2.4 Automata-Based De ision Algorithms
2.4.2 Axiomati Satisability of LTL Formulae
In order to de ide axiomati satisabilityof LTL formulae, we will onstru t an
au-tomatonwhosesu essfulruns orrespondto omputations fortheinput. Noti ethat
a omputation :N !
P
(P) an be seen also as a unarytree, that is,a tree where every node hasexa tlyone su essor. Morepre isely,ea h node representsone pointin time and the su essor relation in this tree is given by the standard ordering of
natural numbers. Thus, the automaton we onstru t will have the unique unlabeled
unarytree asinput. Thestatesofthisautomatonwillbesets ofLTLformulae,whi h
intuitivelyrepresenttheset ofall formulae thataresatisedata givenpointintime.
In that sense, these states orrespond to the Hintikka sets dened in the previous
subse tion. Noti e nonetheless that this orresponden e willnot bepre ise sin e for
LTLwe willfollow theideasofprevious automata onstru tions(e.g. [WVS83 ℄), and
hen e willnotassumethat theformulaeare innegationnormal form. Givenan LTL
formulaandasetofLTLformulaeR,wedenethe losureof(;R)asthesetofall
subformulae of and R, and their negations, where doublenegations are an elled.
Thisset isdenoted by l(;R).
The statesofourautomatonareso- alledelementarysets offormulae,whi hplay
theroleoftheHintikkasets oftheprevioussubse tion; thatis,they aremaximaland
onsistent sets of subformulae in l(;R).
Denition 2.21 (Elementary set). Aset H l(;R) is alled an elementaryset
for (;R) if it satises the following onditions:
:2H i 2= H;
^ 2H i f; gH;
2H implies U 2H;
if U 2H and 2= H,then 2H
As we have said before,theautomaton forsatisabilityof LTLformulaewilltake
unary trees asinputs; i.e., its runs willbe innite words over the set of states. The
transitionrelationisthusbinary. Thistransitionrelationmakessurethatthetemporal
next formula inthelabelofanode,thenitssu essornodemust ontain . This
is formalisedbythefollowing denition.
Denition 2.22 (Compatible). A tuple (H;H 0
) of elementary sets is alled
om-patible i it satises the following onditions:
for all 2 l(;R), 2H i 2H 0
; and
for all
1 U
2
2 l(;R),
1 U
2
2H i either (i)
2
2H or (ii)
1
2H and
1 U
2 2H
0
.
The runs of our automaton will be sequen es of elementary sets where ea h two
onse utiveonesforma ompatibletuple. In ontrastto the aseforSI,thepresen e
ofarunofthisautomatondoesnotimplytheexisten eofa omputation. Thereason
is that one an delay the satisfa tion of an until formula indenitely; that is, every
node inthe run may have theformula
1 U
2
whilenone has
2
, violating this way
thelast onditioninthedenitionofa omputationfortheinput(seeDenition2.9).
In order to rule out these kinds of runs and make sure that ea h until formula is
eventuallysatised,we willimposea generalised Bu hi onditionwhi h introdu esa
set of nalstates forea h untilformulain l(;R). Intuitively,ea h su h set ofnal
states isin harge ofenfor ing theeventual satisfa tionof one spe i untilformula.
Denition 2.23 (Automaton A sat
;R
). Let and R be an LTL formula and a set
of LTL formulae, respe tively, and let
1 U
1
;:::;
n U
n
be all the until formulae in
l(;R). Thegeneralised B u hiautomaton A sat
;R
:=(Q;;I;F
1
;:::;F
n
) is givenby
Q is the setof all elementary sets for (;R);
onsists of all ompatible pairs (H;H 0
)2QQ;
I :=fH 2QjR[fgHg;
for 1in;F
i
:=fH 2Qj
i
2H or
i U
i
= 2Hg.
The su essfulruns of thisautomaton whoseroot is labelledwithan initial state
orrespondtothe omputationsfortheinput(;R). Fromthis,weobtainthe
follow-ing result[WVS83℄.
Theorem 2.24. Let be an LTL formula and R a set of LTL formulae. The
au-tomaton A sat
;R
has a su essfulrun r with r(")2I i isaxiomati satisable w.r.t.
R.
From thistheorem itfollowsthat axiomati satisabilityof LTL formulae an be
de idedbyan emptiness teston theautomaton A sat
;R .
Inthis hapter we have des ribedseveralpreviouslyknownalgorithmsfor
reason-in lusionof more omplex onstru tors and axiomsrestri ting theinterpretationsfor
on epts and roles in DLs. We then left the DL familyto in lude also the temporal
operatorsfor LTL.
Broadly, we showed the main hara teristi s of two dierent approa hes for
on-stru ting de isionpro edures. Onone hand,the tableau-based method,that tries to
onstru t a model whilekeepingtherestri tions imposedby theaxioms(in luded as
expansion rules). On the other hand is the automata-based approa h that tries to
onstru t an automaton forwhi han emptiness test leadsto a orre t de ision.
The parti ularinstan esofde isionpro edurespresentedinthis hapterwillhelp
us formalise the notions of general tableau algorithms (in Chapter 3) and so- alled
axiomati automata (in Chapter 5), respe tively. We will then show how ea h of
these de ision pro edures an be modied to obtain what is alled a pinpointing
pro edure; intuitively, one that will allow us to dedu e how the presen e of ertain
axioms in uen es the property being tested. The output of a pinpointing pro edure
willbe the so- alled pinpointing formula, from whi h all explanationsand diagnoses
Tableaux and Pinpointing
The previous hapter introdu ed pro edures that allow us to de ide if a property,
su h as subsumptionorsatisabilityof on ept names, follows from a set of axioms.
Thesets ofaxiomsused ould take verydierent shapes;namely, on ept denitions,
assertional axioms, or GCIs, in the ase of DLs, or LTL formulae. The de ision
pro edures we presented ame in two avours: the tableau-like and the
automata-based pro edures. It is thegoal of thiswork to show how to extend them insu h a
waythat, on eade isionismade,weareable tojustify itbyretrievingthoseaxioms
thatarerelevantfortheobtainedanswer. Theapproa hfollowedinthiswork onsists
onndingamonotoneBooleanformula,whi hwe allpinpointingformula,fromwhi h
the desired sets of axioms an be dedu ed. The present and following hapters will
deal withthe tableau-like methods, whilewe delay thetreatment of automata-based
pro edures untilChapter 5.
Before we an begin with the task of extending any kind of algorithm, we need
to formally des ribe the problemthat we are trying to solve; namely, theproperties
thatshouldbesatisedbythepinpointingformula. Thisinturnwillrequireaformal
denition of the kinds of properties that the original pro edures de ide. All these
notionsare introdu ed inSe tion3.1.
Afterwards,we pro eedto des ribe extensions of tableau-like de isionpro edures
that ompute the desired pinpointing formula. In order to improve understanding,
thisisdoneintwosteps. We rstfo usinthespe ial aseofgroundtableauxofwhi h
thesubsumptionalgorithmof Se tion2.3.1 isan instan e. We thengeneralise allthe
notions and results to what we all general tableaux in Se tion 3.3. This notion
en- ompassesthepro eduresdes ribedinSe tions2.3.2and2.3.3, butisnotabletodeal
with blo king onditions as des ribed in the last two se tions of the previous
hap-ter. Thepinpointingextensionsofgeneraltableaux areshownto orre tly omputea
pinpointingformulawheneverthey terminate.
Theextensionpresentedinthis hapterfollowstheideasintrodu edbyBaaderand
Hollunder in[BH95 ℄. There, the onsisten y algorithm forALC ABoxes is extended
by a labellingte hnique that ultimately omputes a pinpointing formula. A similar
approa h was followed by S hloba h and Cornet [SC03 ℄ for on ept unsatisability
withrespe tto so- alledunfoldableALC terminologies. The maindieren ebetween
BaaderandHollunder'sapproa handthatbyS hloba handCornetisthatthelatter
tries to nd the sets of axioms that are relevant to unsatisability dire tly, rather
than by usingthe intermediarypinpointing formulaas donein theformer approa h.
Inreality,theresultobtainedusingthemethodin[SC03℄ anbeseenasapinpointing
formulawrittenindisjun tivenormalform. Althoughtheseideashavebeenextended
to in lude additional onstru tors or usedierent kinds of axioms (see, forinstan e,
[PSK05, MLBP06℄), ea h of these extensions has beenmade to work spe i ally for
the language being studied. Nonetheless, ex ept for the ase dealing with blo king
[LMP06 ℄ thatneeds spe ialattention, they all followthesame basi ideas.
Unfortunately, as shown at the end of this hapter, there is no warranty that
the extended algorithm will stop after a nite number of steps, even if the original
tableaudoes. Thisfa tisspe iallyrelevantsin enoneof thepapers itedsofardeals
with termination of the extensions they present. A tually, termination is usually
disregarded astrivially followingfrom thesame ausesof terminationof the original
tableau, givingno furtherinsightinto whi h these ausesare inreality. It willbe the
taskofChapter4tointrodu eaframeworkwhereboth,tableauxandtheirpinpointing
extensions,areguaranteedtoterminate. Itisinthat haptertoothatwewillintrodu e
thenotionof blo kingforgeneraltableaux and theirpinpointingextensions.
3.1 Basi Notions for Pinpointing
Webeginthisse tionbydeningthegeneralform oftheinputsforthede ision
algo-rithms used along thiswork. These inputs, alled axiomatised inputs, onsist of two
parts. Intuitively,one part orrespondsto a knowledge base, that is,a set of axioms
possibly restri tedto satisfy additionalinternal restri tions, and the other expresses
the instan e of theinferen e problemthat needs to be tested againstthis knowledge
base. The internal restri tions in the set of axioms are ne essary for modelling e.g.
a y li - or SI-TBoxes, where not every set of axioms is allowed. Indeed, a y li
TBoxes requireevery on ept name to appear at most one in theleft-hand-side of a
on ept denition, and SI-TBoxes are restri ted to allow the use of ea h role name
in at most one inverse axioms. But noti e that in both ases, if a set of axioms is
allowedtobeusedasaknowledge base,thenanyof itssubsetsisalsoallowed. Inour
general approa hwekeep thisproperty.
The onsequen es in whi h we are interested need to satisfy a monotoni ity
re-stri tion inthe sensethatadding axiomsto theknowledge base an onlymake more
onsequen es true, but not falsify any that already follows from the original set of
axioms. A property is merely a set of axiomatised inputs, and the de ision
prob-lemasso iatedwithsu h property onsist onde iding,fora givenaxiomatised input,
whether itbelongsto theset ornot. A propertythatmodels onsequen es satisfying
themonotoni ity restri tionstatedabove willbe alled onsequen e property.
Denition 3.1 (Axiomatised input, -property). Let I be a set, alled the set
of inputs, T be a set, alled the set of axioms, and let
P
admis
(T)
P
fin
(T) be a
set of nite subsets of T.
P
admis
(T) is alled admissible if T 2
P
admis
(T) implies
T 0
2
P
(T) for all T 0T. An axiomatised inputfor I and
P
(T) is of theform (I;T) where I2I and T 2
P
admis (T).
A onsequen eproperty(or -propertyfor short)isa setP I
P
admis
The idea behind -properties on axiomatised inputs is to model onsequen e
re-lationsin logi , i.e., the -propertyP holdsifthe inputI \follows" from theaxioms
inT. The monotoni ityrequirement on -properties orresponds to thefa t thatwe
want to restri t the attention to onsequen e relations indu ed by monotoni logi s.
Infa t, fornon-monotoni logi s,lookingat minimalsetsof axiomsthathave agiven
onsequen edoesnotmake mu h sense.
To illustrate Denition 3.1, onsiderthe set N
C
of on ept names. Assume that
I is the set of ordered pairsN
C
N
C
and that T onsists of all H L-GCIs over these
on ept names. Then thefollowingis a -propertya ording to the above denition:
P := f((C ;D);T) j C v
T
Dg: This property represents subsumptionw.r.t. general
H L-TBoxes. As a on rete example, onsider := ((A;B);T) where T onsists of
It is easy to see that 2 P. Note that Denition 3.1 is general enough to apture
other variants of the example above, for instan e, where I 0
onsist of tuples of the
form (C ;D;T
1
)2I
P
fin
(T)and the -property isdenedas
P
For example, if we take the axiomatised input 0
Due to the monotoni ity of -properties, it may well be that some axioms are
irrelevant fordedu inga onsequen e. If we areinterested injustifying su h a
onse-quen e,wewouldneedto getridofallthose irrelevantaxiomsandpresent aminimal
knowledge basefromwhi hthe onsequen e stillfollows. If, onthe ontrary,the
on-sequen e isdete tedasan error,wemightwant to removeonlyenoughaxiomstoget
rid ofit butnotmore,sin ethat might alsoremove some desired onsequen es.
Denition 3.2 (MinA,MaNA). Given an axiomatised input =(I;T) and a
-property P, a setof axioms S T is alled a minimalaxiom set(MinA)for w.r.t.
Note that the notionsof MinA and MaNA areonlyinterestingin the ase where
2 P. In fa t, otherwise the monotoni ity property satised by P implies that
MIN
P( )
=; and MAX
P( )
=fTg. In the above example, where we have 2 P, it
is easy to seethat MIN
P( )
gg. In thevariant ofthe
examplewhereonlysubsetsof fax
1
ThesetMAX
P( )
anbeobtainedfromMIN
P( )
by omputingtheminimalhitting
sets of MIN
P( )
,and then omplementing these sets [SC03, LS05 ℄. A set S T is a
hittingset ofMIN
P( )
ifithasanonemptyinterse tionwitheveryelementofMIN
P( ) ,
and is a minimal hitting set if no stri t subset of S is itself a hitting set. In our
example, the minimal hitting sets of MIN
P( )
gg. The intuition behind this
redu tionsisthat,togetasetofaxiomsthatdoesnothavethe onsequen e,wemust
remove from T at leastone axiom forevery MinA,and thus theminimalhitting sets
give ustheminimalsets to be removed.
Theredu tionwehavejustsket hedshowsthatitisenoughtodesignanalgorithm
for omputing all MinAs, sin e the MaNAs an then be obtained by a hitting set
omputation. It should be noted, however, that this redu tion is not polynomial:
there may be exponentially many hitting sets of a given olle tion of sets, and even
de iding whether su h a olle tion has a hitting set of ardinality nis already an
NP- omplete problem[GJ79 ℄. Also note that there is a similar redu tion involving
hitting sets for omputingtheMinAsfrom allMaNAs.
Instead of omputing MinAs or MaNAs, one an also ompute the pinpointing
formula.
10
To dene the pinpointingformula, we assume that every axiom t2 T is
labeled with a unique propositional variable, whi h we denote as lab(t). Let lab(T)
betheset ofall propositional variableslabelinganaxiom inT. A monotone Boolean
formula over lab(T) isa Booleanformulausing(someof) thevariablesinlab(T) and
onlythe onne tives onjun tionanddisjun tion. Wefurtherassumethattheformula
>, whi h is always evaluated as true, is a monotone Boolean formula. As usual, we
identifyapropositionalvaluation withthesetofpropositionalvariablesitmakestrue.
Foravaluation V lab(T),let T
V
:=ft2T jlab(t)2Vg.
Denition 3.3 (Pinpointing formula). Given a -property P and an axiomatised
input =(I;T), a monotone Boolean formula over lab(T) is alled a pinpointing
formulaforP and ifthefollowingholdsfor everyvaluationV lab(T): (I;T
V
g as the set of propositional
variables. Itis easyto see that(ax
1
isapinpointingformula forP
and .
Valuations have a natural partial order by means of set in lusion, whi h allows
us to speak about minimal and maximal valuations. The following is an immediate
onsequen e ofthe denitionof apinpointingformula[BH95 ℄.
Lemma 3.4. Let P be a -property, =(I;T) an axiomatised input, and a
pin-pointingformula for P and . Then
MIN
P( )
= fT
V
jV isa minimal valuation satisfying g
MAX
P( )
= fT
V
jV isa maximal valuation falsifyingg
10
This orrespondsto whatwas alled the lashformula in[BH95 ℄. Here, wedistinguishbetween
the pinpointingformula, whi h anbe denedindependentlyof atableau algorithm, and the lash
This lemma shows that itis enough to design an algorithm for omputinga
pin-pointing formula to obtainall MinAsand MaNAs. However, like theprevious
redu -tion for omputingMAX
P( )
from MIN
P( )
,theredu tionsuggested bythelemma is
notpolynomial. Forexample,to obtainMIN
P( )
from,one anbringinto
disjun -tivenormalformandthenremovedisjun tsimplyingotherdisjun ts. Itiswell-known
thatthis an ausean exponentialblowup. Conversely,however,thesetMIN
P( ) an
dire tly be translatedintothe pinpointingformula
_
S2MIN
P( )
^
s2S
lab(s): (3.2)
Returning to our example, the pinpointing formula obtained in this fashion from
MIN
P( )
=ffax
1
;ax
2
;ax
4 g; fax
2
;ax