Smolka,
1992).
The normalization proceeds by discovering all the possibles interactions between nodes
with the intention of making explicit all the implicit information. A crucial temporal in-
teraction occurs when a node is always coincident with a set of nodes in every possible
interpretation of the temporal network.
De nition 6.2 (Covering) Given a temporal constraint network hX; Tci, let y 2 X and
Z = fz1 ; z2 ; : : : ; zp g X , with p 1, and y 62 Z . Z is a Covering for y if 8V 2 hX; TciE ,
V (y) 2 fV (z1 ); V (z2 ); : : : ; V (zp)g and for each W Z , W is not a covering for y. If Z = ;,
then y is called uncovered, otherwise y is said covered by Z .
Proposition 6.3 (Covering procedure) Given a temporal constraint network hX; Tci in
minimal form (see, e.g., (van Beek & Manchak, 1996)) and a node y 2 X then the procedure
described in Figure 12 returns all the possible coverings for y with size 2.
The idea behind the covering is that whenever a set of nodes fz1 ; z2 ; : : : ; zp g is a covering
for y the disjunctive concept expression (Qz1 t : : : t Qzp ) should be conjunctively added to
the concept expression Qy . Actually, since in TL-F concept disjunction is not allowed it will
be sucient to add to the node y the Least Commom Subsumer (lcs) of (Qz1 t : : : t Qzp )
as de ned below.
De nition 6.4 (lcs) Let Q1; : : : ; Qn; Q; C be F concept expressions. Then, the concept
Q = lcsfQ1 ; : : : ; Qn g is such that: Q1 v Q ^ : : : ^ Qn v Q and there is no C such that
Q1 v C ^ : : : ^ Qn v C ^ C < Q.
Given a concept in existential form, the temporal completion of the constraint network is
computed as described below.
De nition 6.5 (Completed existential form) The temporal completion of a concept in
existential form { the Completed Existential Form, CEF { is obtained by sequentially ap-
plying the following steps:
481
Artale & Franconi
(closure) The transitive closure of the Allen temporal relations in the conceptual
temporal constraint network is computed, obtaining a minimal temporal network (see,
e.g., (van Beek & Manchak, 1996)).
(= collapsing) For each equality temporal constraint, collapse the equal nodes by
applying the following rewrite(rule:
3(X ) Tc (xi = xj ). Q ! 3 3(X n fxj g) Tc[xj =xi]. Q[xj =xi ] if xi 6= xj and xj 6= ].
(X n fxi g) Tc[xi=]] . Q[xi =]] if xi 6= xj and xj = ].
Then apply exhaustively the rst rule of Figure 11.
(covering) For each y 2 X let compute the covering = fZ 1; : : : ; Z ng following the
procedure showed by proposition 6.3. Whenever the covering is not empty, translate
Qy applying the following rewrite rule: Qy ! Qy ui=1:::nlcsfQi1 ; : : : ; Qim g where Z i =
fzi1 ; : : : ; zim g, and Qij @zij 2 hX; Tc; Q@X i.
(parameter introduction) New information is added to each node because of the pres-
ence of parameters, as the following rules show. The ; symbol is intended so that,
each time the concept expression in the left hand side appears in some node of the
temporal constraint network, possibly conjoined with other concepts, then the right
hand side represents the concept expression that must be conjunctively added to all the
other nodes; square brackets point out optional parts; the letters f (?f ) and g (?g),
possibly with subscripts, denote atomic (parametric) features while p and q stand for
generic features.
?g1 : : : ?gn [ f [ p]] : C ; ?g1 : : : ?gn : >.
?g1 : : : ?gn [ f [ p]] # g [ q] ; ?g1 : : : ?gn : >.
?g1 : : : ?gn # ?f1 : : : ?fm ; ?g1 : : : ?gn # ?f1 : : : ?fm .
?g1 : : : ?gn g [ p] # ?f1 : : : ?fm [ f [ q]] ; ?g1 : : : ?gn : > u
?f1 : : : ?fm : >.
Proposition 6.6 (Equivalence of CEF) Every concept in existential form can be re-
duced into an equivalent completed existential concept.
Both the covering and the parameter introduction steps can be computed independently
after the =-collapsing step and then conjoining the resulting concept expressions. Observe
that, to obtain a completed existential concept, the steps of the normalization procedure
require linear time with the exception of the computation of the transitive closure of the
temporal relations, and the covering step. Both these steps involve NP-complete temporal
constraint problems (van Beek & Cohen, 1990). However, it is possible to devise reasonable
subsets of Allen's algebra for which the problem is polynomial (Renz & Nebel, 1997). The
most relevant properties of a concept in CEF is that all the admissible interval temporal
relations are explicit and the concept expression in each node is no more re nable without
changing the overall concept meaning; this is stated by the following proposition.
Proposition 6.7 (Node independence of CEF) Let hX; Tc; Q@X i be a conceptual tem-
poral constraint network in its completed form (CEF); then, for all Q 2 Q and for all
482
F concept expressions C such that C 6w Q, there exists an interpretation I such that

hX; Tc; (Q u C )@X iIt 6= hX; Tc; Q@X iIt , for some interval t.
Proof. The proposition states that the information in each node of the CEF is independent
from the information in the other nodes. In fact, hX; Tc; (Q u C )@X iIt = hX; Tc; Q@X iIt if
the concept expression in one node implies new information in some other node. Two cases
can be distinguished.
i) Covered Nodes. Both the (= collapsing) rule and the (covering) rule provide to restrict a
covered node with the most speci c F concept expression. Indeed, the (= collapsing) rule
provides collapsing two contemporary nodes conjoining the concept expressions of each of
them. On the other hand, the (covering) rule adds to the covered node the most speci c
F concept expression that subsumes the disjunctive concept expression that is implicitly
true at the covered node. Note that, thanks to the (Closure) rule, all the possible equal
temporal relations are made explicit. So these two normalization rules cover all the possible
cases of temporal interactions between nodes.
ii) No coincident nodes. Every time-invariant information should spread over all the nodes.
Both parametric features and the > concept have a time-invariant semantics: the only time-
invariant concept expressions are >, ?g1 : : : ?gn : >, ?g1 : : : ?gn # ?f1 : : : ?fm , with
n; m 1, or an arbitrary conjunction of these terms. The (parameter introduction) rule
captures all the possible syntactical cases of completion concerning time-invariant concept
expressions. By induction on the syntax, it can be proven that adding to a node any other
concept expression changes the overall interpretation. 2
The last normalization procedure eliminates nodes with redundant information. This
nal normalization step ends up with the concept in the essential graph form, that will be
the normal form used for checking concept subsumption.
De nition 6.8 (Essential graph) The subgraph of the CEF of a conceptual temporal con-
straint network T = hX; Tc; Q@X i obtained by deleting the nodes labeled only with time-
invariant concept expressions { with the exception of the ] node { is called essential graph
of T : (ess T ).
Proposition 6.9 (Equivalence of essential graph) Every concept in completed existen-
tial form can be reduced in linear time into an equivalent essential graph form.
Theorem 6.10 (Equivalence of normal form) Every concept expression can be reduced
into an equivalent essential graph form. If a polynomial fragment of Allen's algebra is
adopted, the reduction takes polynomial time.
As an example, the normal form is shown { i.e., the essential graph { of the previously
introduced Stack action (see Section 5.2):
:
Stack = 3(x y v w z )(x ])(y mi ])(z mi ])(w f ])(v o ])(y mi x)(z mi x)(w f x)
(v (o; d; s) x)(z (=; s; si) y)(w m y)(v b y)(w m z )(v b z )(w mi v).
((?OBJECT2 : Clear-Block u ?OBJECT1 : >)@x u
(?OBJECT1 ON # ?OBJECT2)@y u
(?OBJECT1 : Clear-Block u ?OBJECT2 : >)@v u
(?OBJECT1 : Hold-Block u ?OBJECT2 : >)@w u
(?OBJECT1 : Clear-Block u ?OBJECT2 : >)@z )
483
Artale & Franconi
In this example, the essential graph is also the CEF of Stack since there are no redundant
nodes.
6.2 Computing Subsumption
A concept subsumes another one just in case every possible instance of the second is also an
instance of the rst, for every time interval. Thanks to the normal form, concept subsump-
tion in the temporal language is reduced to concept subsumption between non-temporal
concepts and to subsumption between temporal constraint networks. A similar general pro-
cedure was rst presented in (Weida & Litman, 1992), where the language for non-temporal
concepts is less expressive { it does not include features or parametric features.
To compute subsumption between non-temporal concepts { which may possibly include
lcs concepts { we refer to (Cohen, Borgida, & Hirsh, 1992). In the following, we will write
\wF " for subsumption between non-temporal F concepts taking into account lcs concepts.
De nition 6.11 (Variable mapping) A variable mapping M is a total function M :
X 1 7! X 2 such that M(]) = ]. We write M(X ) to intend fM(X ) j X 2 X g, and M(Tc)
to intend f(M(X ) (R) M(Y )) j (X (R) Y ) 2 Tcg.
De nition 6.12 (Temporal constraint subsumption) A temporal constraint (X1 (R1 )Y1)
is said to subsume a temporal constraint (X2 (R2 )Y2 ) under a generic variable mapping M,
written (X1 (R1 )Y1 ) wM (X2 (R2 ) Y2 ), if M(X1 ) = X2 , M(Y1 ) = Y2 and (R1 )E (R2 )E
for every temporal interpretation E .
Proposition 6.13 (TC subsumption algorithm) (X1 (R1 )Y1) wM (X2 (R2 )Y2) if and
only if M(X1 ) = X2 , M(Y1 ) = Y2 and the disjuncts in R1 are a superset of the disjuncts
in R2 .
Proof. Follows from the observation that the 13 temporal relations are mutually disjoint
and their union covers the whole interval pairs space. 2
De nition 6.14 (Temporal constraint network subsumption) A temporal constraint
network hX 1 ; Tc1 i subsumes a temporal constraint network hX 2 ; Tc2 i under a variable map-
ping M : X 1 7! X 2 , written hX 1 ; Tc1 i wM hX 2 ; Tc2 i, if hM(X 1 ); M(Tc1 )iE hX 2 ; Tc2 iE
for every temporal interpretation E .
Proposition 6.15 (TCN subsumption algorithm) hX 1; Tc1 i wM hX 2 ; Tc2i i , after
computing the temporal transitive closure, there exists a variable mapping M : X 1 7! X 2
such that for all X1i ; Y1j 2 X 1 exist X2m ; Y2n 2 X 2 which satisfy (X1i (R1i;j ) Y1j ) wM
(X2m (R2m;n ) Y2n ).
Proof. $ " Since from de nition 6.12 (X1i (R1i;j ) Y1j ) wM (X2m (R2m;n ) Y2n ) implies that
(R1i;j )E (R2m;n )E for every E , then, from the de nition of interpretation of a temporal
constraint network, it is easy to see that each assignment of variables V in the interpretation
of hX 2 ; Tc2 i is also an assignment in the interpretation of hM(X 1 ); M(Tc1 )i.
$ " Suppose that one is not able to nd such a mapping; then, by hypothesis, for each
possible variable mapping there exists some i; j such that R1i;j is not a superset of R2m;n .
484
Since, by assumption, the temporal constraint networks are minimal, the temporal relation
R2m;n cannot be further restricted. So, for each variable mapping and each temporal inter-
pretation E , we can build an assignment V such that hV (X2m ); V (X2n )i 2 (R2m;n )E while
hV (X1i ); V (X1j )i 62 (R1i;j )E . Now, we can extend the assignment V in such a way that
V 2 (hX 2; Tc2 i)E while V 62 (hM(X 1 ); M(Tc1)i)E . This contradicts the assumption that
hX 1; Tc1 i wM hX 2; Tc2 i. 2
De nition 6.16 (S-mapping) An s-mapping from a conceptual temporal constraint net-
work hX 1 ; Tc1 ; Q@X 1 i to a conceptual temporal constraint network hX 2 ; Tc2 ; Q@X 2 i is a
variable mapping S : X 1 7! X 2 such that the non-temporal concept labeling each node
in X 1 subsumes the non-temporal concept labeling the corresponding node in S (X 1 ), and
hX 1; Tc1 i wS hX 2; Tc2i.
The algorithm for checking subsumption between temporal concept expressions reduces the
subsumer and the subsumee in essential graph form, then it looks for an s-mapping be-
tween the essential graphs by exhaustive search. To prove the completeness of the overall
subsumption procedure it will be showed that the introduction of lcs's preserves the sub-
sumption. A model-theoretic characterization of the lcs will be given for showing this
property. Let's start to build an Herbrand model for an F concept. Let C 0(x) denote the
rst order formula corresponding to a concept C (see proposition 2.1), while the function-
ality of features can be expressed with a set of formul F . By syntax induction it easy to
show that C 0 (x) is an existentially quanti ed formula with one free variable. Moreover, the
matrices of such formula is a conjunction of positive predicates. F [ fC 0 (x)g is logically
equivalent to F [ fC 00 (x)g where the functionality axioms allow to map every subformula
y 9y.Ff (x; y) into 9!y.Ff (x; y). Then C (x) is such that all the existential quanti ers in
V 00
C 0 (x) (which come from the rst order conversion of features) are replaced by 9! quanti ers.
Now, F [ fC 000 (a)g { where a is a constant substituting the free variable x and C 000 (a) is
obtained by skolemizing the 9! quanti ed variables { is a set of de nite Horn clauses.
De nition 6.17 (Herbrand model) Let C be an F concept expression. Then we de ne
its Minimal Herbrand Model HC as the Minimal Herbrand Model of the above mentioned
set of de nite Horn clauses F [ fC 000 (a)g.
Lemma 6.18 (F concept subsumption) Let C; D be F concept expressions, and HC ; HD
their minimal Herbrand models obtained by skolemizing the rst order set F [fC 000 (a); D000 (a)g.
Then, C v D i HD HC .
Proof. C v D i F [ fC 0 (x)g j= D0 (x), i F [ fC 00 (x)g j= D00 (x), where C 00 and D00
are obtained by applying the functionality axioms to the set fC 0 (x); D0 (x)g (i.e., uni-
fying the variables in the functional predicates) and then replacing all the existential
quanti ers by 9! quanti ers. Now, C 000 (x) and D000 (x) are obtained by skolemizing the
9! quanti ed variables in the following way: let C 00(x) = 9!y1; : : : ; yn(x; y1 ; : : : ; yn) and let
D00(x) = 9!y1 ; : : : ; yk ; z1 ; : : : ; zm (x; y1 ; : : : ; yk ; z1 ; : : : ; zm ), with 0 k n, then skolemize
the formula: = 9!y1 ; : : : ; yn ; z1 ; : : : ; zm (x; y1 ; : : : ; yn ) ^ (x; y1 ; : : : ; yk ; z1 ; : : : ; zm ), and
let 0 (x) indicate its skolemized form. Then, C 000 (x) = 0 (x) and D000 (x) = 0 (x). Now,
since every existential quanti cation in C 00 (x); D00 (x) was of type 89! then the thesis is true
485
Artale & Franconi
i F [ fC 000 (a)g j= D000 (a), where a is a constant substituting the free variable x (see (van
Dalen, 1994)). Now, as showed by lemma 6.17, both C 000 (a) and D000 (a) have minimal Her-
brand models HC ; HD that verify the lemma hypothesis. Then, F [ fC 000 (a)g j= D000 (a) i
HD HC . 2
We are now able to give a model-theoretic characterization of the lcs that will be crucial
to prove the subsumption-preserving property.
Lemma 6.19 (lcs model property) Let Q1; : : : ; Qn be F concept expressions, and HQ1 ;
: : : ; HQn their minimal Herbrand models obtained by skolemizing the rst order set F [
fQ0001 (a); : : : ; Q000n (a)g. Then, Q = lcsfQ1 ; : : : ; Qn g i HQ = HQ1 \ : : : \ HQn .
Proof. First of all, let show that HQ is the minimal Herbrand model of a concept Q in
the language F . Every HQi can be seen as a rooted directed acyclic graph where nodes
are labelled with (possible empty) set of atomic concepts and arcs with atomic features
while equality constraints between nodes correspond to features agreement. Whithout loss
of generality let us consider the case where HQ = HQ1 \ HQ2 . It is sucient to show
that HQ is a rooted directed acyclic graph. Let a be the root of HQ1 ; HQ2 , then will be
proved by induction that if Fi (ai 1 ; ai ) 2 HQ (where Fi is the rst order translation of a
feature, ai 1 ; ai are obtained as a result of the skolemization process, and a0 = a) then
fF1 (a; a1 ); : : : ; Fi (ai 1; ai )g HQ. The case i = 1 is trivial. Let i > 1. Now, Fi (ai 1 ; ai) 2
HQ i Fi(ai 1 ; ai ) 2 HQ1 \ HQ2 . But ai 1 is uniquely de ned by the skolem function
fFi 1 (where, the function symbols fFi are newly generated for each feature Fi by the
skolemization process). Then, Fi (ai 1 ; ai ) 2 HQ1 \ HQ2 i Fi (ai 1 ; fFi 1 (ai )) 2 HQ1 \ HQ2
i Fi 1 (ai 2 ; fFi 1 (ai )) 2 HQ1 \ HQ2 . Then the thesis is true by induction.
Let us now prove the $" direction. Suppose by absurd that there is an F concept C
such that: Q1 v C ^ Q2 v C ^ C < Q. Then, Q1 v C i HC HQ1 , and Q2 v C , i
HC HQ2 . But then HC HQ1 \ HQ2 , i.e., HC HQ. Then Q v C which contradicts the
hypothesis.
The $" direction can be proved with analogous considerations. 2
Proposition 6.20 (lcs subsumption-preserving property) Let A; B; C; D be F con-
cepts, then A u (B t C ) v D i A u lcsfB; C g v D.
Proof. A u (B t C ) v D i A u B v D and A u C v D. Now, A u B v D i F [
fA000 (a); B 000 (a)g j= D000(a) i HA [ HB j= D000 (a) i HD HA [ HB . For the same reasons,
A u C v D i HD HA [ HC . But then, HD HA [ HB and HD HA [ HC , i.e.,
HD HA [ (HB \ HC ), i.e., HD HA [ HlcsfB;C g . But, HD HA [ HlcsfB;C g i
A u lcsfB; C g v D. 2
The following theorem provides a sound and complete procedure to compute subsump-
tion. The completeness proof takes into account that the temporal structure is dense and
unbounded. This allows us to introduce any new node to a conceptual temporal constraint
network without changing its meaning. Remember that, for each of these redundant nodes,
time-invariant information holds.
Theorem 6.21 (TL-F concept subsumption) A concept C1 subsumes a concept C2 i
there exists an s-mapping from the essential graph of C1 to the essential graph of C2 .
486
Proof. Let T1 = hX 1 ; Tc1 ; Q@X 1 i be the essential graph of C1 , and T2 = hX 2 ; Tc2 ; Q@X 2 i
be the essential graph of C2 .
$ " (Soundness). Follows from the fact that the essential graph form is logically equiv-
alent to the starting concept, and from the soundness of the procedures for computing
both the TCN subsumption (proposition 6.15) and the subsumption between non-temporal
concepts (Cohen et al., 1992).
$ " (Completeness). Suppose that such an s-mapping does not exist. Two main cases
can be distinguished.
i) There is not a mapping M such that hX 1 ; Tc1 i wM hX 2 ; Tc2 i. By adding redundant
nodes to T2 , an equivalent conceptual temporal constraint network T2 = hX 2 ; Tc2 ; Q@X 2 i
may be obtained. Let us consider such an extended network
in a way that there exists a

variable mapping M such that hX 1 ; Tc1 i wM hX 2 ; Tc2 i. Now, for all possible M , there
is a node X1i 2 X 1 such that M (X1i ) = X2j with X2j 62 X 2 . Now, Q1i 6wF Q2j , since X2j
cannot coincide with other nodes in X 2 neither can have a covering otherwise the hypothesis
that the mapping M does not exist would be contradicted. Then from proposition 6.7 Q2j
is in a time-invariant node, whereas Q1i is not since T1 is an essential graph. Then, although
the construction of M allows for the existence of a unique V 3 for both networks (follows
from proposition 6.15), it is possible to build an instance of T2 that is not an instance of T1 .
ii) For each possible mapping M such that hX 1 ; Tc1 i wM hX 2 ; Tc2 i there will be always
two nodes X1i and X2j such that M(X1i ) = X2j and Q1i 6wF Q2j . Now, the concept ex-
pression Q2j cannot be re ned (looking for a subsumption relationship with Q1i ) by adding
to it an F concept since from proposition 6.7 this would change the overall interpretation.
On the other hand, the lcs introduction { which would substitute the more speci c con-
cept disjunction implicitly presents because of a node covering { is a subsumption-invariant
concept substitution, as showed by lemma 6.20.
Both cases contradict the assumption that T1 subsumes T2 . 2
6.2.1 Complexity of Subsumption
Now it is shown that checking subsumption between TL-F concept expressions in the es-
sential graph form is an NP-complete problem. Therefore, a polynomial reduction from the
NP-complete problem of deciding whether a graph contains an isomorphic subgraph is pre-
sented. It is then shown that the subsumption computation, as proposed in theorem 6.21,
can be done by a non-deterministic algorithm that takes polynomial time in the size of the
concepts involved. First of all let us consider the complexity of computing subsumption
between non-temporal concepts.
Lemma 6.22 (F subsumpion complexity) Let C; D be F concept expressions that can
contain lcs's. Then, checking whether C vF D takes polynomial time.
Proof. See (Cohen et al., 1992). 2
Here the problem of subgraph isomorphism is brie y recalled. Given two graphs, G1 =
(V1 ; E1 ) and G2 = (V2 ; E2 ), G1 contains a subgraph isomorphic to G2 if there exists a
3. Since subsumption is computed with respect to a xed evaluation time, V maps the di erent occurrences
of ] to the same interval; this justi es the choice that M(]) = ].
487
Artale & Franconi
subset of the vertices V 0 V1 and a subset of the edges E 0 E1 such that j V 0 j=j V2 j,
j E 0 j=j E2 j, and there exists a one-to-one function f : V2 7! V 0 satisfying fu; vg 2 E2 i
ff (u); f (v)g 2 E 0.
Given
: a graph G = (V; E ), with V = fv1 ; : : : ; vn g associate a temporal concept expression:
C = 3(v1 ; : : : ; vn ) : : : (vi (b; a) vj ) : : : . (A@v1 u : : : u A@vn ), where A is an atomic concept
and fvi ; vj g 2 E . This transformation allows us to prove that the problem of subgraph
isomorphism can be reduced to the subsumption of temporal concepts.
Proposition 6.23 Given two graphs G1 and G2, G1 contains a subgraph isomorphic to G2
i C2 w C1 , where C1 and C2 are the corresponding temporal concepts expressions.
Proof. A temporal network with edges labeled only with the (before _ after) relation is always
consistent, minimal and non-directed4 (Gerevini & Schubert, 1994). Then, each temporal
concept is in the essential graph form. Now the proof easily follows since, every time G2 is
an isomorphic subgraph of G1 the one-to-one function f is also an s-mapping from C2 to
C1 , and it is true that C2 w C1 . On the other hand, the s-mapping that gives rise to the
subsumption is also the one-to-one isomorphism from G2 to G1 . 2
Theorem 6.24 (NP-hardness) Concept subsumption between TL-F concept expressions
in normal form is an NP-hard problem.
Proof. Follows from proposition 6.23 and the reduction being clearly polynomial. 2
Now the NP-completeness is proven.
Theorem 6.25 (NP-completeness) Concept subsumption between TL-F concept expres-
sions in normal form is an NP-complete problem.
Proof. To prove NP-completeness it is necessary to show that the proposed calculus can
be solved by a nondeterministic algorithm that takes polynomial time. Now, given two
temporal concepts, T1 and T2 , in their essential graph form, let j X 1 j= N1 and j X 2 j= N2 .
Then, to check whether T1 w T2 , the algorithm guesses one of the N2N1 variable mapping
from T1 to T2 and veri es whether it is an s-mapping, too. This last step can be done in
deterministic polynomial time since, given a mapping M, it is possible to determine whether
hX 1; Tc1 i wM hX 2; Tc2i by checking at most N1(N1 1)=2 edges looking for subsumption
between the corresponding temporal relations (solved by a set inclusion procedure); while
the N1 non-temporal concept subsumptions can be computed in polynomial time. 2
7. Extending the Propositional Part of the Language

The propositional part of the temporal language can be extended to have a more powerful,
but still decidable, Description Logic. It is possible either to add full disjunction, both at
the temporal and non-temporal levels (TLU -FU ), or to have a propositionally complete
language at the non-temporal level only (TL-ALCF ).
Please note that in these languages it is not possible to express full negation, and
in particular the negation of the existential temporal quanti er. This is crucial, and it
4. If (vi (b; a) vj ) then (vj (b; a) vi ), too.
488
(C t D)@X ! C @X t D@X
p : (C t D ) ! p:Ctp:D
(C1 t C2 ) u D ! (C1 u D) t (C2 u D)
3(X ) Tc. (C t D) ! 3(X ) Tc. C t 3(X ) Tc. D
Figure 13: Rewrite rules for computing the disjunctive form.
makes the di erence with other logic-based approaches (Schmiedel, 1990; Bettini, 1997;
Halpern & Shoham, 1991). The dual of 3 (i.e., the universal temporal quanti er 2) makes
the satis ability problem { and the subsumption { for propositionally complete languages
undecidable in the most interesting temporal structures (Halpern & Shoham, 1991; Venema,
1990; Bettini, 1993). For the representation of actions and plans in the context of plan
recognition, the universal temporal quanti er is not strictly necessary. This limitation makes
these languages decidable, with nice computational properties, and capable of supporting
other kinds of useful extensions. The examples shown throughout the paper may serve as a
partial validation of the claim. Section 8.1 proposes the introduction of a limited universal
temporal quanti cation that maintains decidability of subsumption.
7.1 Disjunctive Concepts: TLU -FU
The language TLU -FU adds to the basic language TL-F the disjunction operator { with
the usual semantics { both at the temporal and non-temporal levels:
C; D ! TL j C t D (TLU )
E; F ! F j E t F (FU )
Before showing how to modify the calculus to check subsumption, let us begin with a
clarifying example. The gain in expressivity allows us to describe the alternative realizations
that a given plan may have. Let us consider a scenario with a robot moving in an empty
room that can move only either horizontally or vertically. Let's call Rect-Move that which
involves a simple sequence of the two basic moving actions. Then, to describe a Rect-Move
plan we can make use of the disjunction operator:
:
Rect-Move = 3(x y ) (] m x)(x m y ). (Hor-Move@x u Ver-Move@y ) t
3(x y) (] m x)(x m y). (Ver-Move@x u Hor-Move@y)
7.1.1 The Calculus for TLU -FU
Normal Form
In computing subsumption, a normal form for concepts is needed. The normalization pro-
cedure is similar to that reported in Section 6.1. Let us start by reducing each concept
expression into an equivalent disjunctive concept of the form:
(3(X 1 ) Tc1 . G1) t t (3(X n) Tcn. Gn) t Q1 t t Qm
489
Artale & Franconi
where Gi are conjunctions of concepts of the form Qik @Xik , and each Q does not contain
neither temporal information, nor disjunctions, i.e., it is an element of the language F .
Proposition 7.1 (Equivalence of disjunctive form) Every concept C can be reduced
into an equivalent disjunctive form (df C ), by exhaustively applying the set of rewrite rules
of Figure 13 in addition to the rules introduced in Figure 11.
It is now possible to compute the completed disjunctive normal form (cdnf C ). Each dis-
junct of such normal form has some interesting properties, which are crucial for the proof
of the theorem 7.4 on concept subsumption: temporal constraints are always explicit, i.e.,
any two intervals are related by a basic temporal relation; there is no disjunction, either
implicit or explicit, neither in the conceptual part nor in the temporal part, i.e., it is a
TL-F concept; the information in each node is independent of the information in the other
nodes and it does not contain time-invariant (i.e., redundant) nodes.
De nition 7.2 (Completed disjunctive normal form) Given a concept in disjunctive
form, the completed disjunctive normal form is obtained by applying the following rewrite
rules to each disjunct:
(Temporal completion) The rules of de nition 6.5 are applied to each disjunct with
the exclusion of the covering step, which is replaced by the t-introduction step. If a
disjunct is unsatis able { i.e., the temporal constraint network associated with it is
inconsistent { then eliminate it.
(Essential form) The rules of de nition 6.8 are applied to each disjunct.
(t introduction) Reduce to concepts containing only basic temporal relationships:
3(X ) (X1 (R,S ) X2 ) Tc.C ! 3(X )(X1 R X2 )Tc.C t 3(X )(X1 S X2 )Tc.C
Proposition 7.3 (Equivalence of CDNF) Every concept expression can be reduced into
an equivalent completed disjunctive normal concept.
Subsumption
The theorem 7.4 reduces subsumption between CDNF concepts into subsumption of disjun-
ction-free concepts, such that the results of theorem 6.21 can be applied. The following
theorem gives a terminating, sound, and complete subsumption calculus for TLU -FU .
Theorem 7.4 (TLU -FU concept subsumption) Let C = C1 t t Cm and D = D1 t
t Dn be TLU -FU concepts in CDNF. Then, C v D if and only if 8i9j . Ci v Dj .
Proof. Since it is easy to show that C1 t : : : t Cn v D i 8i.Ci v D we need only to prove the
restricted thesis: Ci v D1 t t Dn i Ci v D1 _ : : : _ Ci v Dn . Every concept expression in
CDNF corresponds to an existential quanti ed formula with two free variables. Moreover,
the matrices of such formul are conjunctions of positive predicates. Let us denote the
formula corresponding to a concept C as C 0 (t; x). Now, the restricted thesis holds i it is
true that F [ fCi000 (a; b)g j= D1000 (a; b) _ D2000 (a; b). Now, let HB the minimal Herbrand model
of F [fCi000 (a; b)g. Then, F [fCi000 (a; b)g j= D1000 (a; b) _ D2000 (a; b) i HB j= D1000 (a; b) _ D2000 (a; b).
Since we are talking of a single model, D1000 (a; b) _ D2000 (a; b) is valid in HB if and only if either
D1000(a; b) or D2000 (a; b) is valid in HB . This proves the theorem.5 2
5. The proof of this theorem comes from an idea of Werner Nutt.
490
As a consequence of the theorems 6.25, 7.4 the following complexity result holds.
Theorem 7.5 (TLU -FU subsumption complexity) Concept subsumption between TLU -FU
concept expressions in normal form is an NP-complete problem.
7.2 A Propositionally Complete Language: TL-ALCF
TL-ALCF uses the propositionally complete Description Logic ALCF (Hollunder & Nutt,
1990) for non-temporal concepts by changing the syntax rules for TL-F in the following
way:
E; F ! FU j ? j :E j p " q j p "j 8P .E j 9P .E (ALCF )
The interpretation functions are extended to take into account roles:
P I T<? I I
PtI = P^t I I j 8a; b. ha; bi 2 P^t $ ht; a; bi 2 P I
As seen in Section 2, ALCF adds to F full negation { thus introducing disagreement (p " q)
and unde nedness (p ") for features, and role quanti cation (8P .E; 9P .E ).
As an example of the expressive power gained, let us re ne the description of the world
states involved in the Stack action (see Section 5.2). Suppose that a block is described by
saying that it has LATERAL-SIDEs (role) and BOTTOM- and TOP-SIDEs (features). Then, the
property of being clear could be represented as follows:
:
Clear-Block = Block u 8LATERAL-SIDE.Clear u TOP-SIDE : HAS-ABOVE "
which says that, in order to be clear, each LATERAL-SIDE has to be clear and nothing has
to be over the TOP-SIDE. Now, the situation in which a block involved in a Stack action is
on top of another one is reformulated with the following concept expression:
(?OBJECT1 TOP-SIDE HAS-ABOVE # ?OBJECT2)
Furthermore, given the above de nition of Clear-Blocks, it can be derived that:
(?OBJECT1 TOP-SIDE HAS-ABOVE # ?OBJECT2) v (?OBJECT1 : :Clear-Block)
i.e., an object, having another object on top of it, is no more a clear object.
In TL-ALCF it is possible to describe states with some form of incomplete knowledge
by exploiting the disjunction among non-temporal concepts. For example, let us say that
the agent of an action can be either a human being or a machine: ?AGENT:(Person t Robot).
7.2.1 The Calculus for TL-ALCF
This Section presents a calculus for deciding subsumption between temporal concepts in
the Description Logic TL-ALCF . Again, the calculus is based on the idea of separating the
inference on the temporal part from the inference on the Description Logic part (\vALCF "),
and adopting standard procedures developed in the two areas.
Normal Form
Once more, the subsumption calculus is based on a normalization procedure. The rst
step reduces a concept expression into an equivalent existential form { 3(X ) Tc. (Q0 u
Q1 @X1 u : : : u Qn @Xn) { by applying the rewrite rules of Figure 11 augmented with the
491
Artale & Franconi
:> ! ?
:? ! >
:(C u D) ! :C t :D
:(C t D) ! :C u :D
: :C ! C
:8P .C ! 9P .:C
:9P .C ! 8P .:C
:f : C ! f " t f : :C
:p : C ! f " t f : (:q : C ) if p = f q
:p # q ! p"tq"tp"q
:p " q ! p"tq"tp#q
(f p ) " ! f " t f : (p ")
Note: By f we denote both an atomic feature and an atomic parametric feature.
Figure 14: Rewrite rules to transform an arbitrary concept into a simple concept.
rule: p : (q1 " q2 ) ! p q1 " p q2 . Each Q is a non-temporal concept, i.e., it is an element

of the language ALCF .
In the following normalization step there will be a need to verify concept satis ability
for non-temporal concept expressions. An ALCF concept E is unsatis able i E vALCF ?.
Algorithms for checking satis ability and subsumption of concepts terms in ALCF are well
known (Hollunder & Nutt, 1990).
De nition 7.6 (Completed existential form) The temporal completion of a concept
in existential form { the Completed Existential Form, CEF { is obtained by sequentially
applying the following steps:
(closure, collapsing, covering) As reported in de nition 6.5. As for the covering,
translate the concept expression Qy applying the rewrite rule: Qy ! Qy ui=1:::n (Qi1 t
: : : t Q i m ).
(parameter introduction) This requires two phases.
1. Each Q is translated in disjunctive normal form. First the simple form6 is ob-
tained by transforming each Q following the rewrite rules reported in Figure 14.
The disjunctive normal form is then obtained by rewriting each Q { which is now
in simple form { using the following rules, which correspond to the rst order
rules for computing the disjunctive normal form of logical formul:
(C1 t C2 ) u D ! (C1 u D) t (C2 u D)
p : (C t D) ! p : C t p : D
6. A simple concept contains only complements of the form :A, where A is a primitive concept, and no
sub-concepts of the form p ", where p is not an atomic (parametric) feature { this corresponds to a rst
order logical formula in negation normal form.
492
?g1 : : : ?gn [ f [ p]] : C ! ?g1 : : : ?gn : >.

?g1 : : : ?gn [ f [ p]] # g [ q] ! ?g1 : : : ?gn : >.
?g1 : : : ?gn # ?f1 : : : ?fm ! ?g1 : : : ?gn # ?f1 : : : ?fm .
?g1 : : : ?gn g [ p] # ?f1 : : : ?fm [ f [ q]] ! ?g1 : : : ?gn : > u
?f1 : : : ?fm : >.
?g1 : : : ?gn " ?f1 : : : ?fm ! ?g1 : : : ?gn " ?f1 : : : ?fm .
?g " ! ?g " .
?g1 : : : ?gn : (?gn+1 ") ! ?g1 : : : ?gn : (?gn+1 ").
?g1 : : : ?gn [ f [ p]] " g [ q] ! ?g1 : : : ?gn : >.
?g1 : : : ?gn g [ p] " ?f1 : : : ?fm [ f [ q]] ! ?g1 : : : ?gn : > u
?f1 : : : ?fm : >.
Figure 15: Rewrite rules that compute the parameter introduction step.
2. For each Qj = Ej1 t : : : t Ejn , on compute its time-invariant part (let us indicate
this particular concept expression as Q~ j ). This gives Q~ j by computing for each
disjunct Eji in Qj its time-invariant information E~ji . If Eji vALCF ?, then
E~ji = ?. Otherwise, rewrite every conjunct in Eji as showed in Figure 15, while
the conjuncts not considered there are rewrote to >. Now, unless there is an
E~ji = >, Q~ j = E~j1 t : : : t E~jn must be conjunctively added to all the other nodes.
Proposition 7.7 (Equivalence of CEF) Every concept in existential form can be re-
duced into an equivalent completed existential concept.
As for the TL-F case, both covering and parameter introduction can be computed inde-
pendently. As a consequence of the above normalization phase, the proposition 6.7 (node
independence) is now true for TL-ALCF concepts in CEF. Observe that, to obtain a CEF
concept, the steps of the normalization procedure require the computation of the transitive
closure of the temporal relations { which is an NP-complete problem (van Beek & Co-
hen, 1990) { and the computation of ALCF subsumption { which is a PSPACE-complete
problem (Hollunder & Nutt, 1990).
Before the presentation of the last normalization phase, which will eliminate redundant
nodes, it is now possible to check whether a concept expression is satis able.
Proposition 7.8 (Concept satis ability) A TL-ALCF concept in CEF, hX; Tc; Q@X i,
is satis able (with the proviso that the temporal constraints are satis able) if and only if the
non-temporal concepts labeling each node in X are satis able. Checking satis ability of a
TL-ALCF concept in CEF is a PSPACE-complete problem.
Proof. Is a direct consequence of the node independence established by proposition 6.7,
which is true also for TL-ALCF concepts in CEF. 2
The normalization procedure now goes on by rewriting unsatis able concepts to ? and
then computing the essential graph form for satis able concepts. This last phase is more
493
Artale & Franconi
complex than for the other temporal languages considered in this paper essentially because
ALCF can express the > concept by means of a concept expression (e.g., > = A t :A).
From this consideration it follows that in TL-ALCF a redundant node can be derived from
a complex concept expression (e.g., both A t:A, and ?g : A t ?g : :A are redundant nodes).
The key idea is that all the time-invariant information is present in the ] node thanks to
the CEF. Thus it is needed only to extract this information from the ] node by computing
the disjunctive normal form of Q] , applying the ~ translation, and then testing whether
Q~ ] vALCF Qi , for a given node xi.
De nition 7.9 (Essential graph) The subgraph of the CEF of a TL-ALCF conceptual
temporal constraint network T = hX; Tc; Q@X i obtained by deleting the nodes xi such that
Q~ ] vALCF Qi { with the exception of the ] node { is called essential graph of T : (ess T ).
Proposition 7.10 (Equivalence of essential graph) Every CEF concept can be reduced
into an equivalent essential graph form (and, obviously, every concept can be reduced into
an equivalent essential graph form).
Subsumption
The overall normalization procedure reduces the subsumption problem in TL-ALCF to the
subsumption between ALCF concepts.
Theorem 7.11 (TL-ALCF concept subsumption) A concept C1 subsumes a concept
C2 if and only if there exists an s-mapping from the essential graph of C1 to the essential
graph of C2 .
The above theorem gives a sound and complete algorithm for computing subsumption be-
tween TL-ALCF concepts (the proof is the same as the one for theorem 6.21). The sub-
sumption problem is now PSPACE-hard, since satis ability and subsumption for ALCF
concepts were proven to be PSPACE-complete (Hollunder & Nutt, 1990).
8. Extending the Expressivity for States

The following suggests how to extend the basic language to cope with important issues in
the representation of states. (i) Homogeneity allows us to consider properties of the world {
peculiar to states { which remain true in each subinterval of the interval in which they hold.
(ii) Persistence guarantees that a state holding as an e ect of an action continues to hold
unless there is no evidence of its falsity at some time. An approach to the frame problem is
then presented, showing a possible solution to one of the most (in)famous problems in AI
literature. The following subsections shall be interested more in semantically characterizing
actions and states than on computational properties. The extensions proposed now to the
temporal languages are for having a full edged Description Logic for time and action.
8.1 Homogeneity
In the temporal literature homogeneity characterizes the temporal behavior of world states:
when a state holds over an interval of time t, it also holds over subintervals of t. Thus, if
494
] -
Simple-Stack(BLOCK)
r x -
OnTable(BLOCK) r
OnBlock(BLOCK)
y
-
Figure 16: Temporal dependencies in the de nition of the Simple-Stack action.
a block is on the table for a whole day, one can conclude that it is also on the table in the
morning. On the other hand, actions are not necessarily homogeneous. In the linguistic
literature a di erence is made between activity and performance verbs. The distinction
comes out in the fact that activity verbs do have sub-events that are denoted by the same
verb, whereas performance verbs do not. Generally, activity verbs represent ongoing events,
for example to eat and to run, and can be described as homogeneous predicates; whereas
performance verbs represent events with a well de ned granularity in time, such as to prepare
spaghetti. Performance verbs are an example of anti-homogeneous events: if they occur over
an interval of time t, then they do not occur over a subinterval of t, as they would not yet
be completed.
The language is extended by introducing the Homogeneity operator:
C; D ! rC (homogeneous concept)
The semantics of homogeneous concepts is easily given in terms of the semantics of the
temporal universal quanti er: rC 2x (x (=; s; d; f ) ]). C @x. This means that rC
is an homogeneous concept if and only if when it holds at an interval it remains true at
each subinterval. In particular, 2x universally quali es the temporal variable x, while the
temporal constraint (x (=; s; d; f ) ]) imposes that x is a generic interval contained in ].
Moreover, it is always true that rC v C , i.e., rC is a more speci c concept than C .
Let us consider as an example a more accurate de nition of the Basic-Stack action
(see Section 4.1.1):
:
Simple-Stack = 3(x y )(x m ])(] m y ). ((?BLOCK : rOnTable)@x u
(?BLOCK : rOnBlock)@y)
Figure 16 shows the temporal dependencies of the intervals in which the Simple-Stack
holds. The di erence with the Basic-Stack action is the use of the homogeneity operator.
In fact, since the predicates OnTable and OnBlock denote states, their homogeneity should
be explicitly declared. The assertion Simple-Stack(i; a) says that a is an individual action
of type Simple-Stack occurred at interval i. Moreover, the same assertion implies that a
is related to a ?BLOCK, say b, which is of type OnTable at some interval j { meeting i { and
at all intervals included in j , while it is of type OnBlock at another interval l { met by i {
and at all intervals included in l:
Simple-Stack(i; a) =) 9b. ?BLOCK(a; b) ^9 j; l. m(j; i) ^ m(i; l) ^
8 ^; ^l. (=; s; d; f )(^; j ) ^ (=; s; d; f )(^l; l) !
OnTable(^; b) ^ OnBlock(^ l; b):
495
Artale & Franconi
] -
Instant-Stack(BLOCK)
r
OnTable(BLOCK) -r y OnBlock(BLOCK)
-
z
Figure 17: Temporal dependencies in the de nition of the Instant-Stack action.
Note that the Simple-Stack action subsumes the Instant-Stack action, whose temporal
dependencies are depicted in Figure 17:
:
Instant-Stack = 3(z y )(] f z )(] m y ). ((?BLOCK : rOnTable)@z u
(?BLOCK : rOnBlock)@y)
Subsumption holds because the class of intervals { obtained by homogeneity of the state
OnTable as de ned in the Simple-Stack action { including x and all its subintervals is a
subset of the class of intervals over which the block is known to be on the table, according
to the de nition of Instant-Stack { this latter class includes all the subintervals of z .
If the Instant-Stack action had been de ned without the r operator, then it would not
specialize any more the Simple-Stack action. In fact, according to such a weaker de nition
of Instant-Stack, specifying that the object is on the table at z does not imply that the
object is on the table at subintervals of z ; in particular, it is not possible to deduce any
more that the object is on the table at x and its subintervals, as speci ed in the de nition of
Simple-Stack action. Moreover, the weak Instant-Stack action type would not specialize
the weak Simple-Stack action type { i.e., Basic-Stack { too. Thus, homogeneity helps
us to de ne states and actions in a more accurate way, such that important inferences are
captured.
As seen above, the de nition of homogeneity makes use of universal temporal quan-
ti cation. Remember that subsumption in a propositionally complete Description Logic
with both existential and universal temporal quanti cation is undecidable and it is still an
open problem if it becomes decidable in absence of negation (Bettini, 1993). The homo-
geneity operator is a restricted form of universal quanti cation. An even more restricted
form interests us here, where the concept C in rC does not contain any other temporal
operator (called simple homogeneous concept). The expressiveness of the resulting logic is
enough, for example, to correctly represent the homogeneous nature of states. In (Artale,
Bettini, & Franconi, 1994) an algorithm to compute subsumption in TL-F augmented with
the homogeneity operator is proposed. Even if a formal proof is still not available, good
arguments are discussed to conjecture its completeness. This would also prove decidability
of this logic and of the corresponding modal logics.
8.2 Persistence
This Section shows how our framework can be successfully extended in a general way to
cope with inertial properties. In the basic temporal language, a property holding, say, as
a post-condition of an action at a certain interval, is not guaranteed to hold anymore at
other included or subsequent intervals. This is the reason why we propose an extended
496
] - x ]

-:
Load(GUN) Loaded(GUN) Fire(GUN,TARGET) Loaded(GUN)
z
:= Loaded(GUN)
x
or
- Dead(TARGET)
y
Figure 18: De nitions of the actions Load and Fire.
formalism, in which states can be represented as homogeneous and persistent concepts.

As a motivation for introducing the possibility of representing persistent properties in the
language, this Section considers how to solve the frame problem, and in particular the
famous example of the Yale Turkey Shooting Scenario (Sandewall, 1994; Allen & Ferguson,
1994), formerly known as the Yale Shooting Problem.
An inertia operator \= " is introduced here. Intuitively, = C is currently true if it was
true at a preceding interval { say i { and there is no evidence of the falsity of C at any
interval between the current one and i. Thus, the property of an individual of being of type
C persists over time, unless a contradiction arises.
The formalization of the inertia operator makes use of the epistemic operator K (Donini,
Lenzerini, Nardi, Schaerf, & Nutt, 1992), in which KC denotes the set of individuals known
to be instances of the concept C 7 .
De nition 8.1 (Inertia) = C (j; a) i
9i. start(i) start(j ) ^ C (i; a) ^
8h. start(h) end(i) ^ end(h) end(j ) ! :K:C (h; a).
where start and end are two functions giving respectively the starting and the ending point
of an interval { conditions on endpoints are simpler and more readable than their equivalents
on interval relations; :K:C (h; a) means that it is not known that a is not of type C at
interval h. Furthermore, the following relation holds: 8a; j . C (j; a) ! = C (j; a); i.e., = C
subsumes C . The above de nition can be captured by a temporal language equipped with
the epistemic operator { K { and the homogeneity operator { r:
= C C t 3(x y) (x (b; m; o; ; di) ])(x (s; si) y)(y ]).(C @x u r(:K:C )@y)
Two action types are de ned, Load { with the parameter ?GUN { and Fire { with the
parameters ?GUN and ?TARGET (Figure 18):
Load =: 3x (] m x). ?GUN : Loaded@x
:
Fire = 3(x y z ) (] f x)(] m y )(] m z ).
(?GUN : := Loaded@x t ?TARGET : Dead@y) u ?GUN : :Loaded@z
The action Load describes loading a gun. The action Fire describes ring the gun against
a target: e ects of ring are that the gun becomes unloaded and either the target is dead
7. An epistemic interpretation isT a pair (I ,W ) in which I is an interpretation and W is a set of interpreta-
tions such that (KC )I;W = J 2W (C J ;W ).
497
Artale & Franconi
fred :
i
Load(gun)
- i1
Loaded( gun)
gun
Fire(
j
,
- j1
) Loaded( gun)
= Loaded( gun -) Dead(fred)

j0 j2
Figure 19: Actions instances in the Yale Shooting Problem.
or the gun was not loaded { possibly by inertia { before ring. The Yale Shooting Problem
considers the situation described by the following set of assertions (ABox):
Load(i; load-action ); ?GUN(load-action ; gun ); a(j; i); Fire(j; re-action );
?GUN( re-action ; gun ); ?TARGET( re-action ; fred ):
i.e., at the beginning the gun is loaded; then, the action of ring the gun against the target
fred is performed. According to the semantics of the language, logical consequences of the
knowledge base are:
j= 9i1 . m(i; i1 ) ^ Loaded(i1 ; gun )
j= 9j1 . m(j; j1 ) ^ :Loaded(j1 ; gun )
j= 9j0 . f (j; j0 ) ^ = Loaded(j0 ; gun )
j= 9j2 . m(j; j2 ) ^ Dead(j1 ; fred ):
i.e., (see also Figure 19) (i) the Load action makes the gun loaded; (ii) the Fire action
makes the gun unloaded at the end; (iii) since there is no evidence to the contrary, the gun
is still loaded at j0 by inertia; (iv) since the gun is not unloaded at j0 , the target fred must
be dead.
Since the inertia operator is useful to describe the behavior of properties, which are
characterized as homogeneous concepts, a simple way of representing persistence in the
context of homogeneous concepts is proposed.
Proposition 8.2 Let P be a property { i.e., P =: rP 0 is an homogeneous concept { and
a knowledge base such that 6j= P (j; a). = P (j; a) is true in { i.e., j= = P (j; a) {
if and only if two intervals i; k exist such that: j= (start(i) start(j ) ^ P (i; a)) and
[ fs(i; k); f (j; k); P (k; a)g is satis able.
Proof. The entailment test veri es the rst part of the de nition of inertia, while the
satis ability test veri es that, between the interval at which the system knows that the
individual a belongs to P { i { and the interval at which P (a) is deduced by inertia { j
{ does not exist an interval h at which the system knows that P (a) is false. Indeed, such
interval h would be related to the interval k by the relation in and since it is supposed
that P is homogeneous, the knowledge base with :P (h; a) ^ P (k; a) ^ in(h; k) would be
inconsistent. 2
The deduction P (j; a) ! = P (j; a) can be obtained as a particular case of the above stated
proposition.
498
9. Related Works
The original formalism devised by Allen (1991) forms, in its very basis, the foundation for
our work. It is a predicate logic in which interval temporal networks can be introduced,
properties can be asserted to hold over intervals, and events can be said to occur at inter-
vals. His approach is very general, but it su ers from problems related to the semantic
formalization of the predicates hold and occur (Blackburn, 1992). Moreover, computa-
tional properties of the formalism are not analyzed. The study of this latter aspect was, on
the contrary, our main concern.
In the Description Logic literature, other approaches for representing and reasoning with
time and action were proposed. In the beginning the approaches based on an explicit notion
of time are surveyed, and then the Strips-like approaches are considered. This Section ends
by illustrating some of the approaches devoted to temporally extend the situation calculus.
Bettini (1997) suggests a variable-free extension with both existential and universal
temporal quanti cation. He gives undecidability results for a class of temporal languages
{ resorting to the undecidability results of Halpern and Shoham's temporal logic { and in-
vestigates approximated reasoning algorithms. Basically, he extends the ALCN description
logics with the existential and universal temporal quanti ers, but, unlike our formalism,
explicit interval variables are not allowed. The temporal quanti cation makes use of a set
of temporal constraints on two implicit intervals: the reference interval and the current one.
In this framework, the concept of Mortal can be de ned as:
:
Mortal = LivingBeing u 3(after). (not LivingBeing)
Schild (1993) proposes the embedding of point-based tense operators in a propositionally
closed Description Logic. He proved that satis ability in ALCT , the point-based temporal
extension of ALC , interpreted on a linear, unbounded and discrete temporal structure, is
PSPACE-complete. His ideas were applied by (Fischer, 1992; Neuwirth, 1993) in the Back
system. Note that a point-based temporal ontology is unable to express all the variety of
relations between intervals.
Baader and Laux (1995) integrate modal operators for time and belief in a terminological
system looking for an adequate semantics for the resulting combined language. The major
point in this paper is the possibility of using modal operators not only inside concept
expressions but also in front of concept de nitions and assertions. The following example
shows the notion of Happy-father, where di erent modalities interact:
[BEL-JOHN](Happy-father =: 9MARRIED-TO.(Woman u [BEL-JOHN]Pretty) u
hfuturei8CHILD.Graduate)
In this case, it is John's belief that a Happy-father is someone married to a woman believed
to be pretty by John, and whose children will be graduates sometime in the future. The
semantics has a Kripke-style: each modal operator is interpreted as an accessibility relation
on a set of possible worlds, while the domain of objects is split into (possible) di erent
domain objects, each one depending on a given world. This latter choice captures the case of
di erent de nitions for the same concept { such as [BEL-JOHN](A =: B ) and [BEL-PETER](A =:
C ) { since the two formul are evaluated in di erent worlds. The main restriction is that
all the modal operators do not satisfy any speci c axioms for belief or time. On the other
hand, the language is provided with a complete and terminating algorithm that should
499
Artale & Franconi
serve, as the authors propose, \...as a basis for satis ability algorithms for more complex
languages".
There are Description Logics intended to represent and reasoning about actions following
the Strips tradition. Heinsohn, Kudenko, Nebel and Pro tlich (1992) describe the Rat
system, used in the Wip project at the German Research Center for AI (DFKI). They use a
Description Logic to represent both the world states and atomic actions. A second formalism
is added to compose actions in plans and to reason about simple temporal relationships. No
explicit temporal constraints can be expressed in the language. Rat actions are de ned by
the change of the world state they cause, and they are instantaneous as in the Strips-like
systems, while plans are linear sequences of actions. The most important service o ered
by Rat is the simulated execution of part of a plan, checking if a given plan is feasible
and, if so, computing the global pre- and post-conditions. The feasibility test is similar
to the usual consistency check for a concept description: they temporally project the pre-
and post-conditions of individual actions composing the plan, respectively backward and
forward. If this does not lead to an inconsistent initial, nal or intermediate state, the plan
is feasible and the global pre- and post-conditions are determined as a side e ect.
Devanbu and Litman (1991, 1996) describe the Clasp system, a plan-based knowledge
representation system extending the notion of subsumption and classi cation to plans, to
build an ecient information retrieval system. In particular, Clasp was used to repre-
sent plan-like knowledge in the domain of telephone switching software by extending the
use of the software information system lassie (Devanbu, Brachman, Selfridge, & Ballard,
1991). Clasp is designed for representing and reasoning about large collections of plan
descriptions, using a language able to express temporal, conditional and looping operators.
Following the Strips tradition, plan descriptions are built starting from states and actions,
both represented by using the Classic (Brachman, McGuiness, Patel-Schneider, Resnick,
& Borgida, 1991) terminological language. Since plans constructing operators correspond
to regular expressions, algorithms for subsumption integrate work in automata theory with
work in concept subsumption. The temporal expressive power of this system can capture
to sequences, disjunction and iterations of actions and each action is instantaneous. Fur-
thermore, state descriptions are restricted to a simple conjunction of primitive Classic
concepts. Like Rat, Clasp checks if an instantiated plan is well formed, i.e., the speci ed
sequence of individual actions are able to transform the given initial state into the goal state
by using the Strips rules.
We end up by reporting on the e orts made by researchers in the situation calculus
eld to overcome the strict sequential perspective inherent to this framework. Recent works
enrich the original framework to represent properties and actions having di erent truth
values depending not only on the situation but also on time. The work of Reiter (1996),
moving from the results showed by Pinto (1994) and by Ternovskaia (1994), provides a
new axiomatization of the situation calculus able to capture concurrent actions, properties
with continuous changes, and natural exogenous actions { those under nature's control. The
notion of uent { which models properties of the world { and situation are maintained. Each
action is instantaneous and responsible for changing the actual situation to the subsequent
one. Concurrent actions are simply sets of instantaneous actions that must be coherent,
i.e., the action's collection must be non empty and all the actions occur at the same time.
Pinto (1994) and Reiter (1996) introduce the time dimension essentially to capture both
500
the occurrence of the natural actions, due to known laws of physics { i.e., the ball bouncing
at times prescribed by motion's equations { and the dynamic behavior of physical objects
{ i.e., the position of a falling ball. This is realized by introducing a time argument for
each action function, while properties of the world are divided into two di erent classes:
classical uents that hold or do not hold throughout situations, and continuous parameters
that may change their value during the time spanned by the given situation.
More devoted to have a situation calculus with a time interval ontology is the work of
Ternovskaia (1994). In order to describe processes { i.e., actions extended in time { she
introduces durationless actions that initiate and terminate those processes. As a matter of
fact, processes become uents, with instantaneous events { Start(Fluent) and Finish(Fluent)
{ which respectively make true or false the corresponding uent, and with persistence
assumptions that make the uent true during the interval. For example, in a blocks world
the picking-up process is treated as a uent with Start(picking-up(x)) and Finish(picking-
up(x)) instantaneous actions that enable or falsify the picking-up uent.
10. Conclusions
The main objective of this paper was the design of a class of logical formalisms for uni-
formly representing time, actions and plans. According to this framework, an action has a
duration in time, it can have parameters, which are the ties with the temporal evolution
of the world, and it is possibly associated over time with other actions. A model-theoretic
semantics including both a temporal and an object domain was developed, for giving both
a meaning to the language formul and a well founded de nition of the various reasoning
services, allowing us to prove soundness and completeness of the corresponding algorithms.
The peculiar computational properties of this logic make it an e ective representation and
reasoning tool for plan recognition purposes. An action taxonomy based on subsumption
can be set up, and it can play the role of a plan library for plan retrieval tasks.
This paper contributes to exploration of the decidable realm of interval-based temporal
extensions of Description Logics. It presented complete procedures for subsumption rea-
soning with TL-F , TLU -FU and TL-ALCF . In addition, the subsumption problem for
TL-F was proven an NP-complete problem. The subsumption procedures are based on
an interpretation preserving transformation that operates a separation between the tem-
poral and the non-temporal parts of the formalism. Thus, the calculus can adopt distinct
standard procedures developed in the Description Logics community and in the temporal
constraints community. To obtain decidable languages the key idea was to restrict the tem-
poral expressivity by eliminating the universal quanti cation on temporal variables. While
a propositionally complete Description Logic with both existential and universal temporal
quanti cation is undecidable, it is still an open problem if it becomes decidable in absence
of negation. With the introduction of the homogeneity operator investigation of the impact
of a restricted form of temporal universal quanti cation in the language TL-F was begun.
Several extensions were proposed to the basic temporal language. With the possibility
to specify homogeneous predicates the temporal behavior of world states can be described
in a more natural way, while the introduction of the non-monotonic inertial operator gives
rise to some forms of temporal prediction. Another extension { not considered in this paper
{ deals with the possibility of relating an action to more elementary actions, decomposing
501
Artale & Franconi
it in partially ordered steps (Artale & Franconi, 1995). This kind of reasoning is found in
hierarchical planners like Nonlin (Tate, 1977), Sipe (Wilkins, 1988) and Forbin (Dean,
Firby, & Miller, 1990).
Acknowledgements
This paper is a substantial extension and revision of (Artale & Franconi, 1994). The work
was partially supported by the Italian National Research Council (CNR) project \Ontologic
and Linguistic Tools for Conceptual Modeling", and by the \Foundations of Data Warehouse
Quality" (DWQ ) European ESPRIT IV Long Term Research (LTR) Project 22469. The
rst author wishes to acknowledge also LADSEB-CNR of Padova and the University of
Firenze for having supported part of his work. Some of the work carried on for this paper
was done while the second author was working at ITC-irst, Trento. This work owes a lot to
our colleagues Claudio Bettini and Alfonso Gerevini, for having introduced us many years
ago to the temporal maze. Special thanks to Achille C. Varzi, for taking time to review the
technical details of the paper and for his insightful comments on the philosophy of events,
and to Fausto Giunchiglia, for useful discussions and feedback. Thanks to Paolo Bresciani,
Nicola Guarino, Eugenia Ternovskaia and Andrea Schaerf for enlightening comments on
earlier drafts of the paper. Werner Nutt and Luciano Sera ni helped us to have a deeper
insight into logic. We would also like to thank Carsten Lutz for the helpful discussions we
had with him about temporal representations. Many anonymous referees checked out many
errors of previous versions of the paper. All the errors of the paper are, of course, our own.
References
Allen, J. F. (1991). Temporal reasoning and planning. In Allen, J. F., Kautz, H. A., Pelavin,
R. N., & Tenenberg, J. D. (Eds.), Reasoning about Plans, chap. 1, pp. 2{68. Morgan
Kaufmann.
Allen, J. F., & Ferguson, G. (1994). Actions and events in interval temporal logic. Journal
of Logic and Computation, 4 (5). Special Issue on Actions and Processes.
Artale, A., Bettini, C., & Franconi, E. (1994). Homogeneous concepts in a temporal de-
scription logic. In F.Baader, M.Lenzerini, W.Nutt, & P.F.Patel-Schneider (Eds.),
Workshop Notes of the Int. Workshop on Description Logics, DL-94, pp. 36{41 Bonn,
Germany. DFKI, Saarbrucken. Tech. Rep. DFKI-D-94-10.
Artale, A., & Franconi, E. (1994). A computational account for a description logic of
time and action. In J.Doyle, E.Sandewall, & P.Torasso (Eds.), Proc. of the 4 th
International Conference on Principles of Knowledge Representation and Reasoning,
pp. 3{14 Bonn, Germany. Morgan Kaufmann.
Artale, A., & Franconi, E. (1995). Hierarchical plans in a description logic of time and
action. In A.Borgida, M.Lenzerini, D.Nardi, & B.Nebel (Eds.), Workshop Notes of the
Int. Workshop on Description Logics. DL-95, pp. 1{5 Roma, Italy. Tech. Rep. 07.95.
Also in the Workshop Notes of the IJCAI-95 Workshop on \The Next Generation of
502
Plan Recognition Systems: Challanges for and Insight from Related Areas of AI",
Montreal, 1995.
Baader, F., Burckert, H.-J., Heinsohn, J., Hollunder, B., Muller, J., Nebel, B., Nutt, W.,
& Pro tlich, H.-J. (1990). Terminological knowledge representation: a proposal for a
terminological logic. Technical memo TM-90-04, DFKI, Saarbrucken, Germany.
Baader, F., & Laux, A. (1995). Terminological logics with modal operator. In Proc. of the
13 th IJCAI, pp. 808{814 Montreal, Canada.
Bettini, C. (1993). Temporal Extensions of Terminological Languages. Ph.D. thesis, Com-
puter Science Department, University of Milan, Italy.
Bettini, C. (1997). Time dependent concepts: Representation and reasoning using temporal
description logics. Data & Knowledge Engineering, 22 (1), 1{38.
Blackburn, P. (1992). Fine grained theories of time. In Working Papers of the 4th Intl.
Workshop on Semantics of Time, Space, Movement, and Spatio-Temporal Reasoning,
pp. 299{320.
Brachman, R. J., McGuiness, D. L., Patel-Schneider, P. F., Resnick, L. A., & Borgida, A.
(1991). Living with classic: When and how to use a kl-one-like language. In Sowa,
J. (Ed.), Principles of Semantic Networks. Morgan Kaufmann.
Buchheit, M., Donini, F. M., & Schaerf, A. (1993). Decidable reasoning in terminological
knowledge representation systems. Information Systems, 1, 109{138.
Calvanese, D., Lenzerini, M., & Nardi, D. (1994). A uni ed framework for class-based
representation formalisms. In Proc. of the 4 th International Conference on Principles
of Knowledge Representation and Reasoning Bonn, Germany.
Cohen, W., Borgida, A., & Hirsh, H. (1992). Computing least common subsumers in
description logics.. pp. 754{760 San Jose, CA.
De Giacomo, G., & Lenzerini, M. (1996). Tbox and abox reasoning in expressive description
logics. In Proc. of the 5 th International Conference on Principles of Knowledge
Representation and Reasoning, pp. 316{327 Boston, MA. Morgan Kaufmann.
De Giacomo, G., & Lenzerini, M. (1995). What's in an aggregate: Foundations for descrip-
tion logics with tuples and sets. In Proc. of the 13 th IJCAI Montreal, Canada.
Dean, T., Firby, J., & Miller, D. (1990). Hierarchical planning involving deadlines, travel
time and resources. Computational Intelligence, 6 (1).
Devanbu, P. T., & Litman, D. J. (1991). Plan-based terminological reasoning. In Proc.
of the 2 nd International Conference on Principles of Knowledge Representation and
Reasoning, pp. 128{138 Cambridge, MA.
Devanbu, P. T., & Litman, D. J. (1996). Taxonomic plan reasoning. Arti cial Intelligence,
84, 1{35.
503
Artale & Franconi
Devanbu, P., Brachman, R., Selfridge, P., & Ballard, B. (1991). LASSIE { a knowledge-
based software information system. Communication of the ACM, 34 (5).
Donini, F. M., Hollunder, B., Lenzerini, M., Spaccamela, A. M., Nardi, D., & Nutt, W.
(1992). The complexity of existential quanti cation in concept languages. Arti cial
Intelligence, 53, 309{327.
Donini, F. M., Lenzerini, M., Nardi, D., & Nutt, W. (1995). The complexity of concept
languages. Tech. rep. RR-95-07, DFKI, Germany. A preliminary version appears in
Proc. of the 2nd International Conference on Principles of Knowledge Representation
and Reasoning (KR-91).
Donini, F. M., Lenzerini, M., Nardi, D., & Schaerf, A. (1994). Deduction in concept
languages: from subsumption to instance checking. Journal of Logic and Computation,
4 (4), 423{452.
Donini, F. M., Lenzerini, M., Nardi, D., Schaerf, A., & Nutt, W. (1992). Adding epistemic
operators to concept languages. In Proc. of the 3 rd International Conference on
Principles of Knowledge Representation and Reasoning, pp. 342{353 Cambridge, MA.
Fikes, R. E., & Nilsson, N. (1971). STRIPS: a new approach to the application of theorem
proving as problem solving. Arti cial Intelligence, 2, 198{208.
Fischer, M. (1992). The integration of temporal operators into a terminological representa-
tion system. Kit-report 99, Technische Universtitat Berlin, Germany.
Gerevini, A., & Schubert, L. (1994). On point-based temporal disjointness. Arti cial
Intelligence, 70, 347{361.
Halpern, J. Y., & Moses, Y. (1985). A guide to the modal logic of knowledge and belief:
Preliminary draft. In Proc. of the 9 th IJCAI, pp. 480{490 Los Angeles, CA.
Halpern, J. Y., & Shoham, Y. (1991). A propositional modal logic of time intervals. Journal
of ACM, 38 (4), 935{962.
Heinsohn, J., Kudenko, D., Nebel, B., & Pro tlich, H. (1992). RAT: representation of
actions using terminological logics. Tech. rep., DFKI, Saarbrucken, Germany.
Hollunder, B., & Nutt, W. (1990). Subsumption algorithms for concept languages. Tech.
rep. RR-90-04, DFKI, Germany.
Kautz, H. A. (1991). A formal theory of plan recognition and its implementation. In Allen,
J. F., Kautz, H. A., Pelavin, R. N., & Tenenberg, J. D. (Eds.), Reasoning about Plans,
chap. 2, pp. 69{126. Morgan Kaufmann.
Lifschitz, V. (1987). On the semantics of strips. In The 1986 Workshop on Reasoning
about Actions and Plans, pp. 1{10. Morgan Kaufman.
McCarthy, J., & Hayes, P. J. (1969). Some philosophical problems from the standpoint
of Arti cial Intelligence. In Meltzer, B., & Michie, D. (Eds.), Machine Intelligence,
Vol. 4, pp. 463{502 Edinburgh, UK. Edinburgh University Press.
504
Nebel, B. (1990). Terminological reasoning is inherently intractable. Arti cial Intelligence,

43, 235{249.
Nebel, B. (1991). Terminological cycles: Semantics and computational properties. In Sowa,
J. F. (Ed.), Principles of Semantic Networks, chap. 11, pp. 331{362. Morgan Kauf-
mann.
Neuwirth, A. (1993). Inferences for temporal object descriptions in a terminological repre-
sentation system: Design and implementation. Kit-report 107, Technische Universtitat
Berlin, Germany.
Pinto, J. A. (1994). Temporal Reasoning in the Situation Calculus. Ph.D. thesis, Department
of Computer Science, University of Toronto.
Reiter, R. (1996). Natural actions, concurrency and continuous time in the situation calculs.
In Proc. of the 5 th International Conference on Principles of Knowledge Representa-
tion and Reasoning Boston, MA.
Renz, J., & Nebel, B. (1997). On the complexity of qualitative spatial reasoning: a maximal
tractable fragment of the region connection calculus. In Proc. of the 14 th IJCAI, pp.
522{527 Nagoya, Japan.
Sandewall, E. (1994). Features and Fluents. The Representation of Knowledge about Dy-
namical Systems, Vol. I. Oxford University Press.
Sandewall, E., & Shoham, Y. (1994). Non-monotonic temporal reasoning. In Gabbay, D.
(Ed.), Handbook of Arti cial Intelligence and Logic programming. Oxford University
Press.
Schaerf, A. (1994). Reasoning with individuals in concept languages. Data & Knowledge
Engineering, 13 (2), 141{176.
Schild, K. D. (1991). A correspondence theory for terminological logics: Preliminary report.
In Proc. of the 12 th IJCAI, pp. 466{471 Sidney, Australia.
Schild, K. D. (1993). Combining terminological logics with tense logic. In Proceedings of
the 6th Portuguese Conference on Arti cial Intelligence, EPIA'93.
Schmidt-Schau, M., & Smolka, G. (1991). Attributive concept descriptions with comple-
ments. Arti cial Intelligence, 48 (1), 1{26.
Schmiedel, A. (1990). A temporal terminological logic. In Proc. of AAAI-90, pp. 640{645
Boston, MA.
Smolka, G. (1992). Feature constraint logics for uni cation grammar. Journal of Logic
Programming, 12, 51{87.
Tate, A. (1977). Generating project networks. In Proc. of the 5 th IJCAI, pp. 888{893
Cambridge, MA.
505
Artale & Franconi
Ternovskaia, E. (1994). Interval situation calculus. In Workshop Notes of the ECAI-94

Workshop \Logic and Change", pp. 153{164 Amsterdam.
van Beek, P., & Cohen, R. (1990). Exact and approximate reasoning about temporal
relations. Computational Intelligence, 6, 132{144.
van Beek, P., & Manchak, D. W. (1996). The design and experimental analysis of algorithms
for temporal reasoning. Journal of Arti cial Intelligence Research, 4, 1{18.
van Dalen, D. (1994). Logic and Structure. Springer-Verlag.
Venema, Y. (1990). Expressiveness and completeness of an interval tense logic. Notre Dame
Journal of Formal Logic, 31 (4), 529{547.
Weida, R. (1996). Closed Terminologies and Temporal reasoning in Descriptions for Plan
Recognition. Ph.D. thesis, Department of Computer Science, Columbia University,
New York, NY.
Weida, R., & Litman, D. (1992). Terminological reasoning with constraint networks and
an application to plan recognition. In Proc. of the 3 rd International Conference on
Principles of Knowledge Representation and Reasoning, pp. 282{293 Cambridge, MA.
Weida, R., & Litman, D. (1994). Subsumption and recognition of heterogeneous constraint
networks. In Proceedings of CAIA-94.
Wilkins, D. (1988). Practical planning. Morgan Kaufmann, San Mateo CA.
506

Temporal Description Logics

Uploaded by

Copyright:

Available Formats

Temporal Description Logics

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Temporal Description Logics

Uploaded by

Copyright:

Available Formats

Journal of Arti cial Intelligence Research 9 (1998) 463-506 Submitted 4/98; published 12/98

A Temporal Description Logic

Figure 1: Syntax rules for the ALCF Description Logic.

Figure 2: The extensional and transformational semantics in ALCF .

A knowledge base is a nite set  of terminological and assertional axioms. An interpre-

3. Towards a Temporal Description Logics

Moreover, it is possible to build temporal structured actions { as opposed to the atomic

4. The Feature Temporal Language TL-F

Relation Abbr. Inverse i j

Figure 3: The Allen's interval relationships.

Figure 4: Syntax rules for the interval Description Logic TL-F

4.1.1 A clarifying Example

(s)E = fh[u; v]; [u1 ; v1 ]i 2 T<? T<? j u = u1 ^ v < v1 g

Figure 6: The temporal interpretation function.

Basic-Stack( i; a) =) 9b. ?BLOCK(a; b) ^ 9j; l. (OnTable(j; b) ^ OnBlock(l; b) ^

AIV ;t;H = fa 2 I j ht; ai 2 AI g = AIt

Figure 7: The interpretation function.

Figure 8: Temporal dependencies in the de nition of the Boil-Spaghetti plan.

An interpretation I satis es the terminological axiom A =: C i AIt = CtI , for every t.

5. Action and plan representation: two examples

Figure 9: Temporal dependencies in the de nition of Assemble-Spaghetti-Marinara.

5.1 The Cooking Domain

Figure 10: Temporal dependencies in the de nition of the Stack action.

However, note that Boil-Spaghetti does not subsume Prepare-Spaghetti, even if it is a

at the interval denoted by x; while (?OBJECT1  ON # ?OBJECT2)@y indicates that at the

6. The Calculus for TL-F

Procedure Covering( hX; Tci; y):

the node y and each of the nodes in s is inconsistent

Figure 12: Procedure which computes a covering.

F concept expressions C such that C 6w Q, there exists an interpretation I such that

7. Extending the Propositional Part of the Language

Figure 13: Rewrite rules for computing the disjunctive form.

rule: p : (q1 " q2 ) ! p  q1 " p  q2 . Each Q is a non-temporal concept, i.e., it is an element

?g1  : : :  ?gn [ f [ p]] : C ! ?g1  : : :  ?gn : >.

8. Extending the Expressivity for States

Figure 16: Temporal dependencies in the de nition of the Simple-Stack action.

Figure 17: Temporal dependencies in the de nition of the Instant-Stack action.

Figure 18: De nitions of the actions Load and Fire.

formalism, in which states can be represented as homogeneous and persistent concepts.

= Loaded( gun -) Dead(fred)

Figure 19: Actions instances in the Yale Shooting Problem.

Nebel, B. (1990). Terminological reasoning is inherently intractable. Arti cial Intelligence,

Ternovskaia, E. (1994). Interval situation calculus. In Workshop Notes of the ECAI-94

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

A knowledge base is a nite set of terminological and assertional axioms. An interpre-

at the interval denoted by x; while (?OBJECT1 ON # ?OBJECT2)@y indicates that at the

rule: p : (q1 " q2 ) ! p q1 " p q2 . Each Q is a non-temporal concept, i.e., it is an element

?g1 : : : ?gn [ f [ p]] : C ! ?g1 : : : ?gn : >.

= Loaded( gun -) Dead(fred)