1 Expressing Default Abduction Problems as Quanti(cid:12)ed Boolean Formulas Hans Tompits esis" as explanation for observed facts, according Institut fu(cid:127)r Informationssysteme 184/3, to known laws modelling a speci(cid:12)c domain under Technische Universita(cid:127)t Wien, consideration [29]. Favoritenstra(cid:25)e 9{11, A{1040 Vienna, Austria Various formalisations of abductive reasoning E-mail: [email protected] have been proposed in the literature, di(cid:11)ering ba- sically in the way the domain knowledge is repre- sented. Besides abduction by set-covering [30] or Abduction is the process of (cid:12)nding explanations for probabilisticabduction [28;30],animportantcat- observed phenomena in accord to known laws about egory of abductive reasoning techniques is logic- a given application domain. This form of reasoning is animportantprincipleofcommon-sensereasoningand based abduction, in which the application domain is particularly relevant in conjunction with nonmono- ismodelledbymeansofalogicaltheory.Although tonicknowledgerepresentationformalisms.Inthispa- classical logic is mainly used for this purpose [32], per, we deal with a model for abduction in which the recent years witnessed an increasing interest in domainknowledgeisrepresentedintermsofadefault employing nonmonotonic reasoning formalisms as theory. We show how the main reasoning tasks asso- underlying language for logic-based abduction. As ciated with this particular form of abduction can be a case in point, abductive logic programming [18; axiomatisedwithinthelanguageofquanti(cid:12)edBoolean 36; 11; 19] is a well-known realisation of this logic. More speci(cid:12)cally, we provide polynomial-time paradigm, and successful implementations of this constructible reductions mapping a given abduction particular approach exist [20; 5; 8]. problemintoaquanti(cid:12)edBooleanformula(QBF)such Anotherimportantnonmonotonicreasoningfor- that the satisfying truth assignments to the free vari- malism relevant for abduction is default logic [33]. ablesofthelatterdeterminethesolutionsoftheorigi- nalproblem.Sincetherearenowe(cid:14)cientQBF-solvers A formal model for abduction from default the- available,thisreductiontechniqueyieldsastraightfor- ories has been introduced by Eiter, Gottlob, and ward method to implement the discussed abduction Leone [12]. In this approach, a default abduction tasks. We describe a realisation of this approach by problem, P, consists of a default theory T = appeal to the reasoning system QUIP. hW;(cid:1)i (where W is a set of formulas and (cid:1) is a Keywords:Defaultlogic,abduction,compilationmeth- set of defaults), a set M of observed phenomena, ods, problem solving, quanti(cid:12)ed Boolean logic andasetH ofhypotheses.Anexplanation forP is any set E (cid:18) H such that M can be inferred from theextendeddefaulttheoryT =hW[E;(cid:1)i,pro- E 1. Introduction viding T has a consistent extension. Inference in E the context of default logic is understood either The philosopher C. S. Peirce identi(cid:12)ed three as brave inference (checking whether M is con- fundamental modes of reasoning: (i) deduction, tained in some extension of T ) or skeptical in- E an analytic process determining necessary conse- ference (checking whether M is contained in all quences from general rules; (ii) induction, a syn- extensions of T ). E theticformofreasoningderivingprobablegeneral- Although a restriced fragment of this approach isations from factual data; and (iii) abduction, an- has been realised by the diagnosis front-end of otherformofsyntheticinference,yieldingrelevant the DLV system [8], no general solver for default premisses from rules and observed consequences. abduction has been put forth so far. In this pa- Thus, following Peirce, abduction can be charac- per, we take up this challenge and describe a gen- terised as the \probational adoption of a hypoth- eralmethodtobuildaprototypereasoningsystem AICommunications ISSN0921-7126,IOSPress.Allrightsreserved 2 H.Tompits/ExpressingDefaultAbductionProblemsasQuanti(cid:12)edBooleanFormulas for the default abduction model from [12], based Reduction methods to QBFs naturally gener- on a reduction approach. The central idea is to alise similar approaches for problems in NP; these translate a given abduction task into a quanti(cid:12)ed latter problems can in turn be solved by translat- Boolean formula (QBF) and then applying some ing them (in polynomial time) to sat, the satis(cid:12)- sophisticated QBF-solver to evaluate the trans- abilityproblemofclassicalpropositionallogic(see latedQBF.Theexistenceofe(cid:14)cientQBF-solvers, e.g., [21] for such an application in Arti(cid:12)cial In- like,e.g.,thesystemsdevelopedbyCadolietal.[3], telligence). Besides the implementation of di(cid:11)er- Giunchiglia et al. [15], Rintanen [35], Letz [23], ent nonmonotonic reasoning tasks as realised by or Feldmann et al. [14], makes this reduction ap- the system QUIP, successful applications based on proach practicably applicable. reductionstoQBFshavealsobeenappliedtocon- Concerningtheparticularencodings,weprovide ditional planning [34]. e(cid:14)cient (polynomial-time constructible) transla- The paper is organised as follows. In the next tions of the following reasoning tasks into QBFs: section, we give some basic notation and recapitu- latetherelevantaspectsofdefaultlogic.Section3 (i) computingallexplanationsforagivenabduc- introduces the default abduction framework, and tion problem; Section 4 lays down the elementary facts about (ii) computing all hypotheses which are relevant QBFs. Section 5 contains our main results, and for a given abduction problem, i.e., which Section 6 deals with implementational issues. Sec- contribute to some explanation; and tion 7 supplies some concluding remarks. (iii) computingallhypotheseswhicharenecessary for a given abduction problem, i.e., which contribute to all explanations. 2. Preliminaries Eachoftheabovetasksisdealtwithforbraveand skeptical default inference, as well as for arbitrary 2.1. Basic Notation explanationsandforsubset-minimalexplanations. From a theoretical point of view, the rationale We deal with propositional languages and use of the current approach relies on the observation the primitive sentential connectives : and ^, to- that the evaluation problem of quanti(cid:12)ed Boolean gether with the logical constant >, to construct formulas, qsat, is PSPACE-complete, so any de- formulas in the standard way. The operators _, cision problem in PSPACE can be polynomially (cid:27), (cid:17), as well as the symbol ?, are de(cid:12)ned from reduced to qsat. In fact, the evaluation problem :, ^, and > as usual. We write L to denote a A for QBFs having prenex normal form with i(cid:0)1 languageoveranalphabetAofpropositional vari- quanti(cid:12)eralternationsiscompleteforthei-thlevel ables or atoms. Formulas are denoted by Greek of the polynomial hierarchy. Since the reasoning lower-case letters (possibly with subscripts). Con- tasksconsideredinthispaperarelocatedbetween junctionsofform (cid:30) areassumedtostandfor i2I i the second and fourth level of the polynomial hi- the logical constanVt > whenever I =;. A literal is erarchy, e(cid:14)cient translations to QBFs with one, eitheranatomporanegatedatom:p.Thesetof two, or three quanti(cid:12)er alternations must exist. all atoms occurring in a formula (cid:30) is denoted by A similar approach for solving various reason- Var((cid:30)). Similarly, for a set S of formulas, Var(S) ing tasks belonging to the area of nonmonotonic is the set of all atoms occurring in elements of S, reasoninghasbeenrealisedinthesystemQUIP [6; i.e., Var(S)= Var((cid:30)). Furthermore, :S de- (cid:30)2S 13; 4; 2; 27]. This prototype implementation cur- notes the set fS:(cid:30)j(cid:30)2Sg. rently handles the computation of the main rea- The (propositional) derivability operator, ‘, is soning tasks for default logic, several modal non- de(cid:12)ned in the usual way, and likewise its seman- monotoniclogics,consistency-basedapproachesto tic counterpart j=. The deductive closure of a set belief revision and paraconsistent reasoning, and S (cid:18) L of formulas is given by Cn (S) = f(cid:30) 2 A A equilibrium logic, a generalisation of the stable L j S ‘ (cid:30)g. We say that S is deductively closed A model semantics for logic programs. We obtain a i(cid:11) S =Cn (S). Furthermore, S is consistent pro- A straightforwardimplementationofthetranslations viding ? 2= Cn (S). If the language is clear from A for default abduction problems by appeal to the the context, we usually drop the index \A" from system QUIP. Cn ((cid:1)) and simply write Cn((cid:1)). A H.Tompits/ExpressingDefaultAbductionProblemsasQuanti(cid:12)edBooleanFormulas 3 Given an alphabet A, we de(cid:12)ne a disjoint al- one of :(cid:12) ;:::;:(cid:12) holds). We call (cid:11) the prereq- 1 n phabet A0 as A0 = fp0 j p 2 Ag: Accordingly, for uisite, the formulas (cid:12) ;:::;(cid:12) the justi(cid:12)cations, 1 n (cid:11) 2 L , we de(cid:12)ne (cid:11)0 as the result of replacing in and (cid:13) the consequent of (cid:14). For notational conve- A (cid:11) each atom p from A by the corresponding atom nience, we also write p((cid:14)) for the prerequisite (cid:11) p0 in A0 (so implicitly there is an isomorphism be- of (cid:14), j((cid:14)) for the set f(cid:12) ;:::;(cid:12) g, and c((cid:14)) for 1 n tween A and A0). This is de(cid:12)ned analogously for the consequent (cid:13) of (cid:14). Accordingly, p((cid:1)) is the sets of formulas. set of prerequisites of all default rules in (cid:1); j((cid:1)) We assume the reader familiar with the basic andc((cid:1))arede(cid:12)nedanalogously.Furthermore,we concepts of complexity theory (see, e.g., [26] for de(cid:12)ne Var((cid:1)) = Var(p((cid:1)) [ j((cid:1)) [ c((cid:1))) and a comprehensive introduction). Relevant for our Var(T) = Var(W)[Var((cid:1)), for T = hW;(cid:1)i. As purposes are the elements of the polynomial hier- well, we extend our priming notation introduced archy, given by the following sequence of classes: for (sets of) formulas to defaults and default the- ories in the obvious way, i.e., (cid:14)0 denotes the result (cid:6)P =(cid:5)P =P; 0 0 of replacing each atom p in default (cid:14) by p0, and and, for i>0, likewise for sets of defaults and default theories. WesaythatadefaulttheoryT =hW;(cid:1)iis(cid:12)nite (cid:6)P =NP(cid:6)Pi(cid:0)1 and (cid:5)P =co-NP(cid:6)Pi(cid:0)1: i i i(cid:11) both W and (cid:1) are (cid:12)nite. Here, P is the class of all problems solvable on a For simplicity, defaults will also be written in deterministic Turing machine in polynomial time; the form ((cid:11):(cid:12)1;:::;(cid:12)n=(cid:13)). NP is similarly de(cid:12)ned but using a nondeter- The semantics of default theories is de(cid:12)ned in ministic Turing machine as underlying computing terms of extensions. Formally, a set S (cid:18) LA is model; and, for complexity classes C and A, the an extension of a default theory T = hW;(cid:1)i i(cid:11) notation CA stands for the relativized version of S = i(cid:21)0Si, where S0 =W and, for i(cid:21)0, C,consistingofallproblemswhichcanbedecided S S =Cn(S )[fc((cid:14))j(cid:14) 2GD g; i+1 i i by Turing machines of the same sort and time boundasinC,onlythatthemachineshaveaccess with to an oracle for problems in A. As well, co-C is GD =f(cid:14) 2(cid:1)jp((cid:14))2S ;:j((cid:14))\S =;g: i i theclassofallproblemswhicharecomplementary to the problems in C. We note that (cid:6)P = NP, In general, a default theory may possess none, 1 (cid:6)P =NPNP, and (cid:5)P =co-NPNP. one,orseveralextensions.Intuitively,anextension 2 2 characterises a possible totality of beliefs to which 2.2. Default Logic an agent may refer on the basis of a given default theory. In default logic [33], knowledge about the world Setting GD((cid:1);S) = i(cid:21)0GDi; it holds that is represented in terms of default theories, which S =Cn(W [c(GD((cid:1);S)S)),foranyextensionS of are ordered pairs of form T =hW;(cid:1)i, where W is T =hW;(cid:1)i.WecalltheelementsofGD((cid:1);S)the a set propositional formulas, called the premisses generating defaults of S. of T, and (cid:1) is a set of default rules (or defaults, Following Marek and Truszczyn(cid:19)ski [24], exten- forshort).ThesetW representscertain(thoughin sions can alternatively be characterised as follows. general incomplete) information about the given For any set (cid:1) of defaults and any S (cid:18) LP, de- applicationdomain,whilst(cid:1)representsdefeasible (cid:12)ne the reduct of (cid:1) with respect to S as the set of knowledge,whichmaybeinvalidatedbynew,more classical inference rules accurate information. Formally, a default, (cid:14), is an p((cid:14)) (cid:1) = (cid:14) 2(cid:1);:j((cid:14))\S =; : inference rule of form S (cid:26)c((cid:14)) (cid:12) (cid:27) (cid:12) (cid:11) : (cid:12) ;:::;(cid:12) (cid:12) 1 n; Furthermore, for (cid:12)any set F of propositional for- (cid:13) mulas and any set K of classical inference rules, where (cid:11);(cid:12) ;:::;(cid:12) ; and (cid:13) are formulas. Intu- let CnK(F) be the set of all formulas derivable 1 n itively, (cid:14) expresses that (cid:13) is asserted whenever fromF usingclassicallogictogetherwiththerules (cid:11) is believed and (cid:12) ;:::;(cid:12) are consistent with in K. Then, S is an extension of T = hW;(cid:1)i i(cid:11) 1 n what is believed (i.e., there is no evidence that S =Cn(cid:1)S(W). 4 H.Tompits/ExpressingDefaultAbductionProblemsasQuanti(cid:12)edBooleanFormulas There are two basic inference operations in the WecallE a braveexplanationforP i(cid:11)(i)hW[ context of default logic, viz. brave inference and E;(cid:1)i (cid:24) M and (ii) hW[E;(cid:1)i has a consistent b skeptical inference. To wit, we say that a formula extension. Similarly, E is a skeptical explanation (cid:30) is a brave consequence of a given default the- forP i(cid:11)(i)hW[E;(cid:1)i (cid:24) M and(ii)hW[E;(cid:1)i s ory T, symbolically T (cid:24) (cid:30), i(cid:11) (cid:30) belongs to has a consistent extension. b some extension of T, and (cid:30) is a skeptical conse- quence of T, symbolically T (cid:24) (cid:30), i(cid:11) (cid:30) belongs The stipulation that hW [ E;(cid:1)i possesses a s to all extensions of T. As shown in [16], both in- consistent extension assures that explanations are ference tasks are intractable in the general case. consistent with the knowledge represented in the More speci(cid:12)cally, given a default theory T and a given default theory hW;(cid:1)i. This requirement is formula (cid:30), checking whether T (cid:24) (cid:30) holds is (cid:6)P- similar to the usual consistency condition in ab- b 2 complete, whilst checking whether T (cid:24) (cid:30) holds duction from theories in classical logic. s is (cid:5)P-complete. Analogously, checking whether a We note that brave explanations for a given ab- 2 given default theory possesses an extension is (cid:6)P- duction problem P = hH;M;W;(cid:1)i can be equiv- 2 alently characterised as sets E (cid:18) H such that complete as well. Let S be a set of formulas. We write T (cid:24) S hW [E;(cid:1)i has a consistent extension containing b to indicate that T (cid:24) (cid:30), for any (cid:30)2S; a similar M.Thisisasimpleconsequenceofthewell-known b notation applies for (cid:24) . fact that if a default theory T has some consis- s tent extension, then all extensions of T must be consistent. The assumption that hypotheses and mani- 3. The Abduction Framework festations in a default abduction problem P = hH;M;W;(cid:1)i are given by literals instead of ar- FollowingEiter,Gottlob,andLeone[12],wede- bitrary formulas is no restriction since for each (cid:12)ne a formal model for abduction from proposi- non-literal hypothesis or manifestation (cid:30) a new tional default theories as follows. propositional atom p can be introduced, and af- (cid:30) ter adding the formula p (cid:17) (cid:30) to W, (cid:30) can be De(cid:12)nition 1 A defaultabductionproblem, P, is a (cid:30) equivalently replaced by p . quadruple hH;M;W;(cid:1)i, where H and M are sets (cid:30) We extend our notation Var((cid:1)) and our priming of literals, and hW;(cid:1)i is a default theory. We call convention to default abduction problems in the the elements of H the hypotheses or abducibles, obvious way, i.e., for P =hH;M;W;(cid:1)i, we de(cid:12)ne and the elements of M the observations or mani- Var(P) = Var(H [M [W)[Var((cid:1)) and P0 = festations of P. hH0;M0;W0;(cid:1)0i. The default abduction problem P is (cid:12)nite i(cid:11) Following Occam’s principle of parsimony, one each constituting member H, M, W, and (cid:1) of P usually prefers simpler explanations over more is (cid:12)nite. complicatedones.Inthepresentcontext,thisidea Informally, in a default abduction problem P = is realised by accepting only those explanations hH;M;W;(cid:1)i, H speci(cid:12)es the universe of admis- which are minimal among the class of explana- sible explanations, the elements of M are the ob- tions. More formally, we call E (cid:18) H a minimal servedphenomena,andhW;(cid:1)irepresentsthecur- brave (resp., minimal skeptical) explanation for P rent world knowledge. A set E (cid:18) H serves as i(cid:11)E isabrave(resp.,skeptical)explanationforP an explanation for the observed manifestations in and there is no F (cid:26)E such that F is also a brave M if the adjunction of E to the world knowl- (resp., skeptical) explanation for P. edge hW;(cid:1)i entails all propositions in M. Given Another interesting property of hypotheses is that there are two di(cid:11)erent inference operations whether they contribute to some explanation or associatedwithdefaulttheories,thereare,accord- toall explanationsforagivenabductionproblem. ingly, two fundamental kinds of explanations for More speci(cid:12)cally, we call a set H0 (cid:18) H relevant observed phenomena, as de(cid:12)ned next. for P = hH;M;W;(cid:1)i under brave (resp., skepti- cal)explanations i(cid:11)H (cid:18)E forsomebrave(resp., 0 De(cid:12)nition 2 LetP =hH;M;W;(cid:1)ibeadefaultab- skeptical) explanation E for P. Dually, H (cid:18)H is 0 duction problem, and let E (cid:18)H. necessary for P under brave (resp., skeptical) ex- H.Tompits/ExpressingDefaultAbductionProblemsasQuanti(cid:12)edBooleanFormulas 5 Table1 Complexityresultsofabductionfromdefaulttheories. AbductionProblemP arbitraryexplanations minimalexplanations DecisionProblem brave skeptical brave skeptical consistency (cid:6)P2 (cid:6)P3 (cid:6)P2 (cid:6)P3 relevance (cid:6)P2 (cid:6)P3 (cid:6)P3 (cid:6)P4 necessity (cid:5)P2 (cid:5)P3 (cid:5)P2 (cid:5)P3 planations i(cid:11) H (cid:18) E for each brave (resp., skep- Abstracting from the particular kind of expla- 0 tical)explanationE forP.Bothconceptsaresim- nation, the main decision problems in the context ilarly de(cid:12)ned for abduction problems under mini- of abductive reasoning are the following: mal explanations. consistency: Does a given (cid:12)nite default abduc- For illustration, let us consider the following tion problem possess an explanation? simple example. relevance: Given some (cid:12)nite default abduction problemP =hH;M;W;(cid:1)iandasetH (cid:18)H, Example 1 Let hW;(cid:1)i be the default theory con- 0 is H relevant for P? sisting of the following items: 0 necessity: Given some (cid:12)nite default abduction problemP =hH;M;W;(cid:1)iandasetH (cid:18)H, W = l ^ (s _ (c ^ r)) (cid:27) m; 0 is H necessary for P? (cid:8)(cid:0) (cid:1) 0 (:s _ :c) ; Concerning the computational complexity of (cid:9) D= (w ::j=r); (w ::r=j) : these tasks, as shown in [12], these problems are (cid:8) (cid:9) locatedbetweenthesecondandfourthlevelofthe Furthermore, assume we are interested in (cid:12)nding polynomial hierarchy. Table 1 gives the speci(cid:12)c explanations for the manifestation M = fmg us- results; each entry C represents completeness of ingH =fl;s;c;wgasthesetofhypotheses.Then, the corresponding problem for the class C.1 As by straightforward calculations, it can be shown canbeseenfromtheseresults,skepticalabduction that the brave explanations for P =hH;M;W;(cid:1)i is always one level harder than brave abduction. are given by fl;sg, fl;s;wg, and fl;c;wg, whilst This contrasts with complexity results for usual only fl;sg and fl;s;wg represent the skeptical ex- nonmonotonicreasoningformalisms,whereskepti- planations for P. The reason for this di(cid:11)erence cal reasoning has complementary complexity than is that m is not contained in all extensions of brave reasoning. Also, minimality of explanations hW [fl;c;wg;(cid:1)i, which are given by is a source of complexity for relevance but nei- therforconsistencynorfornecessity.Therea- S1=Cn(fl;c;w;r;mg) and sonforthisphenomenonisthefactthatanabduc- tion problem P has a minimal brave (resp., mini- S =Cn(fl;c;w;jg): 2 malskeptical)explanationi(cid:11)ithasabrave(resp., Observe that flg is the only set necessary for skeptical) explanation; likewise, a set is necessary P under brave explanations, whilst fl;sg is neces- for P under minimal brave (resp., minimal skep- saryforP underskepticalexplanations.Moreover, tical) explanations i(cid:11) it is necessary for P under fcgisrelevantforP underbraveexplanations but brave (resp., skeptical) explanations. Finally, the not under skeptical ones. Also, fl;sg and fl;c;wg 1Strictlyspeaking,[12]de(cid:12)nesrelevantandnecessaryhy- aretheminimalbraveexplanationsforP,whereas potheses as single formulas (actually, literals) whereas we fl;sg is the single minimal skeptical explanation allow here sets for this purpose; but this di(cid:11)erence is im- for P. 2 material. 6 H.Tompits/ExpressingDefaultAbductionProblemsasQuanti(cid:12)edBooleanFormulas results for relevance contrast with correspond- 1. if (cid:8)=>, then (cid:23) ((cid:8))=1; I ing results for abduction from theories in classical 2. if (cid:8) = p, for some atom p, then (cid:23) ((cid:8)) = 1 if I logic where minimality is not a source of complex- p2I, otherwise (cid:23) ((cid:8))=0; I ity [10]. 3. if (cid:8) = :(cid:9), then (cid:23) ((cid:8)) = 1 if (cid:23) ((cid:9)) = 0, I I The main contribution of this paper is to show otherwise (cid:23) ((cid:8))=0; I howeachoftheabovedecisionproblems,aswellas 4. if (cid:8)=((cid:8) ^(cid:8) ), then (cid:23) ((cid:8))=1 if (cid:23) ((cid:8) )= 1 2 I I 1 eachcorrespondingsearchproblem,canbemapped (cid:23) ((cid:8) )=1, otherwise (cid:23) ((cid:8))=0; I 2 I in polynomial time into a quanti(cid:12)ed Boolean for- 5. if (cid:8) =8p(cid:9), then (cid:23) ((cid:8)) =1 if (cid:23) ((cid:9)[p=>]) = I I mula such that the models of the latter determine (cid:23) ((cid:9)[p=?])=1, otherwise (cid:23) ((cid:8))=0. I I theanswersoftheformer.Infact,thetranslations Thetruthconditionsfor?, _, (cid:27),(cid:17),and9follow arerealisedinsuchawaythatthestructureofthe from the above in the usual way. Obviously, we constructed formulas precisely matches the com- have that plexityoftheencodedreasoningtasks.Inthenext section,webrie(cid:13)yrecapitulatetherelevantaspects (cid:23) (8p(cid:9))=(cid:23) ((cid:9)[p=>] ^ (cid:9)[p=?]) and I I of quanti(cid:12)ed Boolean logic. (cid:23) (9p(cid:9))=(cid:23) ((cid:9)[p=>] _ (cid:9)[p=?]): I I We say that (cid:8) is true under I i(cid:11) (cid:23) ((cid:8)) = 1, I 4. Quanti(cid:12)ed Boolean Logic otherwise(cid:8)isfalseunderI.If(cid:23) ((cid:8))=1,thenI is I amodel of(cid:8).Thesetofallmodelsof(cid:8)isdenoted Quanti(cid:12)edBoolean logicisanextensionofclas- by Mod((cid:8)). If Mod((cid:8)) 6= ;, then (cid:8) is satis(cid:12)able. sicalpropositionallogicinwhichformulasareper- If (cid:8) is true under every interpretation, then (cid:8) is mitted to contain quanti(cid:12)cations over proposi- valid. As usual, we write j=(cid:8) to express that (cid:8) is tional variables. More formally, the language of valid. quanti(cid:12)ed Boolean logic comprises the symbols of It is easily seen that the truth value of a QBF propositional logic, together with unary operators (cid:8) under interpretation I depends only on the free of form 8p (where p is some atom), called univer- variables in (cid:8). Hence, without loss of generality, sal quanti(cid:12)ers. Similar to (cid:12)rst-order logic, 9p is for determining the truth value of QBFs, we may de(cid:12)ned as the operator :8p: and is called an ex- restrictourattentiontointerpretationswhichcon- istential quanti(cid:12)er. Formulas of this language are tain only atoms occurring free in the given QBF. referredtoasquanti(cid:12)ed Boolean formulas (QBFs) In particular, closed QBFs are either true under and are denoted by Greek upper-case letters. every interpretation or false under every interpre- Informally, a QBF of form 8p9q(cid:8) means that tation, i.e., they are either valid or unsatis(cid:12)able. for all truth assignments of p there is a truth as- So, for closed QBFs there is no need to refer to signment of q such that (cid:8) is true. For instance, particularinterpretations.Aswell,ifaclosedQBF under this reading, it is easily seen that the QBF (cid:8) is valid, we say that (cid:8) evaluates to true, and, 9p9q((p (cid:27) q) ^ 8r(r (cid:27) q)) evaluates to true. correspondingly, if (cid:8) is unsatis(cid:12)able, we say that The precise semantical meaning of QBFs is de- (cid:8) evaluates to false. Two sets of formulas (i.e., or- (cid:12)nedasfollows.First,someancillarynotation.An dinary propositional formulas or QBFs) are logi- occurrence of a propositional variable p in a QBF cally equivalent i(cid:11) they possess the same models. (cid:8) is free i(cid:11) it does not appear in the scope of a Thus,formulas(cid:8)and(cid:9)arelogicallyequivalenti(cid:11) quanti(cid:12)er Qp (Q 2 f8;9g), otherwise the occur- (cid:8)(cid:17)(cid:9) is valid. rence of p is bound. If (cid:8) contains no free variable Inthesequel,weusethefollowingabbreviations occurrences, then (cid:8) is closed, otherwise (cid:8) is open. in the context of QBFs: Let R = f(cid:30) ;:::;(cid:30) g 1 n Furthermore, (cid:8)[p =(cid:30) ;:::;p =(cid:30) ] denotes the re- and S = f ;:::; g be indexed sets of formu- 1 1 n n 1 n n sultofuniformlysubstitutingin(cid:8)eachfreeoccur- las. Then, R(cid:20)S abbreviates ((cid:30) (cid:27) ), and i=1 i i renceofavariablepibyaformula(cid:30)i,for1(cid:20)i(cid:20)n. R < S is a shorthand for (RV(cid:20) S) ^ :(S (cid:20) R). By an interpretation, I, we understand a set of Furthermore, for an indexed set P = fp ;:::;p g 1 n atoms. Informally, an atom p is true under I i(cid:11) of variables and Q2f8;9g, we let QP (cid:8) stand for p2I.Ingeneral,thetruthvalue,(cid:23) ((cid:8)),ofaQBF the formula Qp Qp (cid:1)(cid:1)(cid:1)Qp (cid:8). To ease notation, I 1 2 n (cid:8) under an interpretation I is recursively de(cid:12)ned we usually identify a (cid:12)nite set of formulas with as follows: the conjunction of its elements. Finally, the logi- H.Tompits/ExpressingDefaultAbductionProblemsasQuanti(cid:12)edBooleanFormulas 7 cal complexity of a formula (cid:8), or the degree of (cid:8), 5.1. Preparatory Characterisations symbolicallyd((cid:8)),isthenumberofoccurrencesof the primitive operators :, ^, and 8p (where p is We start with some basic results concerning the some variable) occurring in (cid:8). operator (cid:20). To wit, we (cid:12)rst describe a particular The operators (cid:20) and < are fundamental tools substitution theorem involving (cid:20), and afterwards for expressing certain tests on sets of atoms. In we discuss two basic QBF modules required sub- particular, the following properties hold: Let P = sequently. fp ;:::;p g be a set of indexed atoms, and let 1 n I ;I (cid:18)P be two interpretations. Then, 1 2 Proposition 2 Let H = f(cid:30) ;:::;(cid:30) g and G = 1 n (i) I (cid:18)I i(cid:11) I0 [I is a model of P0 (cid:20)P; and fg ;:::;g gbeindexedsetsofformulasandatoms, 1 2 1 2 1 n (ii) I (cid:26)I i(cid:11) I0 [I is a model of P0 <P. respectively, such that G \ Var(H) = ;, and let 1 2 1 2 (cid:8) be a QBF containing one or more desig- We say that QBF (cid:8) is in prenex form if (cid:8) = G(cid:20)H Q p :::Q p (cid:30), where Q 2 f8;9g, for 1 (cid:20) i (cid:20) natedoccurrencesofG(cid:20)H suchthattheelements 1 1 n n i of G occur free in (cid:8) and, moreover, only in n, and (cid:30) is some propositional formula. Using G(cid:20)H quanti(cid:12)er-shiftingrulessimilartothoseofclassical the designated subformulas G(cid:20)H of (cid:8)G(cid:20)H. Fur- (cid:12)rst-orderlogic,anyQBFcanbee(cid:11)ectivelytrans- thermore, let I (cid:18)Var((cid:8)G(cid:20)H)nG and J (cid:18)G, and formed into an equivalent QBF in prenex form. In let HJ = gi2J(cid:30)i. fact, this transformation can be done in polyno- Then, (cid:8)VG(cid:20)H is true under I[J i(cid:11) (cid:8)HJ is true mial time. under I, where (cid:8)HJ comes from (cid:8)G(cid:20)H by replac- QBFs are closely related to the constituting ing all designated occurrences of G(cid:20)H by H . J membersofthepolynomialhierarchy,asdescribed by the following well-known result: Proof.Foraformula(cid:9)(cid:11) containingdesignatedoc- currences of a formula (cid:11) as consecutive parts, de- Proposition 1 ([40]) Givenapropositionalformula (cid:12)ne d0((cid:9) ) like d((cid:9) ), but count the speci(cid:12)c oc- (cid:30) whose atoms are partitioned into i (cid:21) 1 sets (cid:11) (cid:11) V ;:::;V , deciding whether 9V 8V 9V :::QV (cid:30) currencesof(cid:11)asatomicformulas.Theproofofthe ev1aluatesito true is (cid:6)P-complete,1wh2ere3Q = 9iif proposition proceeds by induction on d0((cid:8)G(cid:20)H). i i is odd and Q = 8 if i is even. Dually, decid- ing whether 8V 9V 8V :::Q(cid:22)V (cid:30) evaluates to true Induction Base. Assume d0((cid:8) ) = 0. Then, 1 2 3 i G(cid:20)H is (cid:5)P-complete, where Q(cid:22) =8 if i is odd and Q(cid:22) =9 (cid:8) = G (cid:20) H and (cid:8) = H . Since, for each i G(cid:20)H HJ J if i is even. 1(cid:20)i(cid:20)n, the truth value of g (cid:27) (cid:30) under inter- i i From this proposition it follows that any deci- pretation I [J coincides with the truth value of sion problem in (cid:6)P can be e(cid:14)ciently reduced to (cid:30)i under I if gi 2 J, and gi (cid:27) (cid:30)i is trivially true i the evaluation problem of prenex QBFs of form under I[J if gi 2= J, we have that 9V18V29V3:::QVi(cid:30), with Q as above, and sim- n ilarly for problems in (cid:5)P. In the next section, (cid:23) ( (g (cid:27) (cid:30) ))=(cid:23) ( (cid:30) ); i I[J i i I i we describe polynomial-time constructible trans- i^=1 g^i2J lations from the main abductive reasoning tasks whichprovesthat(cid:8) istrueunderI[J i(cid:11)H intoQBFssuchthatthequanti(cid:12)erstructureofthe G(cid:20)H J is true under I. resultant formulas precisely matches the inherent complexity of the encoded problem. Induction Step. Assume d0((cid:8) )>0, and let G(cid:20)H the statement hold for all formulas (cid:9) , con- G(cid:20)H 5. Translations taining speci(cid:12)ed occurrences of G(cid:20)H, such that d0((cid:9) )<d0((cid:8) ). Inthissection,wepresente(cid:14)cientreductionsof G(cid:20)H G(cid:20)H Since d0((cid:8) )>0, (cid:8) is a compound for- the main reasoning tasks in the context of abduc- G(cid:20)H G(cid:20)H tionfromdefaulttheoriesintoQBFs.Thesereduc- mula having as its main operator either :, ^, or tions are constructed in such a way that the solu- a universal quanti(cid:12)er (recall that _, (cid:27), (cid:17), and tions of the given abduction problems are deter- existential quanti(cid:12)ers are de(cid:12)ned operators). We mined by the models of the corresponding QBFs. only deal with the case where (cid:8)G(cid:20)H is a univer- In what follows, we tacitly assume that all con- sallyquanti(cid:12)edformula;theremainingcasesfollow sidered default abduction problems are (cid:12)nite. by similar arguments. 8 H.Tompits/ExpressingDefaultAbductionProblemsasQuanti(cid:12)edBooleanFormulas Assume therefore that (cid:8)G(cid:20)H = 8p(cid:9)G(cid:20)H. Re- 9U R ^ (cid:30)i (1) call that (cid:23)I[J(8p(cid:9)G(cid:20)H) is given by (cid:16) g^i2I (cid:1) (cid:23)I[J((cid:9)G(cid:20)H[p=>] ^ (cid:9)G(cid:20)H[p=?]): isvalid.ButQ= gi2I(cid:30)i,hence,bythesemantics ofexistentialquanVti(cid:12)cation,wegetthat(1)isvalid Sincebothd0((cid:9)G(cid:20)H[p=>])andd0((cid:9)G(cid:20)H[p=?])are i(cid:11) R[Q is satis(cid:12)able. It follows that I is a model strictly less than d0((cid:8)G(cid:20)H), by induction hypoth- of C[R;S;G] i(cid:11) R[Q is satis(cid:12)able. esis it follows that Part 2 is an immediate consequence of Part 1, by observing that R[Q j= i(cid:11) (R[f: g)[Q (cid:23) ((cid:9) [p=>])=(cid:23) ((cid:9) [p=>]) and I[J G(cid:20)H I HJ is unsatis(cid:12)able, and since (cid:23) ((cid:9) [p=?])=(cid:23) ((cid:9) [p=?]); I[J G(cid:20)H I HJ :9V (R ^ : ) ^ (G(cid:20)S) from which we immediately get, by the semantics is equival(cid:0)ent to (cid:1) of conjunction and universal quanti(cid:12)cation, that 8V (R ^ (G(cid:20)S)) (cid:27) . (cid:23)I[J(8p(cid:9)G(cid:20)H)=(cid:23)I(8p(cid:9)HJ); (cid:0) (cid:1) Note that C[R;S;G] is an open QBF having G i.e., (cid:8) is true under I [J i(cid:11) (cid:8) is true un- G(cid:20)H HJ asitssetoffreevariables.Thesevariablesfacilitate der I. the selection of those elements of S which deter- minethesetsQ(cid:18)S suchthatR[Qissatis(cid:12)able. Next, we describe two fundamental QBF mod- Infact,C[R;S;G]isdesignedtoexpressall poten- ules, capturing the following tasks, respectively: tial subsets Q (cid:18) S such that R[Q is satis(cid:12)able. Similar considerations apply to D[R;S; ;G]. 1. Given(cid:12)nitesetsRandS ofpropositionalfor- mulas, compute all Q(cid:18)S such that R[Q is 5.2. Encodings for Default Logic satis(cid:12)able. 2. Given (cid:12)nite sets R and S of propositional In order to express default abduction problems formulas, and some propositional formula , in terms of QBFs, we need suitable QBF encod- compute all Q(cid:18)S such that R[Qj= . ings capturing the underlying default inference mechanism. To this end, we use results from [6], These tasks can be characterised as follows. where QBF reductions for default reasoning are given. More speci(cid:12)cally, that paper contains two Proposition 3 Let R and S = f(cid:30) ;:::;(cid:30) g be (cid:12)- 1 n di(cid:11)erent kinds of encodings for default logic: one nite sets of propositional formulas, let be a translation is based on the characterisation of propositional formula, and let G=fg ;:::;g g be 1 n extensions following the method of Marek and a set of new variables. Furthermore, consider any Truszczyn(cid:19)ski [24], and the other translation ex- Q(cid:18)S and any I (cid:18)G such that (cid:30) 2Q i(cid:11) g 2I, i i ploits the so-called full-set characterisation of ex- for 1(cid:20)i(cid:20)n. tensions, due to Niemela(cid:127) [25]. Here, we adopt the Then, former method because it is more suitable for our 1. R[Qissatis(cid:12)ablei(cid:11)I isamodeloftheQBF purposes. Thefollowingresultcontainstherequiredreduc- C[R;S;G]=9U R ^ (G(cid:20)S) ; tions. (cid:0) (cid:1) whereU =Var(R[S)andsuchthatG\U = Proposition 4 ([6; 39]) Let T = hW;(cid:1)i be a de- ;; and fault theory with (cid:1) = f(cid:14) ;:::;(cid:14) g, let (cid:30) be a 1 n 2. R[Qj= i(cid:11) I is a model of the QBF propositional formula, and let C = fc ;:::;c g, 1 n D = fd ;:::;d g, and D0 = fd0;:::;d0 g be sets D[R;S; ;G]=8V (R ^ (G(cid:20)S)) (cid:27) ; 1 n 1 n of pairwise distinct new variables. (cid:0) (cid:1) where V = Var(R[S [f g) and such that Furthermore, consider the QBFs from Fig. 1 G\V =;. with V = Var(W [ c((cid:1))), P = Var((cid:30)) n V, U = Var(p((cid:14) ))nV, and Z = Var((cid:12) )nV, for i i i;j j Proof. Consider Q (cid:18) S and I (cid:18) G such that (cid:30)i 2 (cid:12)j 2j((cid:14)i) and 1(cid:20)i(cid:20)n. Q i(cid:11) gi 2 I, for 1 (cid:20) i (cid:20) n. For proving Part 1, Then, for any F (cid:18)(cid:1) and any I (cid:18)D such that according to Proposition 2, we have that I is a (cid:14) 2 F i(cid:11) d 2 I, for 1 (cid:20) i (cid:20) n, the following i i model of 9U R ^ (G(cid:20)S) i(cid:11) conditions hold: (cid:0) (cid:1) H.Tompits/ExpressingDefaultAbductionProblemsasQuanti(cid:12)edBooleanFormulas 9 E[W;(cid:1);D]=9D0((D (cid:20)D0) ^ (cid:8) ^ (cid:8) ^ (cid:8) ); 1 2 3 M [W;(cid:1);(cid:30);D]=E[W;(cid:1);D] ^ 8V 8P W ^ (D (cid:20)c((cid:1))) (cid:27) (cid:30) ; b (cid:16)(cid:0) (cid:1) (cid:17) M [W;(cid:1);(cid:30);D]=E[W;(cid:1);D]^9V 9P W ^ (D (cid:20)c((cid:1))) ^ :(cid:30) ; s (cid:16) (cid:17) n (cid:8) = d0 (cid:17) 9V 9Z W ^ D (cid:20)c((cid:1)) ^ (cid:12) ; 1 i i;j j i^=1h (cid:12)j2^j((cid:14)i) (cid:16) (cid:0) (cid:1) (cid:17)i n (cid:8) = (d0 ^ :d ) (cid:27) 9V 9U W ^ D (cid:20)c((cid:1)) ^ :p((cid:14) )) ; 2 i i i i i^=1h (cid:16) (cid:0) (cid:1) (cid:17)i (cid:8) =8C (C (cid:20)D0) (cid:27) 8V W ^ (C (cid:20)c((cid:1))) (cid:27) (D (cid:20)c((cid:1))) _ 3 n h (cid:16)(cid:0) (cid:1) (cid:17) n d0 ^:c ^8V 8U W ^ (C (cid:20)c((cid:1))) (cid:27) p((cid:14) ) : i i i i i_=1(cid:16) (cid:16)(cid:0) (cid:1) (cid:17)(cid:17)io Fig.1.QBFmodulesforexpressingdefaultreasoning. 1. Cn(W [c(F)) is an extension of T i(cid:11) I is a and M [W;(cid:1);(cid:30);D]. More speci(cid:12)cally, the vari- s modelofE[W;(cid:1);D],providing(C[D[D0)\ ables in D0 take care of selecting the members Var(T)=;; in the reduct (cid:1) , and the variables in D check S 2. Cn(W [c(F)) is an extension of T contain- whether the consequent of a selected rule is con- ing (cid:30) i(cid:11) I is a model of M [W;(cid:1);(cid:30);D], pro- tained in S =Cn(cid:1)S(W). To that end, the follow- b viding D\(Var(T)[Var((cid:30))) = (C [D0)\ ingtestsareperformed,expressedbythesubmod- Var(T)=;; and ules (cid:8)1, (cid:8)2, and (cid:8)3: 3. Cn(W [c(F)) is an extension of T not con- { (cid:8) tests whether each default in the guessed 1 taining (cid:30) i(cid:11) I is a model of M [W;(cid:1);(cid:30);D], s set (cid:1) is consistent with the guess for the S assuming the same proviso as in 2. extension S; { (cid:8) tests whether no applicable default in (cid:1) Here, the module E[W;(cid:1);D] characterises the 2 S is missing with respect to the guessed assign- extensionsofthegivendefaulttheoryT =hW;(cid:1)i, ment for D; and and { (cid:8) utilises the test whether, for each default, 3 D[W;c((cid:1));(cid:30);D]= it holds that its consequent is actually con- tained in Cn(cid:1)S(W). 8V 8P W ^ (D (cid:20)c((cid:1))) (cid:27) (cid:30) (cid:16)(cid:0) (cid:1) (cid:17) The latter formula encodes a result due to Got- checks whether a given formula (cid:30) is contained in tlob [17] stating that a formula 2= Cn(cid:1)S(W) i(cid:11) a selected extension. Note that, dually, there exists a subset B (cid:18) f(cid:13) j (cid:11) =(cid:13) 2 (cid:1) g such i i i S that (i) W [B 6j= , and (ii) for each (cid:11) =(cid:13) 2(cid:1) C[W [f:(cid:30)g;c((cid:1));D]= i i S with (cid:13) 2= B, we have that W [B 6j=(cid:11) . i i 9V 9P W ^ (D (cid:20)c((cid:1))) ^ :(cid:30) ; Encodings for brave and skeptical default infer- (cid:16) (cid:17) ence are obtained from Proposition 4 in the fol- which is equivalent to :D[W;c((cid:1));(cid:30);D], checks lowing way. whether (cid:30) is not contained in the selected ex- Corollary 1 Let T, (cid:30), and D be as in Proposi- tension. The selection of an extension S, in turn, tion 4. is done by means of the variables in D, which Then, constitute the set of free variables of each of the three main modules E[W;(cid:1);D], M [W;(cid:1);(cid:30);D], 1. T (cid:24) (cid:30) i(cid:11) j=9DM [W;(cid:1);(cid:30);D]; and b b b 10 H.Tompits/ExpressingDefaultAbductionProblemsasQuanti(cid:12)edBooleanFormulas 2. T (cid:24) (cid:30) i(cid:11) j=:9DM [W;(cid:1);(cid:30);D]. 1. E isabraveexplanationforP i(cid:11)I isamodel s s of the QBF Tb [P;U ], given by cons H We note that 9DM [W;(cid:1);(cid:30);D] can be trans- b formed in polynomial time into a QBF of prenex 9D M+b [W;(cid:1);M;D;UH]^ form 9X 8X , whilst :9DM [W;(cid:1);(cid:30);D] can (cid:16) 1 2 s 9V W ^ (U (cid:20)H) ^ (D (cid:20)c((cid:1))) ; be transformed, in the same manner, into a QBF H ofprenexform8Y 9Y ’.Hence,inviewofPropo- (cid:0) (cid:1)(cid:17) 1 2 and sition1,theseencodingspreciselymatchtheinher- 2. E is a skeptical explanation for P i(cid:11) I is a ent computational complexity of brave and skep- model of the QBF Ts [P;U ], given by tical default reasoning, respectively. cons H 9D E+[W;(cid:1);D;U ]^ H 5.3. Encodings of the Basic Abduction Tasks (cid:16) 9V W ^ (U (cid:20)H) ^ (D (cid:20)c((cid:1))) ^ H (cid:0) (cid:1)(cid:17) The following de(cid:12)nition gives the core modules :9DM+[W;(cid:1);M;D;U ]: s H for our encodings. Proof. For Part 1, recall that E is a brave expla- De(cid:12)nition 3 Let P = hH;M;W;(cid:1)i be a default nation for P i(cid:11) there is a consistent extension of abduction problem with (cid:1) = f(cid:14) ;:::;(cid:14) g, and 1 n hW [E;(cid:1)i containing M. Consider any J (cid:18) D let D = fd ;:::;d g and U = fu j h 2 1 n H h and any F (cid:18) (cid:1) such that di 2 J i(cid:11) (cid:14)i 2 F, for Hg be sets of pairwise distinct variables with 1 (cid:20) i (cid:20) n. We show that Cn(W [E[c(F)) is a (D[UH)\Var(P) = ;. Furthermore, let E[(cid:1);(cid:1);(cid:1)] consistent extension of hW [E;(cid:1)i containing M and M(cid:27)[(cid:1);(cid:1);(cid:1);(cid:1)] be the translations from Proposi- i(cid:11) I[J is a model of tion 4, for (cid:27) 2 fb;sg, and consider the values E[W ^ (UH (cid:20) H);(cid:1);D] and M(cid:27)[W ^ (UH (cid:20) M+b [W;(cid:1);M;D;UH]^ H);(cid:1);M;D]: 9V W ^ (U (cid:20)H) ^ (D (cid:20)c((cid:1))) : H Then, (cid:0) (cid:1) Fromthis,bythesemanticsofexistentialquanti(cid:12)- E+[W;(cid:1);D;UH] cation, Part 1 follows. Firstofall,accordingtoPart1ofProposition3, istheresultofdroppinginE[W ^(U (cid:20)H);(cid:1);D] H we obviously have that Cn(W [E[c(F)) is con- allquanti(cid:12)erswhichbindvariablesfromU .Like- H sistent i(cid:11) I[J is a model of wise, M+[W;(cid:1);M;D;U ] 9V W ^ (UH (cid:20)H) ^ (D (cid:20)c((cid:1))) : (cid:27) H (cid:0) (cid:1) It remains to show that Cn(W [E[c(F)) is an is the result of dropping in formula M [W ^ (cid:27) extension of hW [E;(cid:1)i containing M i(cid:11) I [J is (U (cid:20) H);(cid:1);M;D] all quanti(cid:12)ers which bind H a model of variables from U . H M+[W;(cid:1);M;D;U ]: (2) Observe that the modules E+[W;(cid:1);D;U ] and b H H M+[W;(cid:1);M;D;U ] have D [ U as their sets This can be seen as follows. (cid:27) H H of free variables. In fact, the variables in U will Tobeginwith,observethat,byconstruction,(2) H beusedforselectingthoseelementsfromH which is a QBF of form (cid:8)W^(UH(cid:20)H), resulting from represent a solution to the given abduction prob- Mb[W ^ (UH (cid:20) H);(cid:1);M;D] by dropping all lem. quanti(cid:12)ers which bind variables from UH. Since The next results gives the speci(cid:12)c encodings for E and I are chosen in such a way that E = brave and skeptical explanations. uh2Ih, Proposition 2 implies that I [ J is a Vmodel of (2) i(cid:11) J is a model of (cid:8)W^E, where Theorem 1 Under the circumstances of De(cid:12)ni- the latter formula is the result of replacing all oc- tion 3, and supposing further that V = Var(H [ currences of W ^ (U (cid:20) H) in (cid:8) by H W^(UH(cid:20)H) W [c((cid:1))), the following conditions hold, for any W ^ E. Now, (cid:8) is a QBF di(cid:11)ering from W^E E (cid:18) H and any I (cid:18) U such that h 2 E i(cid:11) M [W ^E;(cid:1);M;D]onlybythepresenceofquan- H b u 2I: ti(cid:12)ersbindingvariablesfromVar(HnE).Butsuch h
Description: