1: \documentstyle[12pt]{article}   

2: \begin{document}

3: \bibliographystyle{plain}

4: \hyphenation{mono-tonicity Mono-tonicity mono-tonic Mono-tonic Mo-notonicity

5: mo-notonicity monoto-nicity Monoto-nicity}

6: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

7: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

8: \newtheorem{theorem}{Theorem}

9: \newtheorem{corollary}{Corollary}

10: \newtheorem{lemma}{Lemma}

11: \newtheorem{exercise}{Exercise}

12: \newtheorem{claim}{Claim}

13: \newtheorem{remark}{Remark}

14: \newtheorem{definition}{Definition}

15: \newtheorem{example}{Example}

16: \newenvironment{notation}{\noindent\bf Notation:\em\penalty1000}{}

17: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

18: \newcommand{\blackslug}{\mbox{\hskip 1pt \vrule width 4pt height 8pt 

19: depth 1.5pt \hskip 1pt}}

20: \newcommand{\QED}{\quad\blackslug\lower 8.5pt\null\par\noindent}

21: \newcommand{\proof}{\par\penalty-1000\vskip .5 pt\noindent{\bf Proof\/: }}

22: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

23: %\newcommand{\SEP}{\makebox[0in]{\rule{.5mm}{4.5mm}}}

24: \newcommand{\ru}{\rule[-0.4mm]{.1mm}{3mm}}

25: \newcommand{\nni}{\ru\hspace{-3.5pt}}

26: \newcommand{\sni}{\ru\hspace{-1pt}}

27: \newcommand{\pre}{\hspace{0.28em}}

28: \newcommand{\post}{\hspace{0.1em}}

29: \newcommand{\NIm}{\pre\nni\sim}

30: \newcommand{\NI}{\mbox{$\: \nni\sim$}}

31: \newcommand{\notNIm}{\pre\nni\not\sim}

32: \newcommand{\notNI}{\mbox{ $\nni\not\sim$ }}

33: \newcommand{\NIVm}{\pre\nni\sim_V}

34: \newcommand{\NIV}{\mbox{ $\nni\sim_V$ }}

35: \newcommand{\notNIVm}{\pre\sni{\not\sim}_V\post}

36: \newcommand{\notNIV}{\mbox{ $\sni{\not\sim}_V$ }}

37: \newcommand{\NIWm}{\pre\nni\sim_W}

38: \newcommand{\NIW}{\mbox{ $\nni\sim_W$ }}

39: \newcommand{\NIWp}{\mbox{ $\nni\sim_{W'}$ }}

40: \newcommand{\notNIWm}{\pre\sni{\not\sim}_W\post}

41: \newcommand{\notNIW}{\mbox{ $\sni{\not\sim}_W$ }}

42: \newcommand{\eem}{\hspace{0.8mm}\rule[-1mm]{.1mm}{4mm}\hspace{-4pt}}

43: \newcommand{\EM}{\eem\equiv}

44: \newcommand{\notEM}{\eem\not\equiv}

45: \newcommand{\R}{\cal R}

46: \newcommand{\notR}{\not {\hspace{-1.5mm}{\cal R}}}

47: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

48: \newcommand{\bK}{{\bf K}}

49: \newcommand{\bKp}{${\bf K}^p$}

50: \newcommand{\oK}{$\overline {\bf K}$}

51: \newcommand{\bM}{{\bf M}}

52: \newcommand{\bP}{{\bf P}}

53: \newcommand{\ga}{\mbox{$\alpha$}}

54: \newcommand{\gb}{\mbox{$\beta$}}

55: \newcommand{\gc}{\mbox{$\gamma$}}

56: \newcommand{\gd}{\mbox{$\delta$}}

57: \newcommand{\gep}{\mbox{$\varepsilon$}}

58: \newcommand{\gf}{\mbox{$\zeta$}}

59: \newcommand{\cA}{\mbox{${\cal A}$}}

60: \newcommand{\cB}{\mbox{${\cal B}$}}

61: \newcommand{\cC}{\mbox{${\cal C}$}}

62: \newcommand{\cE}{\mbox{${\cal E}$}}

63: \newcommand{\cF}{\mbox{${\cal F}$}}

64: \newcommand{\cK}{\mbox{${\cal K}$}}

65: \newcommand{\cL}{\mbox{${\cal L}$}}

66: \newcommand{\cM}{\mbox{${\cal M}$}}

67: \newcommand{\cR}{\mbox{${\cal R}$}}

68: \newcommand{\cS}{\mbox{${\cal S}$}}

69: \newcommand{\cT}{\mbox{${\cal T}$}}

70: \newcommand{\cU}{\mbox{${\cal U}$}}

71: \newcommand{\ab}{\mbox{\ga \NI \gb}}

72: \newcommand{\cd}{\mbox{\gc \NI \gd}}

73: \newcommand{\ef}{\mbox{\gep \NI \gf}}

74: \newcommand{\xe}{\mbox{$\xi$ \NI $\eta$}}  

75: \newcommand{\pht}{\mbox{$\varphi$ \NI $\theta$}}

76: \newcommand{\rt}{\mbox{$\rho$ \NI $\tau$}}

77: \newcommand{\Cn}{\mbox{${\cal C}n$}}

78: \newcommand{\CF}{\mbox{${\cal C}_{\cal F}$}}

79: \newcommand{\CG}{\mbox{${\cal C}_{\sim}$}}

80: \newcommand{\CW}{\mbox{${\cal C}_{W}$}}

81: \newcommand{\Pf}{\mbox{${\cal P}_{f}$}}

82: \newcommand{\leC}{\mbox{${\preceq_{\cC}}$}}

83: \newcommand{\leF}{\mbox{${\preceq_{\cF}}$}}

84: \newcommand{\lC}{\mbox{${\prec_{\cC}}$}}

85: \newcommand{\notlC}{\mbox{$\not \! \! \lC$}}

86: \newcommand{\lF}{\mbox{${\prec_{\cF}}$}}

87: \newcommand{\Z}{\mbox{$Z_{\cF}$}}

88: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

89: \newcommand{\ra}{\rightarrow}

90: \newcommand{\Ra}{\Rightarrow}

91: \newcommand{\eqdef}{\stackrel{\rm def}{=}}

92: \newcommand{\absv}[1]{\mid #1 \mid}

93: \newcommand{\vstar}{\mbox{$V\sstar_{\infty}$}}

94: \newcommand{\sumstar}{\mbox{$\sum\sstar$}}

95: \newcommand{\tilh}{\mbox{$\tilde{h}$}}

96: \newcommand{\tilep}{\mbox{$\tilde{\varepsilon}$}}

97: \newcommand{\tilf}{\mbox{$\tilde{f}$}}

98: \newcommand{\gahat}{\mbox{$\hat{\ga}$}}

99: \newcommand{\gafalse}{\mbox{$\ga\NI{\bf false}$}}

100: \newcommand{\sstar}{^{*}}

101: \newcommand{\calR}{\mbox{${\cal R}\sstar$}}

102: \newcommand{\subseteqf}{\mbox{$\subseteq_{f}$}}

103: 

104: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

105: \title{Another perspective on Default Reasoning

106: \thanks{

107: This work was 

108: partially supported 

109: by the Jean and Helene Alfassa fund for 

110: research in Artificial Intelligence}

111: }

112: \author{Daniel Lehmann \\ Institute of Computer Science, \\

113: Hebrew University, Jerusalem 91904 (Israel)

114: }

115: \date{}

116: 

117: \maketitle

118: \begin{abstract}

119: The lexicographic closure of any given finite set $D$ of normal defaults

120: is defined. A conditional assertion \mbox{$a \NIm b$} is in this lexicographic

121: closure if, given the defaults $D$ and the fact $a$, one would conclude

122: $b$. The lexicographic closure is essentially a rational extension of

123: $D$, and of its rational closure, defined in a previous paper.

124: It provides a logic of normal defaults that is different from the one

125: proposed by R. Reiter and that is rich enough not to require the consideration

126: of non-normal defaults.

127: A large number of examples are provided to show that the lexicographic

128: closure corresponds to the basic intuitions behind Reiter's logic of 

129: defaults.

130: \end{abstract}

131: 

132: \section{Plan of this paper}

133: \label{sec:plan}

134: Section~\ref{sec:intro} is a general introduction, describing the goal

135: of this paper, in relation with Reiter's Default Logic and the program

136: proposed in~\cite{LMAI:92} by Lehmann and Magidor.

137: Section~\ref{sec:nonmonreas} first discusses at length some general

138: principles of the logic of defaults, with many examples, and, then, 

139: puts this paper in perspective relatively to previous work.

140: Section~\ref{sec:what} sets the stage for this paper by describing

141: the intuitive meaning of default information and the formal representation

142: used in this paper for defaults. It singles out two different

143: possible interpretations for defaults: a {\em prototypical} and

144: a {\em presumptive} reading.

145: Section~\ref{sec:defvsmat} briefly discusses the relation between

146: defaults and material implications.

147: This paper proposes a meaning to any set $D$ of defaults.

148: This meaning is presented in a complex construction, that is

149: described in full only in Section~\ref{sec:full}.

150: The different aspects of this construction are presented

151: separately at first.

152: In Section~\ref{sec:single}, the meaning of a set consisting of a single

153: default will be studied. Reiter's proposal does not enable the use

154: of a default the antecedent of which is not known to hold.

155: The new perspective of this paper allows much more sophisticated ways of using 

156: default information.

157: In particular the default $(a : b)$ may be used to conclude that,

158: if $b$ is known to be false, then $a$ should be presumed to be false too.

159: Section~\ref{sec:seminormal} is a short digression on non-normal defaults.

160: It is shown that such defaults can never be understood if one 

161: requires that the closure of a set of defaults be rational.

162: Section~\ref{sec:Dsys} studies interacting normal defaults that

163: have the same rank (or strength). We propose that, in the case of 

164: contradictory defaults of the same rank, we try to satisfy as many

165: as possible. This proposal is in disagreement with D.~Poole's~\cite{Poole:88},

166: but in agreement with the Maximal Entropy approach of~\cite{GMP:90}.

167: It is shown that this idea guarantees rationality.

168: In Section~\ref{sec:full} a formal description of our complete 

169: proposal is given.

170: First, a model-theoretic construction is presented:

171: given a finite set $D$ of normal defaults, a modular model is defined and

172: the lexicographic closure of $D$ is the

173: rational consequence relation defined by this model.

174: Then, an equivalent characterization in terms of maxiconsistent sets is given.

175: Section~\ref{sec:exc} presents examples and the description of the answer 

176: provided by our proposal.

177: One of those shows how and why this proposal disagrees with 

178: the Maximal Entropy approach.

179: Section~\ref{sec:properties} is a concluding discussion.

180: 

181: \section{Introduction}

182: \label{sec:intro}

183: In~\cite{Reiter:80}, R. Reiter proposed a formal framework for 

184: Default Reasoning. Its focal point is the definition of an extension.

185: In~\cite{ReiterCri:83}, R. Reiter and G. Criscuolo found that, 

186: in this framework, one must consider non-normal defaults.

187: Non-normal defaults have, since then, been taken as the basic 

188: piece of {\em default} information by the logic programming community.

189: An alternative point of view is propounded here.

190: An answer is provided to the following question: 

191: given a set $D$ of normal defaults, what are the normal defaults that should

192: be considered as following from $D$, or entailed by $D$?

193: This answer provides a Logic of Defaults that does not suffer from the

194: problems of multiple extensions or the inabilty of Reiter's system to

195: cope satisfactorily with disjunctive information. There is no need to consider

196: non-normal defaults. 

197: In~\cite{LMAI:92}, M. Magidor and the author proposed, as their first

198: thesis [Thesis~1.1 there],

199: that the set of defaults entailed by any set $D$ be rational. 

200: This requirement is met.

201: The rational closure of a set $D$, defined there,

202: is not the set looked for, since it does not provide for inheritance

203: of generic properties to exceptional subclasses. 

204: In their second thesis [Thesis~5.25 in Section~5.9], they

205: proposed to look for some uniform way of constructing a rational superset

206: of the rational closure of a knowledge base.

207: The answer provided here, the lexicographic closure, is almost such a set,

208: and a simple variation meets the condition in full.

209: Independently, in~\cite{BCDLP:lexi}, Benferhat\&al. proposed a similar 

210: lexicographic construction based on an unspecified ordering of single 

211: defaults. When one applies their construction to the ordering on single 

212: defaults defined in~\cite{LMAI:92}, one obtains the lexicographic closure

213: presented in this paper.

214: Its computational complexity has been studied in~\cite{CL:comp} 

215: and~\cite{L:priv}: it is in

216: \mbox{$\Delta^{p}_{2}$} and is NP-hard and co-NP-hard.

217: The lexicographic closure is a syntactic construction in the sense 

218: of~\cite{Nebel:syntax}, i.e., it is sensitive to the presentation of the 

219: default information.

220: 

221: \section{Nonmonotonic Reasoning}

222: \label{sec:nonmonreas}

223: %Discuss Default Reasoning, normal and semi-normal defaults.

224: %Compare defaults statements and typicality statements.

225: %Explain the plan of the paper.

226: \subsection{The Rational Enterprise}

227: \label{subsec:rat}

228: We shall briefly summarize~\cite{KLMAI:89} and~\cite{LMAI:92}

229: and set up the stage.

230: This section was prepared in collaboration with David Makinson.

231: Some nonmonotonic inference relations are better behaved than others.

232: In particular, there are some simple closure conditions that appear

233: highly desirable: {\em reflexivity, left-logical-equivalence, 

234: right-weakening, and, or,} and {\em cautious monotonicity}.

235: The family of relations that satisfy those properties is closed under

236: intersection. Therefore, given a set \bK\ of ordered pairs 

237: \mbox{$(a , b)$} of formulas (which we shall write \mbox{$a \NIm b$} 

238: to remind us

239: that they are meant to be elements of an inference relation),

240: there is a natural and convincing way of defining a distinguished superset

241: of \bK\ that satisfies those conditions: simply put \bKp, 

242: called the preferential closure of \bK, to be the intersection of all 

243: supersets of \bK\ that satisfy the six conditions above.

244: 

245: However, there are other desirable ``closure'' (in a broad sense) properties

246: that are much more difficult to deal with.

247: {\em Rational monotonicity} defines a family of relations that is not closed

248: under intersection.

249: Other desirable conditions appear to be incapable of a purely formal

250: expression, but may be conveyed intuitively and are illustrable

251: by examples. Because of their informal nature, their identification is

252: not cut and dried, but four seem to be of particular interest:

253: \begin{enumerate}

254: \item \label{typ} the presumption of typicality,

255: \item \label{ind} the presumption of independence, 

256: \item \label{prio} priority to typicality, and

257: \item \label{spec} respect for specificity.

258: \end{enumerate}

259: There may be other desirable properties.

260: 

261: (\ref{typ}) The {\em presumption of typicality} begins where rational

262: monotonicity leaves off. Suppose \mbox{$ p \NIm x \in \bK$}.

263: By rational monotonicity, the {\em closure} of \bK, $\bK^{+}$, will contain

264: either \mbox{$p \wedge q \NIm x$} or \mbox{$p \NIm \neg q$}.

265: But which? No guideline is given. The presumption of typicality

266: (it may as well be called ``a weak presumption of monotonicity'')

267: tells us that, in the absence of a convincing reason to accept

268: the latter, we should prefer the former. 

269: \begin{example}

270: {\rm 

271: if \bK\ has \mbox{$p \NIm x$} as its sole element,

272: there is no apparent reason why the relation  \mbox{$\bK^{+} \supseteq \bK$}

273: that we regard as ``generated'' by \bK\ should contain \mbox{$p \NIm \neg q$}.

274: Hence, it should contain \mbox{$p \wedge q \NIm x$}.

275: {\em Note}: In this and all examples, 

276: $p , q , r \ldots x , y , z$ are understood to be 

277: {\em distinct} atomic formulas, i.e., propositional variables.

278: }

279: \end{example}

280: 

281: (\ref{ind}) The {\em presumption of independence} is a

282: sharpening of the presumption of typicality, and thus a stronger presumption

283: of monotonicity. For,

284: even if typicality is lost with respect to one consequent, we may still

285: presume typicality with respect to another, ``unless there is reason

286: to the contrary''.

287: \begin{example}

288: \label{indep}

289: {\rm

290: Suppose \mbox{$\bK = \{p \NIm x , p \NIm \neg q\}$}.

291: Presumption of typicality cannot be used to support \mbox{$p \wedge q \NIm x$},

292: since $\bK^{+}$ is known to contain \mbox{$p \NIm \neg q$}.

293: Presumption of independence tells us we should expect $x$ to be independent

294: of $q$, and therefore unaffected by the truth of $q$.

295: Therefore, it tells us, we should accept \mbox{$p \wedge q \NIm x$}.

296: }

297: \end{example}

298: \begin{example}

299: \label{indep2}

300: {\rm

301: Suppose \mbox{$\bK = \{p \NIm x , p \wedge q \NIm \neg x, p \NIm y\}$}.

302: Notice that \mbox{$p \NIm \neg q$} is in \bKp, the preferential closure

303: of \bK, and, therefore, the presumption of typicality cannot convince us

304: to accept \mbox{$p \wedge q \NIm y$}.

305: But, we should presume that $x$ is independent from $y$, as there is no reason

306: to think otherwise, and put \mbox{$p \wedge q \NIm y$} in the desired

307: consequence relation \mbox{$\bK^{+} \supseteq \bK$}.

308: }

309: \end{example}

310: 

311: {\em Remark:} The two conditions above may be interpreted 

312: as related to and strengthening the condition of rational monotonicity.

313: The difference between rational monotonicity, on one hand, and the

314: presumptions of typicality and independence is subtle, 

315: and may be easily overlooked. 

316: Rational monotonicity is a constraint on the product 

317: \mbox{$\bK^{+} \supseteq \bK$}, whereas presumptions of typicality and

318: independence are

319: best understood as rough and partial guides to the construction of a

320: desirable $\bK^{+}$.

321: 

322: (\ref{prio}) {\em Priority to typicality} tells us that, 

323: in a situation of clash

324: between two inferences, one of them based on the presumption of typicality,

325: the other one based on the presumption of independence, then 

326: we should prefer the former. Two examples are provided now.

327: \begin{example}

328: \label{strict}

329: {\rm

330: Suppose \mbox{$\bK = \{p \NIm x , p \wedge q \NIm \neg x\}$}.

331: The presumption of typicality offers \mbox{$p \wedge q \wedge r \NIm \neg x$},

332: since there is no compelling reason to accept \mbox{$p \wedge q \NIm \neg r$}.

333: The presumption of independence offers both 

334: \mbox{$p \wedge q \wedge r \NIm \neg x$} and

335: \mbox{$p \wedge q \wedge r \NIm x$}.

336: It clearly would not be justified to draw both conclusions.

337: Priority to typicality, tells us to prefer the former.

338: }

339: \end{example}

340: \begin{example}

341: \label{defeasible}

342: {\rm

343: Suppose \mbox{$\bK = \{p \NIm x , {\bf true} \NIm q , q \NIm \neg x\}$}.

344: The presumption of independence, acting on the last assertion of \bK, 

345: offers \mbox{$q \wedge p \NIm \neg x$}.

346: This is in conflict with

347: \mbox{$p \wedge q \NIm x$} that is offerred by presumption of typicality, from

348: the first assertion.

349: Priority to typicality says we should prefer the latter conclusion.

350: }

351: \end{example}

352: 

353: (\ref{spec}) {\em Respect for specificity} tells us that, in case of clash

354: between two presumptions, one of them based on an assertion with a more

355: specific antecedent than the other,

356: we should prefer the conclusion based on the more specific antecedent.

357: This principle is generally accepted and has been discussed in the literature.

358: It is somewhat difficult to formalize: what does ``based on'' mean?

359: It is closely related to the priority to typicality principle described

360: just above, but the exact relationship between those two principles

361: still needs clarification.

362: In examples~\ref{strict} and~\ref{defeasible}, the priority given

363: to typicality achieves precisely the respect for specificity we are looking 

364: for.

365: In example~\ref{strict}, we prefer to use \mbox{$p \wedge q \NIm \neg x$}

366: to \mbox{$p \NIm x$} also because $p \wedge q$ is strictly more specific

367: than $p$, i.e., \mbox{$p \wedge q \models p$}.

368: In example~\ref{defeasible}, we prefer to use \mbox{$p \NIm x$}

369: to \mbox{$q \NIm \neg x$} also because $p$ is defeasibly more specific

370: than $q$, since, from \mbox{${\bf true} \NIm q$}, we shall conclude

371: \mbox{$p \NIm q$} by presumption of typicality, or, preferably, presumption

372: of independence. Another, more technical, reason to view $p$ as more specific 

373: than $q$ is that the rank (the definition found in~\cite{LMAI:92} is explained

374: at the end of~\ref{subsec:introfull}) of $p$ is strictly greater than the 

375: rank of $q$.

376: 

377: Of course, along with the above principles, one should also not forget

378: {\em avoidance of junk}: the desired \mbox{$\bK^{+} \supseteq \bK$}

379: should avoid gratuitous additions (otherwise, e.g. the total relation

380: would do). In other words, $\bK^{+}$ should be, in some sense,

381: ``least'' among the supersets of \bK\ satisfying the desired conditions.

382: ``Least'' should certainly imply minimal in the set-theoretic sense,

383: i.e., no strict subset is acceptable,

384: but cannot mean ``included in any acceptable superset'', since our family

385: is not closed under intersection.

386: 

387: In~\cite{LMAI:92}, a construction is given, that, given any (finite) set \bK\

388: of pairs \mbox{$a \NIm b$} provides a rational extension \oK\ such that

389: \mbox{$\bK \subseteq \bK^{p} \subseteq \overline {\bK} = 

390: {\overline \bK}^{p} = \overline{\overline \bK}$}

391: that behaves well so far as the presumption of typicality and the respect

392: for specificity are concerned. However it does not pay much heed to the

393: presumption of independence. For example it does not legitimize the

394: conclusion \mbox{$p \wedge q \NIm y$} in example~\ref{indep2} above.

395: The purpose of the present paper is to propose a different construction

396: that performs better in this last respect, whilst not losing satisfaction

397: of the other formal and informal properties.

398: \subsection{Related Work}

399: R.~Reiter's~\cite{Reiter:80} was certainly one of the most influential

400: papers in the field of knowledge representation. 

401: It proposed a way of dealing with {\em default} information.

402: In short, it proposed to represent such information as normal defaults

403: and to define the meaning of a set of normal defaults as the set

404: of extensions it provides to any set of sentences.

405: In a follow-up paper~\cite{ReiterCri:83}, R.~Reiter and G.~Criscuolo

406: remarked that, in many instances, the simple-minded formalization of

407: situations involving more than one (normal) default was not adequate: 

408: the extension semantics enforced some unexpected and undesirable 

409: consequences. 

410: They proposed to cure this problem by considering an extended

411: class of defaults: semi-normal defaults.

412: 

413: In this paper, a different perspective on default reasoning is proposed.

414: Normal defaults are considered and sets of normal defaults are given

415: a meaning that is different from the one proposed in~\cite{Reiter:80}.

416: With this meaning, the interactions between defaults are as expected

417: and the consideration of non-normal defaults is superflous.

418: This perspective is in line with the first thesis of~\cite{LMAI:92},

419: that requires a set of defaults to define a rational consequence relation.

420: It is also almost in line with the second thesis of the same paper, 

421: that requires

422: a set of defaults to define a consequence relation that extends the rational

423: closure of the set of defaults, and a straightforward variation will

424: be shown to extend rational closure.

425: This goal of implementing Reiter's program, but with different

426: techniques, is similar to David Poole's~\cite{Poole:88}.

427: The present paper also shares some technical insights with Poole's. 

428: It may be considered also as a close relative of the maximal entropy

429: approach of~\cite{GMP:90,GMP:93}, but the semantics proposed here is different

430: from the one obtained from maximal entropy considerations.

431: This paper is a descendant of~\cite{LehTech:92}.

432: The main ideas of the lexicographic construction proposed in this paper

433: have been, independently, proposed in~\cite{BCDLP:lexi}. There, the 

434: initial ordering of single defaults was left for the user to chose.

435: A specific ordering of single defaults is used here.

436: 

437: \section{What is default information?}

438: \label{sec:what}

439: Default information is information about the way things usually are.

440: The paradigmatical example of such information,

441: that has been used by most researchers in the field, is 

442: {\em birds fly}. 

443: Syntactically, a default is a pair of propositions that will be written as

444: \mbox{$(a:b)$}, where $a$ and $b$ are formulas (of a propositional calculus

445: for this paper). Remember that only normal defaults are considered,

446: so that \mbox{$(a:b)$} is our notation for Reiter's

447: \mbox{${a : b} \over {b}$}.

448: The default \mbox{$({\bf true}:b)$} will written as \mbox{$( : b)$}.

449: Given a set $D$ of defaults representing some background information about 

450: the way things typically behave and a formula $a$ representing our

451: knowledge of the situation at hand, we shall ask what formulas should

452: be accepted as presumably true.

453: The meaning of a set of defaults $D$ will therefore be understood as

454: the set of pairs (conditional assertions) 

455: \mbox{$ c \NIm d $} it {\em entails},

456: i.e., for which $d$ should

457: be presumed to be true if $c$ is our knowledge about the specific situation,

458: i.e. represents the conjunction of the facts we know to be true.

459: It is probably reasonable to expect that the conditional assertion

460: \mbox{$ c \NIm d $} be entailed by a set $D$ containing the default

461: \mbox{$( c : d )$}, but this will be discussed in the sequel.

462: Notice, that we may, as well, consider that a set of normal defaults

463: entails a set a defaults, confusing ``snake'' (\NI) and colon (:).

464: 

465: The meaning of defaults is a delicate affair and it will be discussed

466: in depth now.

467: In~\cite{ReiterCri:83}, a {\em prototypical} reading is proposed: 

468: {\em birds fly}

469: being understood as {\em typical birds fly}. But, there is another possible

470: reading: {\em birds are presumed to fly unless there is evidence to the 

471: contrary}. This second reading will be called the {\em presumptive} reading.

472: The conclusions of this paper may be summarized in three sentences.

473: The two readings above are {\em almost} equivalent when isolated defaults are 

474: concerned,

475: they are {\em not} when sets of defaults are concerned. 

476: The rational closure construction of~\cite{LMAI:92} is the correct

477: formalization of the prototypical reading. The presumptive 

478: reading is the one

479: intended by Default Logic and its formalization is the topic of this paper.

480: The distinction between the two readings will be explained with an example.

481: This example is formally equivalent to the {\em Swedes} example described 

482: informally in~\cite[page 4]{LMAI:92}.

483: In this example, as in all other examples of this paper, the formulas

484: appearing in the defaults will be represented by meaningless letters

485: and not, as customary in the field, by meaningful sentences.

486: The remark that logic, the study of deductive processes, may

487: be concerned only with the form of the propositions, and not

488: with their meaning, dates back to Aristotle.

489: The use of semantically loaded formulas and the import of the reader's

490: knowledge of the world may only hamper the study of the formal 

491: properties of nonmonotonic deduction (that should perhaps be called induction).

492: When a given example is formally isomorphic to some well known folklore

493: example (or at least to some possible formalization of it), it will be 

494: pointed out.

495: \begin{example}[Swedes]

496: \label{ex:swedes}

497: {\rm Let $p$ and $q$ be different propositional variables.

498: Let $D$ be the set of two defaults: \mbox{$\{( :p) , ( :q) \}$}.

499: Accepting $D$ means that we believe, by default, that $p$ is true, and also

500: that $q$ is true.

501: Following the prototypical reading, then, 

502: {\em typically $p$ is true} and 

503: {\em typically $q$ is true}.

504: Following the presumptive reading {\em $p$ is presumed to be true unless 

505: there is evidence to the contrary} and {\em $q$ is presumed to be true unless 

506: there is evidence to the contrary}.

507: Suppose now that we have the information that \mbox{$\neg p \vee \neg q$} 

508: is true,

509: i.e., at least one of $p$ or $q$ is false.

510: 

511: Using the prototypical reading, we shall conclude that the situation at hand

512: is {\em not} typical. In such a case none of our two defaults is applicable:

513: {\em typically $p$ is true, but this is not a typical situation}, and therefore

514: we shall not conclude, even by default (i.e. defeasibly) that 

515: \mbox{$p \vee q$} holds true.

516: 

517: Using the presumptive reading, on the contrary, we shall conclude that

518: \mbox{$p \vee q$} should be presumed to be true unless there is evidence to 

519: the contrary, and since there

520: is no evidence of this sort, it should be presumed to be true.

521: We should therefore presume that exactly one of $p$ and $q$ holds.

522: }

523: \end{example}

524: 

525: \section{Default vs. Material Implication}

526: \label{sec:defvsmat}

527: A very natural feeling is that the meaning of any single default

528: \mbox{$( a : b )$} should be closely related to the meaning of the material

529: implication \mbox{$a \ra b$}. 

530: This last formula will be called the {\em material

531: counterpart} of the default \mbox{$( a : b )$}.

532: Similarly the meaning of a set of defaults $D$

533: should be related to the meaning of the set of its material counterparts.

534: 

535: It turns out that, both in the rational closure construction of~\cite{LMAI:92}

536: and in the construction proposed in this paper, the meaning of a default

537: \mbox{$( a : b )$} (that is an element of the set $D$ of defaults accepted

538: by a reasoner)

539: in the presence of knowledge $c$, either its material

540: counterpart \mbox{$a \ra b$} or void 

541: (i.e. equivalent to a tautology: {\bf true}).

542: Both constructions may therefore be described by pinpointing, given specific

543: information $c$, which of the defaults of $D$ are meaningful.

544: If $D_{c}$ is this set and $M_{c}$ the set of material counterparts

545: of $D_{c}$, then $d$ should be presumed true iff

546: $d$ is a logical consequence of \mbox{$c \cup M_{c}$}, i.e.,

547: \mbox{$M_{c} , c \models d$}.

548: This semantics fits well into the {\em implicit content} framework proposed

549: in~\cite{Stal:92}.

550: 

551: \section{Single Defaults}

552: \label{sec:single}

553: %Default statement =Typicality judgement

554: %Almost Rational Closure is our answer.

555: %Full description of the rational closure of (a:b).

556: Let $D$ be the singleton set \mbox{$\{ ( a : b ) \}$},

557: where $a$ and $b$ are arbitrary formulas.

558: We propose the following meaning to $D$:

559: \begin{itemize}

560: \item if the information at hand $c$ is consistent with \mbox{$a \ra b$},

561: i.e., \mbox{$c \not \models a \wedge \neg b$}

562: then the default is meaningful and $d$ is presumed iff 

563: \mbox{$ c , a \ra b \models d$}

564: \item otherwise, the default is meaningless and $d$ is presumed iff

565: \mbox{$c \models d$}.

566: \end{itemize}

567: 

568: An equivalent, more model-theoretic description, of the consequence

569: relation determined by $D$ is the following:

570: the rational consequence relation that is defined by the

571: modular model in which the propositional models are ranked in two levels:

572: on the bottom level (the more normal one) all models that satisfy

573: the material implication $a \ra b$, on the top level all other models.

574: 

575: This is the most natural understanding of the default information

576: {\em if $a$ is true then $b$ is presumably true}, 

577: and completely in line with D.~Poole's~\cite{Poole:88} treatment of

578: defaults.

579: Notice, though, that it does not always agree with Reiter's treatment and

580: only almost agrees with rational closure.

581: If the information at hand $c$ logically implies $a$, then the

582: perspective proposed here agrees with Reiter's: $d$ is presumed to be true iff

583: $d$ is an element of the unique extension of \mbox{$(D , \{c\})$}.

584: 

585: To see the difference with Reiter's treatment, suppose $a$ and $b$ are

586: different propositional variables and consider $c$ to be $\neg b$.

587: The perspective defended here will support the claim that 

588: \mbox{$\neg a$} should be presumed to be true, i.e., $a$ should be

589: presumed to be false.

590: For Reiter, on the contrary, there is a unique extension: \mbox{$\Cn(\neg b)$}

591: ($\Cn$ is the logical consequence operator) 

592: and therefore we should not presume 

593: anything about $a$.

594: Similarly if $c$ is {\bf true}, the present perspective will support

595: \mbox{$a \ra b$}, whereas Reiter will not.

596: 

597: 

598: The comparison with rational closure is more subtle.

599: Our perspective agrees with rational closure except when

600: \mbox{$a \models \neg b$}.

601: This is quite an out of the ordinary situation: $a$ is logically equivalent

602: to something of the form

603: \mbox{$\neg b \wedge e$}, and the default is of the form

604: {\em if $b$ is false and $e$ is true, then assume $b$ is true}.

605: Such a default will probably never be used in practice,

606: but its consideration is nevertheless enlightening.

607: In such a situation, the present perspective claims that the meaning of the

608: default \mbox{$(\neg b \wedge e : b)$} is that all models that

609: satisfy \mbox{$\neg b \wedge e$} are on the top level.

610: In other terms, if 

611: \mbox{$c \not \models \neg b \wedge e$}, the default is meaningful

612: and means \mbox{$\neg b \wedge e \ra b$}, which is logically 

613: equivalent to \mbox{$e \ra b$}, but if \mbox{$c \models \neg b \wedge e$},

614: then the default is meaningless.

615: The treatment of this last case is well in line with the presumptive

616: reading: if $b$ is known to be false, then $b$ should not be assumed

617: to be true.

618: If we look at the way rational closure deals with this case, we see

619: that it agrees with the present perspective in the first case (i.e.,

620: if \mbox{$c \not \models \neg b \wedge e$}) but disagrees with it

621: in the second case. Rational closure accepts any conclusion from the

622: information that $e$ is true and $b$ is false.

623: This is in line with the prototypical reading of the default:

624: {\em if $e$ is true and $b$ is false, then typically $b$ is true}

625: may only mean that it is inconsistent for \mbox{$e \wedge \neg b$} to be true 

626: and therefore one should conclude anything when this happens.

627: 

628: The new perspective does not always support each member of the rational 

629: closure,

630: but the reader may check that the solution it supports is always rational

631: (in the technical sense of~\cite{LMAI:92}). A proof of this, in a more general

632: setting, will be given in Section~\ref{sec:full}.

633: How come our proposal is different from rational closure, that seemed

634: to be the only reasonable one?

635: Let \bK\ be the conditional 

636: knowledge base containing the single assertion 

637: \mbox{$e \wedge \neg b$ \NI $b$}.

638: The rational relation proposed here in place of the rational closure

639: does not contain \mbox{$e \wedge \neg b$ \NI $b$}.

640: It is not an extension of \bK\ and therefore does not satisfy Thesis~5.25:

641: ``The set of assertions entailed by any set of assertions \bK\ is a

642: rational superset of the rational closure of \bK'' 

643: of~\cite{LMAI:92}.

644: This departure from Thesis~5.25 is not central to our proposal and a slight

645: variant of it would satisfy Thesis~5.25 by treating differently only those 

646: useless defaults discussed just above. This variant does not seem to fully fit

647: the presumptive reading of defaults, though.

648: If we denote by $\bK^{l}$ the (lexicographic) construction proposed 

649: in this paper, the variant we have in mind may be defined in the following way:

650: accept \mbox{$a \NIm b$} iff either $a$ has a rank for \bK\ and 

651: \mbox{$a \NIm b \in \bK^{l}$}, or $a$ has no rank.

652: This variant gives a superset (sometimes strict) of $\bK^{l}$, 

653: that is

654: also a superset (sometimes strict) of the rational closure \oK.

655: \section{Seminormal defaults}

656: \label{sec:seminormal}

657: This paper will show that, if one accepts a semantics that

658: is different from Reiter's, the reasons that compelled him to introduce

659: non-normal defaults disappear, and one may restrict oneself to normal defaults.

660: The reader may well ask whether one would not like to consider, anyway,

661: a more general form of defaults: the semi-normal defaults.

662: A semi-normal default \mbox{${a : e \wedge b}\over{b}$} means that 

663: {\em if $a$ is known to be true and there is no evidence that 

664: \mbox{$e \wedge b$} is false,

665: then $b$ should be presumed to be true}.

666: Let $a$ be a tautology, i.e., {\bf true} and $e$ and $b$ be different

667: propositional variables ($q$ and $p$ respectively).

668: Suppose we accept

669: the semi-normal default \mbox{${ : q \wedge p}\over{p}$}.

670: There is general agreement about the following points:

671: \begin{itemize}

672: \item if the information at hand $c$ is a tautology, 

673: i.e., we have no specific information, we should presume that $p$ is true,

674: since there is no evidence that \mbox{$q \wedge p$} does not hold,

675: \item if $c$ is $\neg q$, i.e., we know for sure that $q$ does not hold,

676: we should not use the default information and therefore we should {\em not}

677: presume $p$,

678: \item but, if we have no specific information, we should {\em not}

679: presume that $q$ holds (why should we?).

680: \end{itemize}

681: The three points above provide a counter-example to the rule of

682: Rational Monotonicity of~\cite{LMAI:92}:

683: we accept \mbox{${\bf true} \NIm p$}, but neither 

684: \mbox{${\bf true} \NIm \neg \neg q$}, nor \mbox{$\neg q \NIm p$}.

685: Even the simplest isolated non-normal default cannot be given 

686: a rational interpretation.

687: This remark is very important in view of the fact that the efforts

688: to harness Logic Programming to Nonmonotonic Reasoning take as their

689: basic component rules of the form

690: \[

691: a \leftarrow b \: , \neg c

692: \]

693: meaning {\em conclude $a$ if

694: $b$ has been concluded and $c$ cannot be concluded}.

695: This is essentially equivalent to considering the semi-normal default

696: \[ {b : a \wedge \neg c}\over{a},\] 

697: or to considering the not even semi-normal default

698: \[ {b : \neg c} \over {a} \]

699: and will lead to a consequence relation that is not rational.

700: All we have shown here is that non-normal defaults or the logic programming

701: approach to nonmonotonic reasoning are incompatible with the property of

702: rational monotonicity, which is central to this and previous papers.

703: \section{Competing but equal defaults}

704: \label{sec:Dsys}

705: %D-systems. Rationality. Musicians example.

706: After dealing, in Section~\ref{sec:single}, with single defaults,

707: we shall treat now the more interesting case of a set of interacting

708: normal defaults.

709: In general, given a set of defaults $D$, this set defines a ranking 

710: of the defaults, as explained in~\cite{LMAI:92}.

711: This ranking will be described in full in Section~\ref{subsec:introfull}.

712: The ranking of a default \mbox{$(a : b)$} relative to $D$ depends only on 

713: its antecedent $a$ and, as we shall see in Section~\ref{sec:exc},

714: defaults of higher ranking (they correspond to exceptions) should be 

715: considered stronger than those of lower ranking.

716: In this Section, we shall deal with the case all defaults have the same

717: rank, i.e., all defaults are equal in strength and none of them

718: correspond to an exception.

719: This happens only when all elements of $D$ have rank zero,

720: as will be clear in Section~\ref{subsec:introfull}.

721: It is clear that, when considering such defaults, we should always

722: assume that as many defaults as possible are satisfied (i.e. not violated).

723: We should therefore always prefer violating a smaller set of defaults

724: to violating a larger one. One may hesitate about the meaning

725: to be given to ``smaller'': set inclusion or smaller size.

726: The main conclusion of our considerations will be that sets of

727: defaults should (for rationality's sake) be compared by their size, 

728: not by set inclusion.

729: 

730: We choose an example isomorphic to the musicians example 

731: of~\cite[Section 4.4]{Gin:86}, but we shall first ask about it

732: questions that are different from those asked traditionally.

733: \begin{example}[Musicians]

734: {\rm 

735: Let $p$, $q$ and $r$ be different propositional variables.

736: Let $D$ contain the three following defaults:

737: \mbox{$\{( : p) , ( : q) , ( : r)\}$}.

738: In other words, $p$, $q$ and $r$ are assumed to hold by default.

739: If we learn that \mbox{$c \eqdef \neg p \wedge \neg q \vee \neg r$} holds,

740: i.e., that either both $p$ and $q$ are false, contrary to expectation,

741: or, also contrary to expectation, $r$ does not hold, what should we

742: assume?

743: Should one of the two possibilities (\mbox{$\neg p \wedge \neg q$}) and

744: \mbox{$\neg r$} be assumed more likely than the other one?

745: 

746: In~\cite{Poole:88}, D.~Poole claims we should not.

747: He claims there are two different maximal subsets of the material counterpart

748: of $D$ consistent with $c$ (two bases for $c$): \mbox{$\{p , q\}$}

749: and \mbox{$\{r\}$} and

750: he proposes that we presume true only those formulas that are both

751: in \mbox{$\Cn(r , c)$} and in \mbox{$\Cn(p , q , c)$}.

752: In particular, we should not presume $p$ to be true.

753: But, we should presume \mbox{$p \leftrightarrow q$} to hold.

754: Also, if we learn that \mbox{$c \wedge \neg p$} holds,

755: we should presume true only those formulas that are both in

756: \mbox{$\Cn(r , c , \neg p) = \Cn(r , c)$} and in 

757: \mbox{$\Cn(q , c , \neg p)$}. In particular, we should not

758: presume the truth of \mbox{$p \leftrightarrow q$}.

759: D.~Poole's proposal, therefore, does not satisfy the Rational

760: Monotonicity principle.

761: 

762: Guided by Thesis~1.1 of~\cite{LMAI:92}, that requires Rational Monotonicity,

763: a slight modification of Poole's ideas will be put forward now.

764: This modification is also supported by the Maximal Entropy approach 

765: of~\cite{GMP:90}.

766: The two bases above should not be considered equivalently plausible.

767: The larger one, which contains two defaults should be considered more 

768: plausible than the one containing only a single default.

769: In other terms, situations that violate two defaults should be considered

770: less plausible than those that violate only one default.

771: Here is a model-theoretic description.

772: We shall consider the (propositional) models of our language, and rank

773: them by the number of defaults of $D$ they violate.

774: A model $m$ violates a default \mbox{$a \NIm b$} iff it does not satisfy

775: the material implication \mbox{$a \ra b$}, i.e., iff 

776: \mbox{$m \models a \wedge \neg b$}.

777: The most normal models are those that violate no default of $D$:

778: they constitute the bottom level (zero) of our modular model.

779: Slightly less normal are those models that violate one single default:

780: they constitute level one of our model.

781: In general, level $i$ is constituted by all models that violate exactly

782: $i$ members of $D$. The nonmonotonic consequence relation defined

783: by this model is the one defined by $D$. It is rational, since the model

784: described is ranked and consequence relations

785: defined by modular models are rational (Lemma 3.9 of~\cite{LMAI:92}).

786: Coming back to the Musicians example: if we learn that 

787: \mbox{$c \eqdef \neg p \wedge \neg q \vee \neg r$} holds, \mbox{$\neg r$}

788: should be presumed true. We should therefore presume $p$ to be true.

789: }

790: \end{example}

791: 

792: We provided, just above, a model-theoretic description of our proposal.

793: An equivalent description in terms of ``bases'', in the spirit 

794: of~\cite{Poole:88}, is provided now. The same bases were also considered 

795: in~\cite{BCDLP:lexi}.

796: Let \mbox{$E = \{e_{i}\}$} be a finite set of formulas.

797: Let $c$ be a formula.

798: \begin{definition}

799: \label{def:maxbase}

800: A subset $F$ of $E$ is said to be a maxbase for $c$ iff $c$ is consistent

801: with $F$ and there is no subset $F^{\prime}$ of $E$,

802: \mbox{$\mid F^{\prime} \mid > \mid F \mid$}

803: that is consistent with $c$.

804: \end{definition}

805: \begin{theorem}

806: \label{the:rat}

807: Let $D$ be a set of defaults such that all elements of $D$ have rank zero

808: (with respect to $D$). Let $E$ be the set of material implications

809: corresponding to the defaults of $D$.

810: The consequence relation defined by the model-theoretic description above

811: is characterized by:

812: \begin{equation}

813: \label{eq1}

814: a \NIm b {\rm \ iff \ for \ every \ maxbase} \ F \ of \ E {\rm \ for} \ a \ \ 

815: F , a \models b.

816: \end{equation}.

817: \end{theorem}

818: Theorem~\ref{the:rat} implies that 

819: the relation defined by Equation~(\ref{eq1}) is rational.

820: \proof

821: First, some remarks.

822: Let $n$ be the size of the set $D$.

823: \begin{enumerate}

824: \item \label{aa}

825: If $a$ is satisfied by some model of level $i$ 

826: (\mbox{$0 \leq i \leq n$}),

827: then, there is a maxbase for $a$, and all maxbases for $a$ are of size

828: larger or equal to $n - i$.

829: \item \label{bb}

830: If $F$ is a maxbase of size $k$ (\mbox{$0 \leq k \leq n$}) for $a$,

831: then, there is a model of level $n - k$ that satisfies $a$, 

832: and no model of lower level satisfies $a$.

833: \item \label{cc}

834: There is no maxbase for $a$ iff $a$ is a logical contradiction.

835: \end{enumerate}

836: Suppose \mbox{$a \NIm b$}.

837: If no model satisfies $a$, $a$ is a logical contradiction and

838: \mbox{$ X , a \models b$} for any $X$.

839: Suppose, then, that $a$ is satisfied by some model of level $i$, 

840: but by no model of lower level. Any model of level $i$ that satisfies

841: $a$, satisfies $b$, by hypothesis. 

842: Let $F$ be a maxbase for $a$.

843: By remark~(\ref{bb}), \mbox{$ n \: - \mid F \mid \: = \: i$}.

844: Any model that satisfies $F$ is obviously of level less or equal to 

845: \mbox{$n \: - \mid F \mid \: = \: i$}.

846: Any model that satisfies $F$ and $a$ is therefore of level $i$

847: and satisfies $b$, by hypothesis. 

848: We conclude that \mbox{$F , a \models b$}.

849: 

850: Suppose, now, that for any maxbase $F$ for $a$ we have

851: \mbox{$F , a \models b$}.

852: If there is no maxbase for $a$, then, by remark~(\ref{cc}), 

853: $a$ is a logical contradiction and \mbox{$a \NIm b$}.

854: Suppose, then, the maxbases for $a$ are of size $k$.

855: There is, by remark~(\ref{bb}),  

856: a model of $a$ of level $n - k$, and there is no model

857: of level less than $n - k$ that satisfies $a$.

858: We must show that any model of $a$ of level $n - k$ satisfies $b$.

859: Let $m$ be such a model. Since $m$ violates $n - k$ defaults, it

860: satisfies a set $M$ of $k$ defaults. But $M$ is consistent with $a$,

861: since $m \models a$. The size of $M$ is the size of the maxbases for

862: $a$, therefore $M$ is a maxbase for $a$ and,

863: since \mbox{$M , a \models b$}, we conclude that \mbox{$m \models b$}.

864: \QED

865: \begin{example}[Musicians, continued]

866: {\rm

867: We shall now describe our solution to the questions traditionally 

868: asked about the musicians' example and generally used to demonstrate

869: that counterfactuals do not satisfy Rational Monotonicity.

870: Suppose our specific information is 

871: \mbox{$c \eqdef p \wedge \neg r \vee \neg p \wedge r$}.

872: There are two maxbases: \mbox{\{$p , q$\}} and \mbox{$\{q , r\}$}.

873: We shall therefore presume that $q$ holds and we shall {\em not} presume that 

874: \mbox{$d \eqdef q \wedge r \vee \neg q \wedge \neg r$} holds.

875: This is the common wisdom and the present

876: perspective subscribes to it.

877: 

878: Suppose now that our specific information is \mbox{$c \wedge \neg d$}, or,

879: equivalently,

880: \mbox{$p \wedge q \wedge \neg r \vee \neg p \wedge \neg q \wedge r$}.

881: The common wisdom, defended in~\cite{Gin:86}, would like to convince us that 

882: we should not presume $q$ to be true.

883: The position defended here, presumes that

884: $q$ is true (and also $p$ and $\neg r$) because this situation violates

885: only one default $( : r )$ whereas the other possible situation:

886: \mbox{$\neg p \wedge \neg q \wedge r$} violates two defaults.

887: }

888: \end{example}

889: 

890: The reader may suspect that our policy gives results that are extremely 

891: sensitive

892: to the way the defaults are presented.

893: Indeed, the way defaults are presented is important,

894: and our perspective on defaults does not enjoy the nice global properties

895: of rational closure described in~\cite[Section 5.5]{LMAI:92} that make

896: it invariant under the addition or deletion of entailed defaults.

897: Two examples of this phenomenon will be described now.

898: The first one shows that the addition to $D$ of a default entailed by $D$ may add new

899: conclusions.

900: The second one shows that the addition to $D$ of a default entailed by $D$ may 

901: force us to withdraw previous conclusions.

902: The examples presented are very simple and natural and should convince the

903: reader that any presumptive reading of defaults leads to a high

904: sensitivity to the presentation of the default information.

905: This sensitivity is, probably, a drawback of the lexicographic closure.

906: The following examples should convince the reader that this problem is 

907: inevitably brought about by a presumptive understanding of defaults.

908: If we had decided to consider multisets of defaults, instead of sets,

909: thus allowing certain (stronger) defaults to appear a number of times in $D$,

910: our construction would had been sensitive to the number of times each default

911: appears in $D$.

912: \begin{example}[Adding entailed defaults may add conclusions]

913: {\rm

914: Let $D$ be the singleton \mbox{$\{( : p \wedge q)\}$}.

915: The default (identifying defaults and conditional assertions)

916: \mbox{$( : p )$} is obviously entailed by $D$.

917: But the default \mbox{$(\neg p \vee \neg q : p)$} is not entailed

918: by $D$, the antecedent being inconsistent with the only default of $D$.

919: Nevertheless \mbox{$(\neg p \vee \neg q : p)$} is entailed by the set

920: \mbox{$\{ ( : p \wedge q ) , ( : p ) \}$}, since its antecedent

921: is consistent with the second default.

922: The behavior of the corresponding Poole system is the same.

923: }

924: \end{example}

925: 

926: \begin{example}[Adding entailed defaults may delete conclusions]

927: {\rm

928: Let $D$ be the set \mbox{$\{ ( : p ) , ( : q ) \}$}.

929: Both defaults \mbox{$( : p \leftrightarrow q )$} and

930: \mbox{$( \neg p : q )$} are entailed by $D$.

931: But \mbox{$( \neg p : q )$} is {\em not} entailed by the set

932: \mbox{$\{ ( : p ) , ( : q ) , ( : p \leftrightarrow q )\}$},

933: since the antecedent is consistent with both the last defaults separately

934: but not together, and there are therefore two maxbases.

935: In this case also, the behavior of the corresponding Poole systems is the same.

936: }

937: \end{example}

938: 

939: \section{Lexicographic closure}

940: \label{sec:full}

941: \subsection{Introduction and definition}

942: \label{subsec:introfull}

943: In the previous Section we discussed the treatment of conflicting

944: defaults that had the same precedence.

945: We shall now treat arbitrary conflicting defaults, and define the construction

946: we propose in full generality.

947: We must take into account the fact that defaults may have

948: different weight, or precedence. Fortunately, the correct definition of the

949: relative precedence of defaults has been obtained in a previous work.

950: Given a finite set of defaults $D$, the precedence of a default is given by 

951: its rank (higher rank means higher precedence), i.e., by the rank of 

952: its antecedent

953: as defined in~\cite[Section 2.6]{LMAI:92}.

954: The definition presented here is equivalent to the original

955: definition, by Corollary 5.22 there.

956: 

957: We shall now remind the reader of this definition.

958: Let $D$ be a finite set of defaults, and $\tilde{D}$ the set of its material

959: counterparts. Let $a$ be a formula.

960: We shall put \mbox{$E_{0} = D$}.

961: If $a$ does not have rank less than $i$, 

962: but is consistent with $\tilde{E_{i}}$,

963: it has rank $i$.

964: The set $E_{i + 1}$ is the subset of $E_{i}$ that contains all defaults 

965: \mbox{$( a : b)$} of $D$ for which $a$ does not have rank less or equal

966: to $i$.

967: We shall put \mbox{$D_{i} = E_{i} - E_{i + 1}$}, and let $D_{\infty}$

968: be the set of all elements of $D$ that have no rank, i.e., have infinite rank.

969: Elements of $D_{\infty}$ have precedence over all other defaults.

970: Notice that, since $D$ is finite, 

971: all $D_{i}$'s are empty after a certain point, except possibly

972: $D_{\infty}$.

973: There is a $k$ such that for any $i$, \mbox{$k \leq i < \infty$}, 

974: \mbox{$D_{i} = \emptyset$}. The smallest such number $k$ will be called the 

975: {\em order} of the set $D$.

976: The set $D$ may be partitioned into 

977: \mbox{$D_{\infty} \oplus D_{k-1} \oplus D_{k-2} \oplus \ldots \oplus D_{0}$}.

978: 

979: We remind the reader that the rational closure of the set $D$,

980: defined in~\cite{Leh:89} and studied in depth in~\cite{LMAI:92} 

981: (see Theorem 5.17 and Lemma 2.24) is the set of

982: defaults $\overline{D}$ that consists of all defaults \mbox{$(a : b )$}

983: such that the rank of $a$ is strictly less than the rank of 

984: \mbox{$a \wedge \neg b$} (equivalently, the rank of \mbox{$a \wedge b$} 

985: is strictly less than the rank of \mbox{$a \wedge \neg b$}), or such that $a$

986: has no rank. We shall now define another closure for $D$, the lexicographic 

987: closure. We define the lexicographic closure by way of a modular model

988: in which every model is ranked by the set of defaults it violates.

989: A similar presentation may be used to define rational closure,

990: it will also be described.

991: 

992: \subsection{The model-theoretic description}

993: As usual, we shall suppose a finite set $D$ of defaults is given.

994: We shall describe the consequence relation defined by $D$,

995: the lexicographic closure of $D$, $D^{l}$ as the consequence relation

996: defined by a certain modular model, $\cM_{D}$.

997: To define this model, we need to order the propositional models

998: by some modular ordering. We shall order the propositional models by ordering

999: the sets of defaults (of $D$) that they violate:

1000: each model $m$ violates a set $D_{m} \subseteq D$ of defaults.

1001: How should we order the subsets of $D$?

1002: Intuitively, we are looking for a ``degree of seriousness''.

1003: We prefer to violate a ``lighter'' set of defaults than a more

1004: serious one, i.e., a propositional model that violates a lighter

1005: set of defaults is more normal than a model that violates a more serious set.

1006: There are two criteria that must be taken into account when

1007: deciding which of two sets is more serious:

1008: \begin{itemize}

1009: \item the size of the set: the smaller the set, the less serious it is.

1010: We have seen in Section~\ref{sec:Dsys}, that ``smaller'' should be taken

1011: here to mean ``of smaller size'', and

1012: \item the seriousness of the elements of the set, it is less serious

1013: to violate a less specific default than a more specific default,

1014: i.e. a default of lower rank than a default of higher rank. 

1015: \end{itemize}

1016: The reader should notice here that our definition is in no way circular.

1017: The lexicographic closure is defined in terms of a specific modular model

1018: that is, in turn, defined in terms of the ranks of the formulas involved.

1019: These ranks have been defined above, by a straightforward inductive definition.

1020: The ranks of the formulas have a close relationship with the ordering

1021: of the modular model that defines the rational closure of $D$, but this is 

1022: a different modular model. In fact, the model we are describing now

1023: is a refinement of the model that defines rational closure

1024: (a level may split into a number of sublevels).

1025: The next question, now, is how should we compose those two criteria?

1026: The principle of rationality will trace the way for us.

1027: We want a modular ordering on the subsets of $D$.

1028: Each one of the criteria above gives a modular ordering.

1029: Is there a general way to combine two modular orderings and obtain

1030: a modular ordering? Yes, a lexicographic (i.e. consider one criterion as

1031: the principal criterion, the other as secondary) composition of modular

1032: orderings is a modular ordering.

1033: Which of the two criteria above should be considered as the major

1034: criterion?

1035: Clearly the second one: specificity. We should prefer violating two

1036: defaults of low specificity to violating one of high specificity.

1037: \begin{example}

1038: {\rm

1039: Let \mbox{$D = \{ (:p) , (:q) , (:x) , (y:\neg x) , (y:r) \}$}.

1040: Suppose our assumptions are 

1041: \mbox{$ y \wedge (\neg p \wedge \neg q \vee \neg r)$}.

1042: You may imagine that $p$, $q$ and $x$ are generic properties (of birds, say),

1043: and that $y$ is a class of birds that are exceptional with respect to $x$.

1044: The property $r$ is a generic property of $y$ birds.

1045: Suppose we have a bird that is part of the $y$ class, and is known to be,

1046: either exceptional with respect to two generic properties of birds,

1047: or exceptional with respect to one generic property of the sub-class $y$.

1048: Presumption of typicality (see Section~\ref{subsec:rat}), from the

1049: last default of $D$ proposes the conclusion $r$ (and therefore

1050: $\neg p$ and $\neg q$).

1051: Presumption of independence proposes the conclusions $p$ and $q$ 

1052: (and therefore $\neg r$).

1053: Priority to typicality convinces us to accept

1054: the former and reject the latter.

1055: }

1056: \end{example}

1057: Therefore, to decide which of two sets of defaults is more serious,

1058: we shall partition those sets into subsets of defaults of equal ranks

1059: and compare (by size) rank by rank, starting with the higher ranks.

1060: As soon (in terms of ranks) as a decision can be made, we decide and stop.

1061: \begin{definition}

1062: \label{def:tnumber}

1063: Let $D$ be a set of defaults and $k$ its order.

1064: To every subset \mbox{$X \subseteq D$} may be associated a $k + 1$-tuple

1065: of natural numbers: \mbox{$\langle n_{0} , \ldots , n_{k} \rangle$},

1066: where \mbox{$n_{0} = \mid D_{\infty} \cap X \mid$},

1067: \mbox{$n_{1} = \mid D_{k - 1} \cap X \mid$},

1068: and in general, for \mbox{$i = 1 , \ldots , k$}, 

1069: \mbox{$n_{i} = \mid D_{k - i} \cap X \mid$}.

1070: In other terms, $n_{0}$ is the number of defaults of $X$ that have no rank

1071: and, for \mbox{$0 < i \leq k$},

1072: $n_{i}$ is the number of defaults of $X$ that have rank $k-i$.

1073: We shall order the subsets of $D$ by the natural lexicographic ordering

1074: on their associated tuples. This is a strict modular partial ordering:

1075: it will be denoted by $\prec$, (the {\em seriousness ordering}).

1076: \end{definition}

1077: The seriousness ordering on sets of defaults is used to order

1078: the propositional models: \mbox{$m \prec m'$} iff \mbox{$V(m) \prec V(m')$},

1079: where $V(m) \subseteq D$ is the set of defaults violated by $m$.

1080: This modular ordering on models defines a modular preferential model,

1081: that, in turn defines a consequence relation, $D^{l}$, 

1082: the lexicographic closure of $D$.

1083: The reader may check that all examples treated in this paper conform to

1084: the definition above.

1085: 

1086: Let us, now, before we give an alternative description of lexicographic 

1087: closure, briefly digress to see that rational closure may be defined

1088: by a specific seriousness ordering, different from the one defined in

1089: Definition~\ref{def:tnumber}.

1090: \begin{definition}

1091: \label{def:tnumber2}

1092: Let $X_{1}$ and $X_{2}$ be subsets of a set $D$ of defaults and $k$ its order.

1093: Let $n^{1}_i$ and $n^{2}_i$, for \mbox{$i = 1 , \ldots , k$},

1094: be the size of the partitions of $X_{1}$ and $X_{2}$ respectively.

1095: Let $m^{j}$ be the smallest $i$ such that \mbox{$n^{j}_{i} \neq 0$},

1096: for \mbox{$j = 1 , 2$}.

1097: We shall write \mbox{$X_{1} \ll X_{2}$} iff \mbox{$m^{1} > m^{2}$}.

1098: \end{definition}

1099: Clearly \mbox{$X_{1} \ll X_{2}$} implies \mbox{$X_{1} \prec X_{2}$},

1100: i.e., $\ll$ is coarser than $\prec$.

1101: \begin{theorem}

1102: \label{the:ll}

1103: Suppose $a$ has a finite rank. 

1104: The conditional assertion \mbox{$a \NIm b$} is a member of

1105: the rational closure of $D$ iff it is satisfied by the modular model in which

1106: each propositional model is ranked by the $\ll$ ordering on the set of

1107: defaults it violates.

1108: \end{theorem}

1109: \proof

1110: Suppose the rank of $a$ is strictly less than that of \mbox{$a \wedge \neg b$},

1111: and that $m$ is a propositional model that satisfies $a$ and is minimal

1112: among those for the $\ll$ ordering. If $a$ has rank $l$, there is a model

1113: that satisfies $a$ and violates no default of $D$ of rank greater or equal to

1114: $l$. We conclude that $m$ violates no such default and therefore satisfies 

1115: no formula of rank strictly greater than $l$.

1116: The model $m$ does not

1117: satisfy \mbox{$a \wedge \neg b$}, and therefore satisfies $b$.

1118: 

1119: Suppose, now, that all propositional models that satisfy $a$ and are

1120: minimal in the $\ll$ ordering for that property, also satisfy $b$.

1121: Let the rank of $a$ be $k$. Since there is a propositional model

1122: that satisfies $a$ and violates no default of rank greater or equal to

1123: $k$, all models that satisfy $a$ and violate no default of rank greater

1124: or equal to $k$ satisfy $b$. We conclude that the rank of 

1125: \mbox{$a \wedge \neg b$} is greater than $k$.

1126: \QED

1127: We may now show that the lexicographic closure is a superset of the rational

1128: closure, at least for defaults of finite rank, thus almost complying with

1129: Thesis~5.25 of~\cite{LMAI:92}.

1130: \begin{theorem}

1131: \label{the:Thesis5.2}

1132: If $a$ has a finite rank and \mbox{$a \NIm b$} is an element of the rational

1133: closure of $D$, then it is an element of its lexicographic closure.

1134: \end{theorem}

1135: \proof

1136: Suppose $a$ has a finite rank and \mbox{$a \NIm b$} is an element of 

1137: the rational closure of $D$.

1138: By Theorem~\ref{the:ll}, \mbox{$a \NIm b$} is satisfied in the modular

1139: model defined by $\ll$. It is therefore satisfied in any modular model

1140: defined by a finer relation. We noticed, just following 

1141: Definition~\ref{def:tnumber2}, that $\prec$ is such a finer relation.

1142: \QED

1143: A characterization of the lexicographic closure in terms of bases will be 

1144: described now.

1145: \subsection{Bases}

1146: \begin{definition}

1147: \label{def:base}

1148: Let $a$ be a formula, and $B$ a subset of $D$. We shall say that

1149: $B$ is a {\em basis} for $a$ iff $a$ is consistent with $\tilde{B}$,

1150: the material counterpart of $B$, and, $B$ is maximal {\em with respect to

1151: the seriousness ordering} for this property.

1152: \end{definition}

1153: The following lemma may help explain the structure of bases, but is not

1154: used in the sequel.

1155: \begin{lemma}

1156: \label{le:bases}

1157: If $a$ has rank $i$ ($a$ has no rank is understood as $a$ 

1158: having an infinite rank)

1159: and $B$ is a basis for $a$, then, for any 

1160: \mbox{$j \geq i$}, \mbox{$D_{j} \subseteq B$}.

1161: \end{lemma}

1162: In other terms, any basis for $a$ is full, for all indexes larger or equal

1163: to the rank of $a$.

1164: \proof

1165: Since $a$ has rank $i$, for any \mbox{$j \geq i$}, $a$

1166: is consistent with \mbox{$E_{j}$} and therefore with

1167: \mbox{$ D_{j} \cup B \cap E_{j + 1} \subseteq E_{j}$}.

1168: \QED

1169: \begin{theorem}

1170: \label{the:entail}

1171: The default \mbox{$(a : b)$} is in $D^{l}$, 

1172: the lexicographic closure $D$ iff,

1173: for any basis $B$ for $a$, \mbox{$\tilde{B} , a \models b$}.

1174: \end{theorem}

1175: \proof

1176: The proof is a generalization of that of Theorem~\ref{the:rat}.

1177: Let $k$ be the order of $D$.

1178: Let \mbox{$d_{i} = \mid D_{i} \mid$} for 

1179: \mbox{$i = 0 , \ldots , k-1 , \infty$}.

1180: \begin{enumerate}

1181: \item \label{aaa}

1182: If $a$ is satisfied by some model of seriousness level

1183: \mbox{$(i_{0} , \ldots , i_{k})$},

1184: then, there is a basis for $a$ of level

1185: \mbox{$(d_{\infty} - i_{0} , d_{k-1} - i_{1} , \ldots , d_{0} - i_{k})$},

1186: and all bases for $a$ have this level.

1187: \item \label{bbb}

1188: If $B$ is a basis of level \mbox{$(l_{0} , \ldots , \l_{k})$} for $a$,

1189: then, there is a model of level 

1190: \mbox{$(d_{\infty} - l_{0} , d_{k-1} - l_{1} , \ldots , d_{0} - l_{k})$}

1191: that satisfy $a$, and no model of strictly smaller seriousness satisfy $a$.

1192: \item \label{ccc}

1193: There is no basis for $a$ iff $a$ is a logical contradiction.

1194: \end{enumerate}

1195: 

1196: Suppose, first, that \mbox{$a \NIm b \in D^{l}$}.

1197: If no model satisfies $a$, $a$ is a logical contradiction and

1198: \mbox{$ X , a \models b$} for any $X$.

1199: Suppose, then, that $a$ is satisfied by some model of level 

1200: \mbox{$(i_{0} , \ldots , i_{k})$}, 

1201: but by no model of lower level. Any model of level 

1202: \mbox{$(i_{0} , \ldots , i_{k})$} that satisfies

1203: $a$, satisfies $b$, by hypothesis. 

1204: Let $B$ be a basis for $a$, of seriousness 

1205: \mbox{$(b_{\infty} , \ldots , b_{0})$}. 

1206: By remark~(\ref{bb}), \mbox{$ d_{j} \: - b_{k-j} \: = \: i_{k-j}$},

1207: for \mbox{$j = 0 , \ldots , k-1$}, and

1208: \mbox{$d_{\infty} - b_{0} = i_{0}$}.

1209: Any model that satisfies $B$ is obviously of seriousness level 

1210: less or equal to

1211: \mbox{$(d_{\infty} - b_{0} ,  d_{k-1} - b_{1} , \ldots , d_{0} - b_{k})$},

1212: i.e., of level less or equal 

1213: \mbox{$i_{0} , \ldots , i_{k}$}.

1214: Any model that satisfies $B$ and $a$ is therefore of level 

1215: \mbox{$i_{0} , \ldots , i_{k}$},

1216: and satisfies $b$, by hypothesis. 

1217: We conclude that \mbox{$B , a \models b$}.

1218: 

1219: Suppose, now, that for any basis $B$ for $a$ we have

1220: \mbox{$B , a \models b$}.

1221: If there is no basis for $a$, then, by remark~(\ref{cc}), 

1222: $a$ is a logical contradiction and \mbox{$a \NIm b \in D^{l}$}.

1223: Suppose, then, the bases for $a$ are of seriousness

1224: \mbox{$(b_{0} , \ldots , b_{k})$}.

1225: There is, by remark~(\ref{bb}),  

1226: a model of $a$ of level 

1227: \mbox{$l = (d_{\infty} - b_{0} , \ldots , d_{0} - b_{k})$}, 

1228: and there is no model of level less than $l$ that satisfies $a$.

1229: We must show that any model of $a$ of level $l$ satisfies $b$.

1230: Let $m$ be such a model. Since $m$ violates $d_{j} - b_{k-j}$ defaults

1231: of rank $j$, 

1232: it satisfies a set $M$ containing $b_{k-j}$ defaults of rank $j$. 

1233: But $M$ is consistent with $a$,

1234: since $m \models a$. The seriousness of $M$ is 

1235: \mbox{$(b_{0} , \ldots , b_{k})$}, the seriousness

1236: of the bases for $a$.

1237: Therefore $M$ is a basis for $a$ and,

1238: since \mbox{$M , a \models b$}, we conclude that \mbox{$m \models b$}.

1239:  

1240: \QED

1241: We shall now describe the lexicographic closure of a number

1242: of sets of defaults, some of them well-known from the literature.

1243: \section{Examples}

1244: \label{sec:exc}

1245: %Exceptions. More exceptional defaults have precedence.

1246: 

1247: In this Section, motivating examples will be described, indicating,

1248: for each of them the conclusions endorsed by the lexicographic closure.

1249: My goal is to, gradually, convince the reader, that each one of the 

1250: decisions taken in the process of defining the lexicographic closure was

1251: reasonable.

1252: My goal is {\em not} to convince the reader that lexicographic

1253: closure provides the {\em intuitively} correct answer once the propositional

1254: variables have been interpreted in some manner that is well-known in the 

1255: folklore of the field, because I believe that, in most cases, we intuitively

1256: treat interpreted formulas in a meaning-dependent manner, not in the formal,

1257: meaning-independent way that is the hallmark of logic.

1258: In other terms, once the variables are interpreted, there is no way of

1259: knowing whether the intuitive conclusions come from formal logical 

1260: considerations or from world knowledge the reasoner has about the situations

1261: or objects the interpreted variables refer to.

1262: A first example exemplifies why one needs to give precedence to defaults

1263: describing exceptional cases over those that describe more normal cases.

1264: \begin{example}[Exceptions]

1265: {\rm

1266: Let \mbox{$D = \{ ( : p ) , ( : q ) , ( \neg p : \neg q )\}$}.

1267: Here $p$ and $q$ hold by default, and when $p$ does not hold, then,

1268: by default $q$ does not hold either.

1269: Suppose we know that $p$ does not hold. Then, obviously we cannot

1270: use the first default. But, we could use the second one to conclude $q$

1271: or the third one to conclude $\neg q$. We obviously want to presume

1272: $\neg q$, and we need to say that the third default has precedence

1273: over the second one. Fortunately, it is not difficult to justify

1274: why the third default has precedence over the other ones.

1275: It (i.e., its antecedent) 

1276: has rank one

1277: whereas the two other defaults have rank zero.

1278: This comes from the fact that the antecedents of the first two defaults

1279: ({\bf true}), describe some unexceptional situation, while $\neg p$

1280: describes an unexpected, exceptional situation since $p$ is, by default,

1281: presumed to be true.

1282: 

1283: The technical description of the lexicographic closure follows.

1284: The order of $D$ is two.

1285: The first two defaults of $D$ have rank zero, the last one has rank one.

1286: The most normal models (those that have level zero), are those that

1287: satisfy $p$ and $q$ (and therefore \mbox{$\neg p \ra \neg q$}).

1288: On level one, we find those models that satisfy \mbox{$\neg p \ra \neg q$}

1289: and violate exactly one of $p$ or $q$: the models that satisfy 

1290: $p$ and $\neg q$.

1291: The third level contains those models that violate both $p$ and $q$,

1292: but satisfy \mbox{$\neg p \ra \neg q$}, i.e. the models satisfying

1293: $\neg p$ and $\neg q$.

1294: The fourth level contains all models violating \mbox{$\neg p \ra \neg q$},

1295: but satisfying $p$ and $q$; it is empty.

1296: The fifth level contains all models violating \mbox{$\neg p \ra \neg q$},

1297: and exactly one of $p$ and $q$: it contains one model.

1298: The sixth level is empty.

1299: }

1300: \end{example}

1301: So, we must give precedence to defaults of higher rank over defaults of

1302: lower rank.

1303: Notice that rank is really all the difference between the default

1304: \mbox{$( a : b )$} and the default \mbox{$( : a \ra b )$}.

1305: The second one has always rank zero, while the first one may have a much higher

1306: rank (if $a$ is presumed to be false) and is therefore more powerful.

1307: The reader may easily check that the first two defaults of $D$ have rank

1308: zero, whereas the third one has rank one.

1309: Let us now treat the similar but more classical penguin example.

1310: \begin{example}[Penguins]

1311: {\rm

1312: Let \mbox{$D = \{ ( p : q ) , (r : p) , (r : \neg q ) \}$}.

1313: The default \mbox{$( : \neg r )$} is entailed by $D$ and

1314: the second and third defaults have rank one, 

1315: whereas the first default has

1316: rank zero.

1317: The defaults \mbox{$( r : \neg q )$} and \mbox{$( p \wedge r : \neg q )$}

1318: are entailed by $D$ whereas \mbox{$( r : q )$} is not.

1319: 

1320: Technically, the rank of $D$ is two.

1321: At level zero: all models satisfying $p \ra q$ and $\neg r$.

1322: At level one: all models satisfying $p$ and $\neq q$, and those 

1323: satisfying $r$, $\neg p$ and $\neg q$.

1324: Level two is empty.

1325: Level three: all models satisfying $r$, $q$ and $p$.

1326: Level four: all models satisfying $r$, $q$ and $\neg p$.

1327: Level five: empty.

1328: }

1329: \end{example}

1330: In the present proposal, 

1331: this precedence of defaults of higher rank is in a sense 

1332: (or in two ways) absolute: one should not trade the violation of a default

1333: of rank $n + 1$ for the violation of any number of defaults of rank less than

1334: or equal to $n$. On this point, the present proposal is in disagreement with

1335: the Principle of Maximal Entropy as proposed in~\cite{GMP:93}.

1336: The following example will exhibit this disagreement, it is not meant

1337: to support one construction against the other.

1338: \begin{example}[Winged Penguins]

1339: {\rm

1340: Let \mbox{$D = \{ ( b : w ) , ( b : f ) , ( p : b ) , ( p : \neg f ) \}$}.

1341: The default \mbox{$( p \wedge ( f \vee \neg w ) : b )$} is entailed

1342: by $D$, since the only basis for \mbox{$ p \wedge ( f \vee \neg w )$} is

1343: the set \mbox{$\{( p : b ) , ( p : \neg f )\}$}, 

1344: containing all defaults of rank one.

1345: In fact, even 

1346: \mbox{$( p \wedge ( f \vee \neg w ) : b \wedge \neg f \wedge \neg w)$}

1347: is entailed by $D$.

1348: But the Principle of Maximal Entropy of~\cite{GMP:93} will consider 

1349: as equivalent

1350: \begin{itemize}

1351: \item

1352: to violate two defaults of rank zero: 

1353: (\mbox{$ ( b : w )$} and \mbox{$ ( b : f )$},

1354: and 

1355: \item to violate one default of rank one: \mbox{$( p : b )$}, 

1356: \end{itemize}

1357: and therefore

1358: will not accept \mbox{$( p \wedge ( f \vee \neg w ) : b )$}.

1359: Notice that the set $D$ is Minimal Core in the sense of~\cite{GMP:93}.

1360: }

1361: \end{example}

1362: \begin{example}[Exceptions again]

1363: {\rm

1364: Let 

1365: \[

1366: D = \{ ( : r ) , ( : p ) , ( : q ) , ( \neg p : \neg q ), 

1367: ( \neg p : \neg r ) \}.

1368: \]

1369: Suppose our specific information is \mbox{$\neg p \wedge q$},

1370: which means the situation is doubly exceptional: $p$ is presumed true

1371: but is in fact false, and when $p$ is false $q$ is presumed false, but

1372: it is true. In other words the rank of \mbox{$\neg p \wedge q$} is two.

1373: Should we presume $r$ to be true or false? 

1374: It is clear we should presume it false, since the default 

1375: \mbox{$( \neg p : \neg r )$} talks about a situation closer to the

1376: one at hand than the default \mbox{$( : r )$}, and should have

1377: precedence over it. But notice that our information shows that

1378: the situation described by \mbox{$\neg p \wedge q$}

1379: is exceptional with respect to the one described by

1380: \mbox{$ \neg p $}.

1381: }

1382: \end{example}

1383: 

1384: \section{Discussion}

1385: \label{sec:properties}

1386: The lexicographic closure $D^{l}$ of a finite set $D$ of defaults

1387: is defined by a modular model in which all propositional models

1388: appear, at some level.

1389: An assertion of the form \mbox{$a \NIm {\bf false}$} will appear in

1390: $D^{l}$ only if $a$ is a logical contradition.

1391: In other terms, lexicographic closure is, in the terminology 

1392: of~\cite{Mak:Handbook},

1393: {\em consistency preserving}. This is indeed one of the hallmarks

1394: of Default Reasoning {\em \`{a} la Reiter} 

1395: as attested by the discussion in~\cite{Reiter:80}, 

1396: and in particular Corollary~2.2 there.

1397: 

1398: Since there has been a lot of discussion in the literature, in particular

1399: in~\cite{Gins:87} and \cite{FLMo:91}, of the principle of Transitivity,

1400: it is probably worth a short discussion.

1401: 

1402: The question of Transitivity is: should we accept \mbox{$(a : c)$}

1403: on the basis of the two defaults  \mbox{$(a : b)$} and \mbox{$(b : c)$}?

1404: The answer proposed here, and which follows from our construction, 

1405: is that, if we have both 

1406: \mbox{$(a : b)$} and \mbox{$(b : c)$}, and if $a$ and $b$ are both of the same

1407: rank, then we should also accept \mbox{$(a : c)$}.

1408: Note that, if we accept \mbox{$(a : b)$}, then $b$ has rank lower or equal

1409: to that of $a$: indeed if the rank of $b$ was larger than that of $a$, 

1410: the rational closure of our set of defaults would include

1411: \mbox{$a \vee b \NIm \neg b$} and therefore \mbox{$a \NIm \neg b$},

1412: and, by Theorem~\ref{the:Thesis5.2}, we would accept \mbox{$( a : \neg b )$}

1413: in the lexicographic closure.

1414: If the rank of $a$ is strictly greater than that of $b$,

1415: then we are not guaranteed that\mbox{$(b : c)$} will be part of all bases for

1416: $a$.

1417: 

1418: Given a finite set $D$ of defaults and a default \mbox{$(a : b)$}, 

1419: how difficult is it to decide whether \mbox{$(a : b)$} is entailed by $D$?

1420: This decision seems to require the computation of the ranks of the defaults

1421: of $D$, but this is relatively easy: a quadratic number of satisfiabilty

1422: problems.

1423: It seems that it also requires the consideration of a possibly

1424: large number of subsets of $D$ and seems therefore inherently exponential.

1425: The rational closure construction of~\cite{LMAI:92} provides a quick

1426: and dirty approximation to this construction in the following sense:

1427: if a default belongs to the rational closure it is entailed

1428: (up to a slightly different treatment of formulas that have no rank).

1429: The case in which all defaults have a Horn structure, needs further study.

1430: One may perhaps avoid the exponential blow-up in this case.

1431: 

1432: \section{Acknowledgements}

1433: Comments, corrections and suggestions by 

1434: Michael Freund, Mois\'{e}s Goldszmidt and David Makinson 

1435: are gratefully acknowledged.

1436: Three anonymous referees (not necessarily disjoint from the previous set)

1437: also helped significantly to make this paper more readable.

1438: \begin{thebibliography}{10}

1439: 

1440: \bibitem{BCDLP:lexi}

1441: Salem Benferhat, Claudette Cayrol, Didier Dubois, J\'{e}r\^{o}me Lang, and

1442:   Henri Prade.

1443: \newblock Inconsistency management and prioritized syntax-based entailment.

1444: \newblock In Ruzena Bajcsy, editor, {\em Proceedings of the 13th I.J.C.A.I.},

1445:   pages 640--645, Chamb\'{e}ry, Savoie, France, August 1993. Morgan Kaufmann.

1446: 

1447: \bibitem{CL:comp}

1448: Claudette Cayrol and Marie-Christine Lagasquie-Schiex.

1449: \newblock Comparaison de relations d'inf\'{e}rence non monotone: \'{E}tude de

1450:   complexit\'{e}.

1451: \newblock Technical Report 93.23.R, Institut de Recherche en Informatique de

1452:   Toulouse, September 1993.

1453: 

1454: \bibitem{FLMo:91}

1455: Michael Freund, Daniel Lehmann, and Paul~H. Morris.

1456: \newblock Rationality, transitivity, and contraposition.

1457: \newblock {\em Artificial Intelligence}, 52(2):191--203, December 1991.

1458: \newblock Research Note.

1459: 

1460: \bibitem{Gin:86}

1461: Matthew~L. Ginsberg.

1462: \newblock Counterfactuals.

1463: \newblock {\em Artificial Intelligence}, 30:35--79, 1986.

1464: 

1465: \bibitem{Gins:87}

1466: Matthew~L. Ginsberg.

1467: \newblock {\em Readings in Nonmonotonic Reasoning}, chapter~1, pages 1--23.

1468: \newblock Morgan Kaufmann, Los Altos, CA, 1987.

1469: 

1470: \bibitem{GMP:90}

1471: Mois\'{e}s Goldszmidt, Paul~H. Morris, and Judea Pearl.

1472: \newblock A maximum entropy approach to nonmonotonic reasoning.

1473: \newblock In {\em Proceedings of AAAI-90}, Boston, August 1990.

1474: 

1475: \bibitem{GMP:93}

1476: Mois\'{e}s Goldszmidt, Paul~H. Morris, and Judea Pearl.

1477: \newblock A maximum entropy approach to nonmonotonic reasoning.

1478: \newblock {\em IEEE Transactions of Pattern Analysis and Machine Intelligence},

1479:   1993.

1480: \newblock To Appear.

1481: 

1482: \bibitem{KLMAI:89}

1483: Sarit Kraus, Daniel Lehmann, and Menachem Magidor.

1484: \newblock Nonmonotonic reasoning, preferential models and cumulative logics.

1485: \newblock {\em Artificial Intelligence}, 44(1--2):167--207, July 1990.

1486: 

1487: \bibitem{L:priv}

1488: Marie-Christine Lagasquie-Schiex, December 1993.

1489: \newblock Private Communication.

1490: 

1491: \bibitem{Leh:89}

1492: Daniel Lehmann.

1493: \newblock What does a conditional knowledge base entail?

1494: \newblock In Ron Brachman and Hector Levesque, editors, {\em Proceedings of the

1495:   First International Conference on Principles of Knowledge Representation and

1496:   Reasoning}, Toronto, Canada, May 1989. Morgan Kaufmann.

1497: 

1498: \bibitem{LehTech:92}

1499: Daniel Lehmann.

1500: \newblock Another perspective on default reasoning.

1501: \newblock Technical Report TR 92-12, The Leibniz Center for Computer Science,

1502:   Institute of Computer Science, Hebrew University, Jerusalem, July 1992.

1503: \newblock Presented at the Bar-Ilan Symposium on Foundations of AI, June 1993.

1504: 

1505: \bibitem{LMAI:92}

1506: Daniel Lehmann and Menachem Magidor.

1507: \newblock What does a conditional knowledge base entail?

1508: \newblock {\em Artificial Intelligence}, 55(1):1--60, May 1992.

1509: 

1510: \bibitem{Mak:Handbook}

1511: David Makinson.

1512: \newblock General patterns in nonmonotonic reasoning.

1513: \newblock In D.~M. Gabbay, C.~J. Hogger, and J.~A. Robinson, editors, {\em

1514:   Handbook of Logic in Artificial Intelligence and Logic Programming}, volume

1515:   Vol. 3, Nonmonotonic and Uncertain Reasoning, pages 35--110. Oxford

1516:   University Press, 1994.

1517: 

1518: \bibitem{Nebel:syntax}

1519: Bernhard Nebel.

1520: \newblock Belief revision and default reasoning: Syntax-based approaches.

1521: \newblock In J.~Allen, R.~Fikes, and E.~Sandewall, editors, {\em Proceedings of

1522:   the Second International Conference on Principles of Knowledge Representation

1523:   and Reasoning}, pages 417--428. Morgan Kaufmann, 1991.

1524: 

1525: \bibitem{Poole:88}

1526: David Poole.

1527: \newblock A logical framework for default reasoning.

1528: \newblock {\em Artificial Intelligence}, 36:27--47, 1988.

1529: 

1530: \bibitem{Reiter:80}

1531: Raymond Reiter.

1532: \newblock A logic for default reasoning.

1533: \newblock {\em Artificial Intelligence}, 13:81--132, 1980.

1534: 

1535: \bibitem{ReiterCri:83}

1536: Raymond Reiter and Giovanni Criscuolo.

1537: \newblock Some representational issues in default reasoning.

1538: \newblock {\em International Journal of Computers and Mathematics with

1539:   Applications}, 9:15--27, 1983.

1540: 

1541: \bibitem{Stal:92}

1542: Robert~C. Stalnaker.

1543: \newblock What is a nonmonotonic consequence relation?

1544: \newblock In {\em Fourth International Workshop on Nonmonotonic Reasoning},

1545:   Plymouth, Vermont, May 1992.

1546: 

1547: \end{thebibliography}

1548: 

1549: \end{document} 

1550: 

1551: \end

1552: