1: \documentclass[a4paper]{article}

2: 

3: % compilation avec latex ou pdflatex

4: \newif\ifpdf

5: \ifx\pdfoutput\undefined

6: \pdffalse

7: \else

8: \pdfoutput=1

9: \pdftrue

10: \fi

11: 

12: \usepackage{amssymb}

13: \ifpdf

14: \usepackage{color}

15: \usepackage{garamond}

16: \renewcommand{\sfdefault}{phv}

17: \renewcommand{\ttdefault}{pcr}

18: \usepackage[pdftex]{graphicx}

19: \pdfcompresslevel=9

20: \usepackage{hyperref}

21: \else

22: \newcommand\hyperpage[1]{#1}

23: \usepackage[dvips]{graphicx}

24: \fi

25: \usepackage{theorem}

26: \usepackage[ruled,vlined]{algorithm2e}

27: 

28: % Theorems 

29: \theoremheaderfont{\scshape}

30: \newtheorem{definition}{Definition}[section]

31: \newtheorem{property}[definition]{Property}

32: \newtheorem{theorem}[definition]{Theorem}

33: \newtheorem{lemma}[definition]{Lemma}

34: \newtheorem{corollary}[definition]{Corollary}

35: \newtheorem{example}{Example}

36: \newtheorem{problem}{Problem}

37: 

38: % Proofs with QED

39: \def\proof{\vskip1sp\noindent{\bf Proof:}\enskip\relax\ignorespaces}

40: \def\endproof{\penalty-250%

41: \hbox to 6pt{\hfill}\hfill\llap{%

42: \vbox{\hrule\hbox{\vrule height6pt\hskip6pt\vrule}\hrule}}\medskip}

43: 

44: % Algorithm2e

45: \SetKwFunction{Proj}{Project}

46: \SetKwFunction{ProjAC}{AC-Project}

47: \SetKwFunction{Ext}{Extend}

48: \SetKwFunction{ExtAC}{AC-Extend}

49: \SetKwFunction{GAC}{GAC}

50: \SetKwFunction{DAC}{DAC}

51: \SetKwFor{Procedure}{Procedure}{}{}%

52: 

53: \renewcommand{\FuncSty}[1]{\textsf{#1}}

54: \renewcommand{\ArgSty}[1]{\textrm{#1}}

55: 

56: % Various symbols

57: \def\setN{{\mathbb{N}}}

58: \def\proj{{\!\downarrow\!}}

59: 

60: % greatest lower absorbing and lowest upper absorbing - To be used in math mode.

61: \newcommand{\gla}[1]{\underline{#1}}

62: \newcommand{\lua}[1]{\overline{#1}}

63: 

64: \title{Arc consistency for soft constraints}

65: 

66: \author{Martin Cooper\\

67: \textsf{cooper@irit.fr}\\

68: IRIT, France \and

69: Thomas Schiex\\

70: \textsf{tschiex@toulouse.inra.fr}\\

71: INRA, France}

72: \date{}

73: \begin{document}

74: 

75: \maketitle

76: 

77: \begin{abstract}

78:   The notion of arc consistency plays a central role in constraint

79:   satisfaction. It is known

80:   since~\cite{Schiex95a,Bistarelli95,Bistarelli97} that the notion of

81:   local consistency can be extended to constraint optimisation

82:   problems defined by soft constraint frameworks based on an

83:   idempotent cost combination operator. This excludes non idempotent

84:   operators such as $+$ which define problems which are very important

85:   in practical applications such as \textsc{Max-CSP}, where the aim is

86:   to minimize the number of violated constraints.

87:   

88:   In this paper, we show that using a weak additional axiom satisfied

89:   by most existing soft constraints proposals, it is possible to

90:   define a notion of \emph{soft arc consistency} that extends the

91:   classical notion of arc consistency and this even in the case of non

92:   idempotent cost combination operators. A polynomial time algorithm

93:   for enforcing this soft arc consistency exists and its space and

94:   time complexities are identical to that of enforcing arc consistency

95:   in CSPs when the cost combination operator is strictly monotonic

96:   (for example \textsc{Max-CSP}).

97:   

98:   A directional version of arc consistency, first introduced

99:   in~\cite{CooperFCSP} is potentially even stronger than the

100:   non-directional version, since it allows non local propagation of

101:   penalties. We demonstrate the utility of directional arc consistency

102:   by showing that it not only solves soft constraint problems on

103:   trees, but that it also implies a form of local optimality, which we

104:   call arc irreducibility.

105: \end{abstract}

106: 

107: \section*{Introduction}

108: 

109: Compared to other combinatorial optimisation frameworks, the CSP

110: framework is essentially characterised by the ubiquitous use of

111: so-called local consistency properties and enforcing algorithms among

112: which arc consistency is certainly preeminent.

113: 

114: The notion of local consistency can be  characterised by a set

115: of desirable properties: 

116: \begin{itemize}

117: \item local consistency is a relaxation of consistency, which means

118:   that for any consistent CSP there is an equivalent non empty locally

119:   consistent CSP.

120: \item this equivalent locally consistent CSP, which is unique, can be

121:   found in polynomial time by so-called enforcing or filtering

122:   algorithms.

123: \end{itemize}

124: Several papers have tried to extend the classical notion of arc

125: consistency to weighted constraint frameworks. In such frameworks, the

126: aim is to find an assignment that minimises combined violations. The

127: first work in this direction is probably~\cite{Rosenfeld76} which

128: defined arc consistency filtering for conjunctive (max-min) fuzzy CSP.

129: 

130: This extension was rather straightforward and one might be tempted to

131: think that this would be the case for other frameworks such as

132: \textsc{Max-CSP}, introduced in~\cite{Shapiro81,Freuder92}, where the

133: aim is to find an assignment which minimises the (weighted) number of

134: violated constraints.  This turned out not to be the case. Later works

135: tried to extend arc consistency in a systematic way using axiomatic

136: frameworks to characterise the properties of the operator used to

137: combine violations:

138: \begin{itemize}

139: \item the Semi-Ring CSP framework was introduced

140:   in~\cite{Bistarelli95,Bistarelli97}. In this work, the extension of

141:   arc consistency enforcing is induced by a generalisation of the

142:   fundamental relational operators such as projection, intersection

143:   and join.  The essential conclusion of this work is that extended

144:   arc consistency works as long as the operator used to combine

145:   violations is idempotent.  This includes the case of conjunctive

146:   fuzzy CSP (in which we try to minimise the violation of the most

147:   violated constraint) and also some other cases with partial orders.

148:   For \textsc{Max-CSP} and other related cases, the algorithm may not

149:   terminate and may also provide non equivalent CSPs.

150:   

151: \item the Valued CSP framework was introduced in~\cite{Schiex95a}.

152:   Here, the extension of the arc consistency property is essentially

153:   based on the notion of relaxation. The same conclusion as in the

154:   Semi-Ring CSP framework was reached for idempotent operators. For

155:   other frameworks such as \textsc{Max-CSP}, it was shown that the

156:   problem of checking the extended arc consistency property defines an

157:   \textsf{NP}-complete problem.

158: \end{itemize}

159: Parallel to these tentative extensions of arc consistency, other

160: research such as~\cite{Wallace94,Bennaceur98,Schiex98,Schiex99} tried

161: to provide improved lower bounds for \textsc{Max-CSP}. The idea of

162: extending arc consistency was abandoned in order to simply provide the

163: most important service, \textit{i.e.} the ability to detect that a CSP has no

164: solution whose cost is below a given threshold.

165: 

166: Globally, each of these proposals violates some of the desirable

167: properties of local consistency. In this paper we show that it is

168: possible, by the addition to the Valued CSP framework of a single

169: axiom, to define an extended arc consistency notion that has all the

170: desirable properties of classical arc consistency except for the

171: uniqueness of the arc consistency closure. It has also the pleasant

172: property that in the idempotent operator cases, it reduces to existing

173: working definitions and uniqueness is recovered.

174: 

175: It has been shown~\cite{Schiex00b} that a lower bound can easily be

176: built from any of the arc consistency closures and that this lower

177: bound generalises and improves upon existing lower

178: bounds~\cite{Wallace94,Bennaceur98,Schiex98,Schiex99}. In this paper,

179: we also consider a directional version of arc consistency that

180: improves lower bounds by propagating partial inconsistencies and not

181: only value deletions as~\cite{Schiex98,Schiex99}. In fact, we show

182: that directional arc consistency, first defined in~\cite{CooperFCSP},

183: defines a locally optimal lower bound.

184: 

185: \section{Notations and definitions}

186: 

187: A constraint satisfaction problem (CSP) is a triple $\langle X,D,C\rangle$. $X$

188: is a set of $n$ variables $X=\{1,\ldots,n\}$. Each variable $i \in X$ has a

189: domain of values $d_i \in D$ and can be assigned any value $a \in d_i$,

190: also noted $(i,a)$. $d$ will denote the cardinality of the largest

191: domain of a CSP. $C$ is a set of constraints. Each constraint $c_P \in

192: C$ is defined over a set of variables $P \subseteq X$ (called the scope of

193: the constraint) by a subset of the Cartesian product $\prod_{i\in P} d_i$

194: which defines all consistent tuples of values.  The cardinality $|P|$

195: is the arity of the constraint $c_P$. $r$ will denote the largest

196: arity of a CSP. We assume, without loss of generality, that at most

197: one constraint is defined over a given set of variables. The set $C$

198: is partitioned into two sets $C = C^1 \cup C^+$ where $C^1$ contains all

199: unary constraints.  For simplification, the unary constraint on

200: variable $i$ will be denoted $c_i$, binary constraints being denoted

201: $c_{ij}$. $e=|C^+|$ will denote the number of non unary constraints in

202: a CSP. If $J\subseteq X$ is a set of variables, then $\ell(J)$ denotes the

203: set of all possible labellings for $J$ \textit{i.e.}, the Cartesian

204: product $\prod_{i\in J}d_i$ of the domains of the variables in $J$. The

205: projection of a tuple of values $t$ onto a set of variables $V \subseteq X$

206: is denoted by $t_{\proj V}$.  A tuple of values $t$ satisfies a

207: constraint $c_P$ if $t_{\proj P} \in c_P$.  Finally, a tuple of values

208: over $X$ is a solution iff it satisfies all the constraints in $C$.

209: 

210: 

211: \section{Valued CSP}

212: 

213: Valued CSP (or VCSP) were initially introduced in~\cite{Schiex95a}. A

214: valued CSP is obtained by associating a valuation with each

215: constraint. The set $E$ of all possible valuations is assumed to be

216: totally ordered and its maximum element is used to represent total

217: inconsistency.  When a tuple violates a set of constraints, its

218: valuation is computed by combining the valuations of all violated

219: constraints using an aggregation operator, denoted by $\oplus$. This

220: operator must satisfy a set of properties that are captured by a set

221: of axioms defining a so-called \emph{valuation structure}.

222: 

223: \begin{definition}

224:   A {\em valuation structure} is defined as a tuple $\langle E,\oplus,\succcurlyeq\rangle$ such

225:   that:

226:   \begin{itemize}

227:   \item $E$ is a set, whose elements are called valuations, which is

228:     totally ordered by $\succcurlyeq$, with a maximum element denoted

229:     by $\top$ and a minimum element denoted by $\bot$;

230:   \item $E$ is closed under a commutative, associative binary

231:     operation $\oplus$ that satisfies:

232:     \begin{itemize}

233:     \item {\em Identity}\/: $\forall \alpha \in E, \alpha \oplus \bot = \alpha$;

234:     \item {\em Monotonicity}\/: $\forall \alpha,\beta,\gamma  \in E, (\alpha \succcurlyeq \beta) \Rightarrow

235:       \bigl((\alpha \oplus \gamma) \succcurlyeq (\beta \oplus \gamma )\bigr)$;

236:     \item {\em Absorbing element}: $\forall \alpha \in E, (\alpha \oplus \top)=\top$.

237:     \end{itemize}

238:   \end{itemize}

239: \end{definition}

240: 

241: When $E$ is restricted to $[0,1]$, this structure of a totally ordered

242: commutative monoid with a monotonic operator is also known in

243: uncertain reasoning, as a triangular co-norm~\cite{DPnormes}.

244: % In the rest of the

245: %paper, we assume that the computation of $\succ$ and $\oplus$ are always polynomial

246: %in the size of their arguments. We also assume $E$ is countable.  

247: 

248: It is now possible to define valued CSPs. Note that, for the sake of

249: generality, rather than considering that a valuation is associated

250: with each constraint, as in~\cite{Schiex95a}, we consider that a

251: valuation is associated with each tuple of each constraint. As

252: observed in~\cite{Bistarelli99}, the two approaches are essentially

253: equivalent.

254: 

255: \begin{definition}

256:   A valued CSP is a tuple $\langle X,D,C,S\rangle$ where $X$ is a set of $n$

257:   variables $X=\{1,\ldots,n\}$, each variable $i \in X$ has a domain of

258:   possible values $d_i \in D$.  $C = C^1 \cup C^+$ is a set of

259:   constraints and $S = \langle E,\oplus,\succcurlyeq\rangle$ is a valuation

260:   structure. Each constraint $c_P \in C$ is defined over a set of

261:   variables $P \subseteq X$ as a function $c_P: \prod_{i\in P} d_i \to E$.

262: \end{definition}

263: 

264: An assignment $t$ of values to some variables $J \subseteq X$ can be simply

265: evaluated by combining, for all assigned constraints $c_P$ (i.e., such

266: that $P \subseteq J$), the valuations of the projection of the tuple $t$ on $P$:

267: \begin{definition}

268:   In a VCSP $V=\langle X,D,C,S\rangle$, the valuation of an assignment

269:   $t$ to a set of variables $J \subseteq X$ is defined by:

270:   \[\mathcal{V}_{V}(t) = \mathop\bigoplus_{c_P \in C, P \subseteq J}[c(t_{\proj P})] \]

271: \end{definition}

272: 

273: The problem usually considered is to find a complete assignment with a

274: minimum valuation.  Globally, the semantics of a VCSP is defined by

275: the valuations $\mathcal{V}(t)$ of assignments $t$ to $X$.

276: 

277: The choice of axioms is quite natural and is usual in the field of

278: uncertain reasoning. The ordered set $E$ simply allows us to express

279: different degrees of constraint violation. The commutativity and

280: associativity guarantee that the valuation of an assignment is

281: independent of the order in which valuations are combined. The

282: monotonicity of $\oplus$ guarantees that assignment valuations cannot

283: decrease when constraint violations increase.  For a more detailed

284: analysis and justification of the VCSP axioms, we invite the reader to

285: consult~\cite{Schiex95a,Schiex99} which also emphasise the difference

286: between idempotent and strictly monotonic aggregation operators $\oplus$.

287: \begin{definition} An operator $\oplus$ is

288:   idempotent if $\,\forall \alpha \in E, (\alpha \oplus \alpha) = \alpha$. It is strictly

289:   monotonic if $\,\forall \alpha,\beta,\gamma \in E, (\alpha \succ \beta) \land (\gamma \neq \top )\Rightarrow (\alpha \oplus

290:   \gamma) \succ (\beta \oplus \gamma)$

291: \end{definition}

292: As shown in~\cite{Schiex95a}, these two properties are incompatible as

293: soon as $|E| > 2$. The only valuation structures with an idempotent

294: operator correspond to classical and possibilistic CSP~\cite{Schiex92}

295: (min-max dual to the conjunctive fuzzy CSP framework) which use $\oplus =

296: \max$ as the aggregation operator.  Other soft CSP frameworks such as

297: \textsc{Max-CSP}, lexicographic CSP or probabilistic CSP use a

298: strictly monotonic operator.

299: 

300: Arc consistency enforcing must yield an equivalent problem, the

301: so-called arc-consistency closure. Several notions of equivalence were

302: introduced in~\cite{Schiex95a,Schiex99} that enabled us to compare

303: pairs of VCSP with different valuations structure.  In this paper, the

304: notion of equivalence will only be used to compare pairs of VCSP with

305: the same valuation structure and can therefore be simplified and

306: strengthened. 

307: \begin{definition}

308:   Two VCSP $V=\langle X,D,C,S\rangle$ and $V'=\langle X,D,C',S\rangle$ are equivalent iff

309:   for all complete assignment $t$ to $X$, we have:

310:   \[\mathcal{V}_{V}(t) = \mathcal{V}_{V'}(t) \]

311: \end{definition}

312: 

313: \section{Fair valuation structures}

314: 

315: We start with an introductory example. In the remainder of the paper,

316: in order to illustrate the notions introduced on concrete examples, we

317: will consider binary weighted \textsc{Max-CSP}s which correspond to

318: valued CSPs using the strictly monotonic valuation structure $\langle\setN \cup

319: \{\infty\}, +, \geq\rangle$.  To describe such problems, we use an undirected graph

320: representation where vertices represent values. For all pairs of

321: variables $i,j \in X$ such that $c_{ij}\in C$, for all values $a\in d_i$,

322: $b\in d_j$ such that $c_{ij}(a,b) \neq \bot =0$, an edge connect the values

323: $(i,a)$ and $(j,b)$. The weight of this edge is set to $c_{ij}(a,b)$.

324: Unary constraints are represented by weights associated with vertices,

325: weights equal to $0$ being omitted.

326: 

327: Let us consider the weighted \textsc{Max-CSP} in

328: figure~\ref{maxcsp1}(a). It has two variables numbered $1$ and $2$,

329: each with two values $a$ and $b$ together with a single constraint.

330: The constraint forbids pair $((1,b),(2,b))$ with cost $1$ and forbids

331: pairs $((1,a),(2,a))$ and $((1,b),(2,a))$ completely (with cost $\infty$).

332: The pair $((1,a),(2,b))$ is completely authorised and the

333: corresponding edge is therefore omitted.

334: 

335: \begin{figure}[bthp]

336:   \begin{center}

337:     \includegraphics{maxcsp}

338:     \end{center}

339: \caption{Four equivalent instances of \textsc{Max-CSP}\label{maxcsp1}}

340: \end{figure}

341: 

342: If we assign the value $b$ to variable $1$, it is known for sure that

343: a cost of $1$ must be paid since all extensions of $(1,b)$ to variable

344: $2$ incur a cost of at least $1$. Projecting this minimum cost down

345: from c$_{12}$ would make this explicit and induce a unary constraint

346: on $1$ that forbids $(1,b)$ with cost $1$. However if we simply add

347: this constraint to the \textsc{Max-CSP}, as was proposed

348: in~\cite{Bistarelli95} for problems with an idempotent operator, the

349: resulting CSP is not equivalent. The complete assignment

350: $((1,b),(2,b))$ which initially had a cost of $1$ would now have a

351: cost of $2$.  In order to preserve equivalence, we must ``compensate''

352: for the induced unary constraint. This can be done by simply

353: subtracting $1$ from all the tuples that contain the value $(1,b)$.

354: The corresponding equivalent CSP is shown in figure~\ref{maxcsp1}(b):

355: the edge $((1,b),(2,b))$ of cost $1$ has disappeared (the associated

356: weight is now $0$) while the edge $((1,b),(2,a))$ is unaffected since

357: it has infinite weight.  We can repeat this process for variable $2$:

358: all extensions of value $(2,a)$ have infinite cost. Thus we can add a

359: unary constraint that completely forbids value $(2,a)$. In this

360: specific case, and because the valuation $\infty$ satisfies $\infty \oplus \infty =

361: \infty$, we can either compensate for this (Figure~\ref{maxcsp1}(c)) or

362: not (Figure~\ref{maxcsp1}(d)).  In both cases, an equivalent

363: \textsc{Max-CSP} is obtained. Between the problems in

364: Figure~\ref{maxcsp1}(c)) and~\ref{maxcsp1}(d), we prefer the problem

365: in Figure~\ref{maxcsp1}(d) because it makes information explicit both

366: at the domain and constraint level.

367: 

368: This type of projection mechanism underlies most of the lower bounds

369: defined for

370: \textsc{Max-CSP}~\cite{Wallace94,Bennaceur98,Schiex98,Schiex99}. To

371: our knowledge, the introduction of a ``compensation'' mechanism for

372: preserving equivalence was first introduced by~\cite{Koster99b} on

373: \textsc{Max-CSP}, independently of any notion of arc consistency. The

374: use of such mechanism for the definition and establishment of arc

375: consistency appeared in~\cite{Schiex00b} and in a related form

376: in~\cite{Horsch00} (for enforcing so-called probabilistic arc

377: consistency).

378: 

379: \begin{figure}[bthp]

380:   \begin{center}

381:     \includegraphics{maxcsp2}

382:     \end{center}

383: \caption{Two equivalent instances of \textsc{Max-CSP}\label{fig2}}

384: \end{figure}

385: 

386: Suppose now that the problem in Figure~\ref{maxcsp1} is part of an

387: instance of \textsc{Max-CSP} on four variables, as shown in

388: Figure~\ref{fig2}(a). As in crisp CSP, inconsistencies can propagate

389: from domains up to constraints. The cost of $\infty$ for $(2,a)$ can be

390: duplicated in the costs of the pairs $((2,a),(3,a))$ and

391: $((2,a),(3,b))$. Since $c_{23}(b,b)=\infty$, this in turn implies that

392: the assignment $(b,3)$ inevitably has a cost of $\infty$. No further

393: propagation of infinite costs can be performed.

394: 

395: A similar process can be applied to finite costs but one must take

396: care to compensate any cost change. The cost $1$ of $(1,b)$ can be

397: first shifted to the constraint $c_{01}$: the costs of the pairs

398: $((0,a),(1,b))$ and $((0,b),(1,b))$ become equal to $1$ and the cost

399: of the value $(1,b)$ is set to $0$.  Since now $c_{01}(a,a) =

400: c_{01}(a,b) = 1$, a cost of $1$ can be projected onto value $(0,a)$.

401: Figure~\ref{fig2}(b) shows the result of such propagations.  In the

402: case of finite costs, the process is obviously not terminated since

403: one could forever shift this cost back and forth between values

404: $(0,a)$ and $(1,b)$.

405: 

406: \subsection{A new axiom for VCSPs}

407: 

408: To formalise and generalise the ideas presented in the previous

409: section to other valuation structures, we have to be able to

410: compensate for the information added by projecting weights down onto

411: domains. This is made possible by the following additional axiom:

412: 

413: \begin{definition}\label{fairdef}

414:   In a valuation structure $S=\langle E,\oplus,\succcurlyeq\rangle$, if $\alpha,\beta \in E$,

415:   $\alpha \preccurlyeq \beta$ and there exists a valuation $\gamma \in E$ such

416:   that $\alpha \oplus \gamma = \beta$, then $\gamma$ is known as a difference of $\beta$ and

417:   $\alpha$.

418:   

419:   The valuation structure $S$ is \emph{fair} if for any pair of

420:   valuations $\alpha,\beta\in E$, with $\alpha \preccurlyeq \beta$, there exists a

421:   maximal difference of $\beta$ and $\alpha$. This unique maximal difference

422:   of $\beta$ and $\alpha$ is denoted by $\beta \ominus \alpha$.

423: \end{definition}

424: 

425: \begin{lemma}\label{lemma1}

426:   Let $S=\langle E,\oplus,\succcurlyeq\rangle$ be a fair valuation structure. Then

427:   $\forall u,v,w, \in E, w\preccurlyeq v$, we have $(v \ominus w) \preccurlyeq v$

428:   and $(u \oplus w) \oplus (v \ominus w) = (u \oplus v)$.

429: \end{lemma}

430: \begin{proof} By definition, $(v \ominus w) \oplus w = v$. From the monotonicity of

431:   $\oplus$, this proves that $(v \ominus w) \preccurlyeq v$ (this inequality

432:   becomes strict if $\oplus$ is strictly monotonic and $v\neq \top$). The

433:   second property follows from the commutativity and associativity of

434:   $\oplus$: we have $(u \oplus w) \oplus (v \ominus w) = u \oplus ((v \ominus w) \oplus w) = (u \oplus

435:   v)$.

436: \end{proof}

437: 

438: Most existing concrete soft constraint frameworks, including all those

439: with either an idempotent or strictly monotonic operator $\oplus$ are

440: fair.

441: \begin{example}

442:   If $\oplus$ is idempotent, then it can easily be shown that

443:   $\oplus=\max$~\cite{Schiex95a}. Classical CSPs can be defined as VCSPs

444:   over the valuation structure $S=\langle\{\bot,\top\},\max,\succcurlyeq\rangle$,

445:   where $\bot$ represents \textit{true} and $\bot$ \textit{false}. The

446:   operator $\oplus$ is also idempotent in possibilistic

447:   CSPs~\cite{Schiex92} which define a $\min$-$\max$ problem which is

448:   dual to the $\max$-$\min$ problem of conjunctive fuzzy

449:   CSPs~\cite{Rosenfeld76,CooperFCSP}. When $\oplus=\max$, we have

450:   $\ominus=\max$, since $\max(\max(\alpha,\beta),\alpha) = \beta$ whenever

451:   $\alpha\preccurlyeq \beta$. When $\alpha=\beta$, then any valuation $\gamma\prec \alpha$ is

452:   also a valid difference of $\beta$ and $\alpha$ but it is clearly not

453:   maximal.

454: \end{example}

455: 

456: \begin{example}

457:   In the strictly monotonic valuation structure

458:   $\langle\mathbb{N}\cup\{\infty\},+,\geq\rangle$, $\ominus$ is defined by $\beta \ominus \alpha = \beta -

459:   \alpha$ for finite valuations $\alpha,\beta\in \mathbb{N}, \alpha\leq \beta$ and $(\infty \ominus

460:   \alpha) = \infty$ for all $\alpha \in \mathbb{N}\cup\{\infty\}$. In the general case of

461:   any strictly monotonic operator $\oplus$, the difference operator may

462:   not exist in $E$, but it has been proved in~\cite{CooperFCSP} that

463:   the difference operator can always be constructed by embedding the

464:   valuation structure in a larger valuation structure derived from the

465:   set $E\times E$, where $(\beta,\alpha)$ represents the imaginary $\beta \ominus \alpha$.

466:   This can be compared with embedding $\mathbb{R}$ in $\mathbb{C}$ so

467:   as to allow us to take square roots of negative numbers. This

468:   construction is interesting for lexicographic CSPs~\cite{Schiex93c}

469:   for which differences are not always defined in the original

470:   valuation structure.  Another possible approach is to transform the

471:   lexicographic CSP into a VCSP on the valuation structure

472:   $\langle\mathbb{N}\cup\{\infty\},+,\geq\rangle$ using the simple transformation

473:   described in~\cite{Schiex95a}.

474: \end{example}

475: 

476: %\begin{theorem}[\cite{CooperFCSP}]

477: %  Every strictly monotonic valuation structure

478: %  $S=\langle E,\oplus,\succcurlyeq\rangle$can be embedded in a valuation structure

479: %  $S'=\langle E',\oplus',\succcurlyeq'\rangle$ in which $\oplus'$ has a difference

480: %  operation $\ominus$ satisfying $b\oplus'(a\ominus b) = a$ in $E'$ for all $a,b \in

481: %  E'$ such that $a \succcurlyeq' b$.

482: %\end{theorem}

483: 

484: %\begin{proof} 

485: %  Define the relation $\equiv$ between pairs of elements of $E\times(E-\{\top\})$

486: %  as follows: $\forall a,c \in E, \forall b,d \in (E-\{\top\}), (a,b) \equiv (c,d) \Leftrightarrow

487: %  a\oplus d = b \oplus c$.

488:   

489: %  That $\equiv$ is an equivalence relation follows from the strict

490: %  monotonicity of $\oplus$: for $(a,b), (c,d), (e,f) \in E\times(E-\{\top\})$, we

491: %  have $(a\oplus d = b\oplus c) \land (c\oplus f = d\oplus e) \Rightarrow a\oplus d\oplus c\oplus f = b\oplus c\oplus

492: %  d\oplus e \Rightarrow a\oplus f = b\oplus e$ if $ c\neq \top$. Otherwise, if $c =\top$, this

493: %  implies that $a=e=\top$ and therefore again $a\oplus f = b\oplus e$.

494:   

495: %  Let $E'$ be the set of equivalence classes of $\equiv$ in $\{(a,b)\in

496: %  E\times(E-\{\top\})~|~a \succcurlyeq b\}$. If we identify $a\in E$ with the

497: %  equivalence class of $(a,\bot)\in E '$, we see that $E '$ is an

498: %  extension of $E$. We write $\bot'$ and $\top'$ for (the equivalence

499: %  classes of) $(\bot,\bot)$ and $(\top,\bot)$ in $E'$. Define an order

500: %  $\succcurlyeq'$ by $(a,b) \succcurlyeq' (c,d) \Leftrightarrow a\oplus d \succcurlyeq

501: %  b\oplus c$. It is easy to see that $\succcurlyeq'$ is well-defined. The

502: %  aggregation $\oplus'$ is defined as $(a,b) \oplus' (c,d) = (a\oplus c,b\oplus d)$.

503: %  The difference $(a,b)\ominus (c,d)$ in $E '$ is defined as $(a\oplus d,b\oplus

504: %  c)$ for $(a,b) \in E', (c,d)\in E'-\{\top'\}$ such that $(a,b)

505: %  \succcurlyeq' (c,d)$ in $E '$. $\top'\ominus \top'$ is defined as $\top'$.

506:   

507: %  It remains to show that $S' = \langle E',\oplus',\succcurlyeq'\rangle$ is a fair

508: %  valuation structure. That $\succcurlyeq'$ is a total order follows

509: %  from:\medskip

510: 

511: %\begin{tabular}{ccl}  

512: %  $(a,b) \succcurlyeq'(c,d) \land (c,d) \succcurlyeq' (a,b)$ & $\Rightarrow$ & $a\oplus d = b

513: %  \oplus c \Rightarrow (a,b)\equiv (c,d)$ and \\

514: %$(a,b) \succcurlyeq' (c,d) \land (c,d)\succcurlyeq'(e,f)$& $\Rightarrow$ & $(a\oplus d\succcurlyeq b\oplus c) \land (c\oplus f \succcurlyeq d\oplus e)$\\

515: % &    $\Rightarrow$ & $(a\oplus d \oplus c \oplus f \succcurlyeq b\oplus c \oplus d\oplus e \land c\neq \top) \lor$\\

516: % &             & $(a=c=\top)$\\

517: %&    $\Rightarrow$ & $a\oplus f\succcurlyeq b \oplus e$ (strict monotonicity)\\

518: %&    $\Rightarrow$ & $(a,b) \succcurlyeq' (e,f)$\\

519: %\end{tabular}

520: 

521: %That $\top'$ and $\bot'$ are the maximum and minimum elements of $E '$

522: %follows from the fact that $\forall (a,b)\in E ', \top\succcurlyeq

523: %a\succcurlyeq b$ which can be rewritten $b\oplus\top \succcurlyeq a\oplus \bot

524: %\succcurlyeq b\oplus\bot$ and therefore $(\top,\bot) \succcurlyeq' (a,b)

525: %\succcurlyeq' (\bot,\bot)$.

526: 

527: %The fact that $E '$ is closed under $\oplus$ follows from $(a,b),(c,d)\in

528: %E' \Rightarrow b\neq\top \land d\neq\top \land a\succcurlyeq b \land c\succcurlyeq d \Rightarrow b\oplus d\neq\top \land

529: %a\oplus c\succcurlyeq b\oplus d \Rightarrow (a\oplus c,b\oplus d)\in E '$.

530: 

531: %The commutativity and associativity of $\oplus'$ in $E '$ follow directly

532: %form the associativity and commutativity of $\oplus$ in $E$. Furthermore,

533: %the identity axiom $\forall (a,b) \in E ', (a,b)\oplus' (\bot,\bot) \equiv (a,b)$ is

534: %obviously satisfied. Strict monotonicity follows from the fact

535: %that $(a,b),(c,d),(e,f) \in E' \land (a,b) \succ' (c,d) \land (e,f)\not\equiv \top' \Rightarrow

536: %a\oplus d \succ b\oplus c \land \top \neq e \succcurlyeq f \Rightarrow a\oplus e\oplus d\oplus f \succ b\oplus f \oplus

537: %c \oplus e \Rightarrow (a\oplus e,b\oplus f) \succ '(c\oplus e,d\oplus f) \Rightarrow (a,b)\oplus'(e,f) \succ '

538: %(c,d)\oplus'(e,f)$ and the absorbing element axiom is a logical

539: %consequence of the other axioms in valuation structures

540: %(see~\cite{Schiex95a}).

541: 

542: %Finally, if we consider $(a,b),(c,d)\in E '$ such that $(c,d)\not\equiv

543: %\top'$ and $(a,b)\succcurlyeq' (c,d)$ then $(c,d)\oplus'((a,b)\ominus(c,d)) \equiv

544: %(c,d)\oplus'(a\oplus d,b\oplus c) \equiv (c\oplus a\oplus d,d\oplus b\oplus c) \equiv (a,b)$ by definition

545: %of $\equiv$ since $d\oplus b \oplus c \neq \top$. When $(c,d)\equiv \top'$, we only need to

546: %consider the case $(a,b) \equiv \top'$ and clearly $\top'\oplus (\top'\ominus\top')\equiv \top'$

547: %as required.

548: %\end{proof}

549: 

550: \subsection{Equivalence preserving transformations}

551: 

552: As it has been demonstrated in the examples of Figures~\ref{maxcsp1}

553: and~\ref{fig2}, it is possible to transform a \textsc{Max-CSP} into an

554: equivalent but different \textsc{Max-CSP} using local transformations

555: (involving only one non unary constraint). Such operations will be

556: called equivalence-preserving transformations:

557: 

558: \begin{definition}

559:   The \emph{subproblem} of a VCSP $V=\langle X,D,C,S\rangle$ on $J\subseteq X$ is the

560:   VCSP $V(J) = \langle J,D_J,C_J,S\rangle$, where $D_J=\{d_j~:~j\in J\}$ and $C_J

561:   = \{c_P \in C~:~P\subseteq J\}$.

562: \end{definition}

563: 

564: \begin{definition}

565:   For a VCSP $V$, an \emph{equivalence-preserving transformation} of

566:   $V$ on $J\subseteq X$ is an operation which transforms the subproblem of

567:   $V$ on $J$ into an equivalent VCSP. If $C_J = \{c_P \in C~:~P\subseteq J\}$

568:   contains only one non unary constraint, such an operation is called

569:   an \emph{equivalence-preserving arc transformation}.

570: \end{definition}

571: 

572: \begin{example} 

573:   The procedures \Proj\ and \Ext\, described in Algorithm~\ref{ext}

574:   are examples of equivalence-preserving transformations.

575:   

576:   \Proj\ transform a VCSP by shifting valuations from the tuples

577:   of a given non unary constraint $c_P$ to the value $(i,a)$, where $i\in

578:   P, a\in d_i$. In order to preserve equivalence, any increase at the

579:   unary level is compensated at the tuple level (line~1 of

580:   Algorithm~\ref{ext}).

581:   

582:   Conversely, \Ext\ shifts the valuation from value $(i,a)$ to the

583:   tuples of the constraint $c_P$ where $i\in P$. Again, the fairness of

584:   the valuation structure allows us to compensate for the possible

585:   increase of the tuple valuations by an operation at the unary level

586:   (line~2 of Algorithm~\ref{ext}).

587: \end{example}

588: 

589: \begin{algorithm}[htbp]

590:   \Procedure{\Proj{$c_P,i,a$}}{

591:     $\beta  \gets \min_{t \in \ell(P-\{i\})}(c_P(t,a))$\;

592:     $c_i(a) \gets c_i(a) \oplus \beta$\;

593:     \ForEach{($t\in \ell(P-\{i\})$)}

594:     {\lnl{comp}$c_P(t,a) \gets c_P(t,a) \ominus \beta$\;}

595:     }

596:   \BlankLine

597:   \BlankLine

598:   \Procedure{\Ext{$i,a,c_P$}}{

599:     \ForEach{($t \in \ell(P-\{i\})$)}{

600:       $c_P(t,a) \gets c_P(t,a) \oplus c_i(a)$\;

601:       }

602:     \lnl{abs}    $c_i(a) \gets c_i(a) \ominus c_i(a)$\;

603:     }

604:   \caption{Two basic equivalence-preserving arc-transformations\label{ext}}

605: \end{algorithm}

606: 

607: \begin{theorem}

608:   Given any fair VCSP $V=\langle X,D,C,S\rangle$, for any $c_P \in C^+$, $i\in P$,

609:   $a\in d_i$, the application of \Proj\ or \Ext\ on $V$ yields an

610:   equivalent VCSP.

611: \end{theorem}

612: \begin{proof}

613:   To demonstrate equivalence, it is sufficient to prove that the value

614:   of $c_P(t,a) \oplus c_i(a)$ is an invariant of \Proj{$c_P,i,a$} and

615:   \Ext{$i,a,c_P$}.  For any $t\in\ell(P-\{i\})$, let $\gamma$ be the initial

616:   value of $c_P(t,a)$ and $\delta$ the initial value of $c_i(a)$. After

617:   the execution of \Proj, we have $(c_P(t,a) \oplus c_i(a)) = (\gamma \ominus \beta)

618:   \oplus (\delta \oplus \beta) = \gamma \oplus \delta$.  After the execution of \Ext, we have

619:   $(c_P(t,a) \oplus c_i(a)) = (\gamma \oplus \delta) \oplus (\delta \ominus \delta) = \gamma \oplus \delta$. This

620:   proves the invariances.

621: \end{proof}

622: 

623: As the example of Figure~\ref{fig2} showed in the case of

624: \textsc{Max-CSP}, the iterated application of equivalence-preserving

625: transformations such as \Proj\ and \Ext\ does not necessarily lead to

626: a quiescent state. The two following sections show how a limited

627: application of carefully designed equivalence-preserving

628: transformations can guarantee that a quiescent state will always be

629: reached.

630: 

631: \section{Soft Arc consistency}

632: 

633: In classical CSPs, arc consistency enforcing always increases the

634: information available on each variable.  In the case of soft arc

635: consistency, application of arc transformations will be limited to

636: operations that either increase the information available at the

637: variable level or that increase information available at the

638: constraint level \emph{as long as they do not lower the information

639:   available at the variable level}.  In the next section, we try to

640: better characterise when this is possible.

641: 

642: \subsection{On the structure of valuation structures}

643: 

644: \begin{definition}

645:   In a valuation structure $\langle E,\oplus,\succcurlyeq\rangle$, an element $\alpha \in

646:   E$ is an absorbing element iff $\alpha \oplus \alpha = \alpha$.

647: \end{definition}

648: Absorbing elements can be duplicated without affecting valuations.

649: They can be propagated, in the same way as inconsistencies are in

650: crisp CSPs. Non-absorbing elements $\alpha$ can be shifted from one

651: constraint to another, but each addition of $\alpha$ must be compensated

652: by a subtraction elsewhere.

653: 

654: In a valuation structure $\langle E,\oplus,\succcurlyeq\rangle$, if $\oplus$ is

655: idempotent then all elements of $E$ are absorbing. If $\oplus$ is a

656: strictly monotonic operator then the only absorbing elements are $\bot$

657: and $\top$. Intermediate cases occur in the following examples:

658: %\begin{example}

659: %  Suppose that a student must obtain $10$ first-year course credits to

660: %  be allowed to follow second-year courses, and then must obtain $10$

661: %  second-year course credits to be awarded his/her degree.

662: 

663: %  This can be modelled by a valuation structure of the form:

664: %  \[\{(N_1,0)~:~N_1\in\{0,\ldots,10\}\} \cup \{(10,N_2)~:~N_2\in \{0,\ldots,10\}\}\]

665: 

666: %where $N_1$ is the number of first-year credits, $N_2$ the number of

667: %second-year credits, and

668: 

669: %\begin{eqnarray*}

670: %(N_1,N_2) < (N_1',N_2') & \Leftrightarrow & (N_2 < N_2') \lor (N_2 = N_2' = 0 \land N_1 <N_1')\\

671: %(N_1,N_2) \oplus (N_1',N_2') & = & (\min\{N_1+N_1',10\}, \min\{N_2+N_2',10\})

672: %\end{eqnarray*}

673: 

674: %There are three absorbing values in this valuation structure, namely

675: %$(0,0)$, $(10,0)$ and $(10,10)$. This is a fair valuation structure

676: %since $\oplus$ has the following inverse operation $\ominus$:

677: 

678: %\[\begin{array}{rcll}

679: %(N_1,N_2) \ominus (N_1',N_2') & = & (N_1-N_1',0\} & (\textrm{if}~(N_2=N_2' = 0) \land N_1<10)\\

680: %(N_1,N_2) \ominus (N_1',N_2') & = & (10,N_2-N_2') & (\textrm{if}~N_1=10)\\

681: %\end{array}\]

682: %\end{example}

683: 

684: \begin{example}

685:   Imagine the possible sentences for driving offences. Suppose that

686:   penalty points (up to a maximum of $12$) are awarded for minor

687:   offences, whereas serious offences are penalised by suspension of

688:   the offender's driving license for a period of $y$ years, for some

689:   positive integer $y$. A driver who accumulates $12$ penalty points

690:   receives an automatic one-year suspension of his/her license. The

691:   set of sentences can be modelled by a valuation structure

692:   $S=\langle E,\oplus,\succcurlyeq\rangle$ of the form:

693:   \[E=\{(p,0)~:~p\in\{0,\ldots,12\}\} \cup \{(0,y)~:~y\in\mathbb{N}^*\cup\{\infty\}\} \]

694: 

695:   \begin{eqnarray*}

696:     (p,y) \prec (p',y') & \Leftrightarrow & (y< y') \lor ((y = y'=0) \land (p< p'))\\

697:     &&\\

698:     (p,0)\oplus(p',0) & = & (\min(p+p',12),0)\\

699:     (p,y) \oplus (p',y') & = & (0,y+y')\hspace*{2cm}\textrm{if $(y+y'\neq 0)$}\\

700:   \end{eqnarray*}

701:   Note that $(12,0)\prec(0,1)$ even though they both give rise to a

702:   one-year license suspension. The penalty $(0,1)$ is deemed to be

703:   worse because it can be cumulated. For example $(0,1)\oplus(0,1) =

704:   (0,2)$, whereas $(12,0)\oplus(12,0) = (12,0)$. Apart from $\bot=(0,0)$ and

705:   $\top=(0,\infty)$, this valuation structure contains another absorbing

706:   valuation, namely $(12,0)$. This is a fair valuation structure since

707:   $\oplus$ has the following inverse operation $\ominus$:

708:   \begin{eqnarray*}

709:     (p,0)\ominus (p',0) & = & (p-p',0)\quad\textrm{if $p < 12$}\\

710:     (12,0) \ominus (p',0) & = & (12,0)\\

711:     (0,y) \ominus (p',y') &=& (0,y-y')

712:   \end{eqnarray*}

713: \end{example}

714: 

715: \begin{example}

716:   Another interesting case occurs if, for example, a company wants to

717:   minimise both financial loss $F$ and loss of human life $H$ if a

718:   fire should break out in its factory. Supposing that the company

719:   considers that no price can be put on human life, we must have

720:   \[(F,H) < (F',H') \Leftrightarrow (H<H') \lor (H=H' \land F<F')\]

721: If a financial loss of $F_{\max}$ represents bankruptcy, then

722: \[ (F,H)\oplus (F',H') = (\min\{F+F',F_{\max}\},H+H')\]

723: and $(F_{\max},0)$ is an absorbing element which is strictly less than

724: $\top$. Note that this valuation structure is not fair, since it is

725: impossible to define $\alpha = (0,1)\ominus (F_{\max},0)$ such that $\alpha\oplus

726: (F_{\max},0) = (0,1)$.

727: \end{example}

728: 

729: \begin{example}

730:   Consider a valuation structure $S=\langle\mathbb{N}\cup\{\infty,\top\},\oplus,\geq\rangle$

731:   composed of prison sentences. Sentences may be of $n$ years, life

732:   imprisonment (represented by $\infty$) or the death penalty

733:   (represented by $\top$). There is a rule that states that two life

734:   sentences lead automatically to a death sentence: in other words

735:   $(\infty\oplus\infty) = \top$.  Otherwise, sentences are cumulated in the obvious

736:   way: $\forall m, n \in\mathbb{N}, (m\oplus n = m+n)$; $\forall n\in \mathbb{N}, (\infty+n =

737:   \infty)$; $\forall \alpha\in E, (\top\oplus\alpha = \top)$. Although every pair $\beta,\alpha \in E,

738:   \alpha\leq \beta$ possesses a difference, this valuation structure is not

739:   fair since the set of differences of $\infty$ and $\infty$ is $\mathbb{N}$

740:   and hence no \emph{maximal} difference of $\infty$ and $\infty$ exists.

741:   However, $S$ can easily be rendered fair by replacing $\mathbb{N}$

742:   by $\{0,1,2,\ldots,150\}$, for example.

743: \end{example}

744: 

745: The following results show that all fair valuation structures are

746: composed of slices separated by absorbing values, each slice being

747: independent of the others.

748: 

749: \begin{lemma}\label{lemma2}

750:   Let $S=\langle E,\oplus,\succcurlyeq\rangle$ be a valuation structure. If $\alpha,\beta

751:   \in E$, $\alpha$ is an absorbing element and $\beta \preccurlyeq\alpha$ then

752:   $\alpha\oplus\beta = \alpha$. If $S$ is fair, then $\alpha \ominus \beta = \alpha$.

753: \end{lemma}

754: \begin{proof}

755:   Since $\beta\preccurlyeq \alpha$, it follows that $\alpha \preccurlyeq \alpha\oplus \beta

756:   \preccurlyeq \alpha\oplus\alpha =\alpha$, by monotonicity. Thus, $\alpha\oplus\beta = \alpha$.

757:   Furthermore, $\alpha\oplus\beta = \alpha$ shows that $\alpha$ is a difference of $\alpha$

758:   and $\beta$. It is the maximal difference since $\alpha\ominus \beta = (\alpha\ominus\beta)\oplus\bot

759:   \preccurlyeq (\alpha\ominus \beta)\oplus\beta = \alpha$, by monotonicity. Thus $\alpha\ominus\beta =\alpha$.

760: \end{proof}

761: 

762: \begin{lemma}\label{lemma3}

763:   Let $S=\langle E,\oplus,\succcurlyeq\rangle$ be a fair valuation structure. If

764:   $\alpha,\beta \in E$, $\alpha$ is an absorbing element and $\beta \succcurlyeq \alpha$

765:   then $\alpha\oplus\beta = \beta$ and $\beta \ominus \alpha = \beta$.

766: \end{lemma}

767: \begin{proof}

768:   Since $\alpha$ is absorbing, $\beta\oplus\alpha =(\beta\ominus \alpha)\oplus \alpha \oplus\alpha = (\beta\ominus \alpha)\oplus

769:   \alpha = \beta$. Furthermore, this shows that $\beta$ is a difference of $\beta$

770:   and $\alpha$.  It is the maximum difference since $\beta \ominus \alpha\preccurlyeq

771:   \beta$, by Lemma~\ref{lemma1}. Thus, $\beta \ominus \alpha = \beta$.

772: \end{proof}

773: %\noindent Notation: Let $S=\langle E,\oplus,\succcurlyeq\rangle$ be a valuation structure and

774: %$\beta\in E$, we denote by $\gla{\beta} = \max\{\alpha \in E~:~(\alpha\preccurlyeq \beta)

775: %\land(\alpha\textrm{~absorbing})\}$, the greatest absorbing lower bound of

776: %$\beta$ in $E$ and by $\lua{\beta} = \min\{\alpha \in E~:~(\alpha\succcurlyeq \beta) \land

777: %(\alpha\textrm{~absorbing})\}$ the lowest absorbing upper bound of $\beta$

778: %in $E$. $\gla{\beta}$ and $\lua{\beta}$ always exist since $\bot$ and $\top$ are

779: %absorbing and respectively the minimum and maximum elements of $E$.

780: 

781: \begin{theorem}[Slice Independence Theorem]\label{slicetheo}

782:   Let $S=\langle E,\oplus,\succcurlyeq\rangle$ be a fair valuation structure. Let

783:   $\beta,\gamma \in E, \beta\preccurlyeq \gamma$, and let $\alpha_0, \alpha_1\in E$ be absorbing

784:   valuations such that $\alpha_0\preccurlyeq \gamma\preccurlyeq\alpha_1$. Then

785:   $\alpha_0\preccurlyeq (\gamma \oplus\beta) \preccurlyeq \alpha_1$ and $\alpha_0

786:   \preccurlyeq (\gamma\ominus\beta) \preccurlyeq \alpha_1$.

787: \end{theorem}

788: \begin{proof}

789:   By monotonicity, $\beta\oplus\gamma \preccurlyeq \alpha_1 \oplus \gamma = \alpha_1$ by

790:   Lemma~\ref{lemma2}. By Lemma~\ref{lemma3}, $\gamma = \gamma \oplus \alpha_0 = (\gamma \ominus

791:   \beta) \oplus \beta \oplus \alpha_0 =(\gamma \ominus \beta) \oplus \alpha_0 \oplus \beta$.  Therefore, $((\gamma \ominus

792:   \beta) \oplus \alpha_0)$ is a difference of $\gamma$ and $\beta$. Since $\gamma\ominus\beta$ is a

793:   maximal difference, $\gamma\ominus\beta \succcurlyeq (\gamma\ominus\beta)\oplus\alpha_0\succcurlyeq

794:   \alpha_0$, by monotonicity. The remaining equalities follow from

795:   monotonicity.

796: \end{proof}

797:  

798: The following results will be useful for the proof of correctness of

799: the arc consistency enforcing algorithm given in the next section.

800: \begin{theorem}\label{maxabs}

801:   Let $S=\langle E,\oplus,\succcurlyeq\rangle$ be a fair valuation structure. For

802:   all $\alpha\in E$, $\alpha \ominus \alpha$ is the maximal absorbing valuation less than

803:   or equal to $\alpha$.

804: \end{theorem}

805: \begin{proof}

806:   Let $\beta=\alpha\ominus\alpha$. Now $\alpha\oplus(\beta\oplus\beta) = (\alpha\oplus\beta)\oplus\beta=\alpha\oplus\beta=\alpha$, which

807:   shows that $\beta\oplus\beta$ is a difference of $\alpha$ and $\alpha$.  By

808:   definition~\ref{fairdef}, $\beta$ is the maximal difference. Therefore,

809:   $\beta\succcurlyeq \beta\oplus\beta$. Since $\beta\preccurlyeq \beta\oplus\beta$ by

810:   monotonicity, we have $\beta=\beta\oplus\beta$ and hence  $\beta$ is

811:   absorbing. Maximality follows from Theorem~\ref{slicetheo}, since

812:   for all absorbing valuations $\alpha_0\preccurlyeq \alpha$, we have

813:   $\alpha_0\preccurlyeq (\alpha\ominus\alpha)$.

814: \end{proof}

815: 

816: \begin{lemma}\label{lemma4}

817:   Let $S=\langle E,\oplus,\succcurlyeq\rangle$ be a fair valuation structure. For

818:   all $\alpha,\beta\in E$, $((\alpha\oplus\beta)\ominus\alpha)\ominus\beta = (\alpha\oplus\beta)\ominus(\alpha\oplus\beta)$.

819: \end{lemma}

820: \begin{proof}

821:   Let $\gamma = (\alpha\oplus\beta)\ominus(\alpha\oplus\beta)$. By Theorem~\ref{maxabs}, $\gamma$ is

822:   absorbing. Now $\gamma\preccurlyeq (\alpha\oplus\beta)$, by Lemma~\ref{lemma1}. Let

823:   $\delta = ((\alpha\oplus\beta)\ominus\alpha)\ominus\beta$. Then the fact that $\delta \succcurlyeq \gamma$

824:   follows from two applications of the Slide Independence Theorem.

825:   But $\delta\oplus(\alpha\oplus\beta) = (\alpha\oplus\beta)$. Therefore $\delta$ is a difference of

826:   $(\alpha\oplus\beta)$ and $(\alpha\oplus\beta)$. $\gamma$ being the maximal difference, this

827:   shows that $\delta \preccurlyeq \gamma$ and $\delta=\gamma$.

828: \end{proof}

829: 

830: \begin{theorem}\label{theoabs}

831:   Let $S=\langle E,\oplus,\succcurlyeq\rangle$ be a fair valuation structure. For

832:   all $\alpha ,\beta \in E$, either $(\alpha\oplus\beta)\ominus \beta = \alpha$ or $(\alpha\oplus\beta)\ominus \beta =

833:   (\alpha\oplus\beta)\ominus(\alpha\oplus\beta)$ which is absorbing and strictly greater

834:   than $\alpha$.

835: \end{theorem}

836: \begin{proof}

837:   Let $\gamma = (\alpha\oplus\beta)\ominus(\alpha\oplus\beta)$. Now $(\alpha\oplus\beta)\ominus \beta =

838:   (((\alpha\oplus\beta)\ominus\beta)\ominus\alpha)\oplus\alpha = \gamma\oplus\alpha$, by Lemma~\ref{lemma4}. Since

839:   $\gamma$ is absorbing, $\gamma\oplus \alpha$ equals either $\alpha$ (if $\alpha \succcurlyeq

840:   \gamma$, Lemma~\ref{lemma3}) or $\gamma$ (if $\gamma \succcurlyeq \alpha$,

841:   Lemma~\ref{lemma2}). In this case, the fact that $\gamma$ is absorbing

842:   follows directly from Theorem~\ref{maxabs}.

843: \end{proof}

844: 

845: \begin{theorem}\label{finivs}

846:   Any VCSP $V=\langle X,D,C,S\rangle$ on a fair valuation structure $S=\langle

847:   E,\oplus,\succcurlyeq\rangle$ is equivalent to a VCSP on a valuation

848:   structure $S'$ with no more than $2ed^r+2$ absorbing valuations.

849: \end{theorem}

850: \begin{proof}

851:   For any $\beta \in E$ define

852:   \[\textit{Slice}(\beta) = \{\alpha\in E~:~\nexists \gamma \textrm{~absorbing in~}E \textrm{~s.t.~} (\alpha \prec \gamma \preccurlyeq \beta) \lor (\alpha \succ \gamma \succcurlyeq \beta) \}\]

853:   

854:   If $\beta$ is absorbing then $\textit{Slice}(\beta) = \{\beta\}$, otherwise

855:   $\textit{Slice}(\beta)$ is the set of valuations $\alpha$ for which there

856:   is no intermediate absorbing valuation $\gamma$ lying between $\alpha$ and

857:   $\beta$. Each $\textit{Slice}(\beta)$ contains at most two absorbing

858:   valuations, namely the maximum absorbing valuation less than or

859:   equal to $\beta$ (which is in fact $\beta \ominus \beta$, by Theorem~\ref{maxabs})

860:   and the minimum absorbing valuation greater than or equal to $\beta$

861:   (which may or may not exist).

862: 

863:   Let $E_0$ be the set of valuations taken on by the cost functions

864:   $c_P\in C$ in the VCSP and let

865: 

866:   \[E' = \{\bot,\top\} \bigcup_{\beta \in E_0} \textit{Slice}(\beta) \]

867:   

868:   Clearly $E'$ contains at most $2ed^r+2$ absorbing valuations. It

869:   is sufficient to show that $E'$ is closed under $\oplus$ and $\ominus$.

870:   

871:   Consider $\alpha,\beta \in E'$ such that $\alpha \preccurlyeq \beta$ and $\beta \in

872:   \textit{Slice}(\beta_o)$ for some $\beta_0 \in E_0$.

873:   

874:   Suppose that $(\alpha\oplus\beta) \not\in E'$. This implies that $\exists \gamma$

875:   absorbing in $E$ such that $\alpha\oplus\beta\succ \gamma\succcurlyeq \beta_0$. But,

876:   since $\beta\preccurlyeq\gamma$ by the definition of

877:   $\textit{Slice}(\beta_0)$, this contradicts Theorem~\ref{slicetheo}.

878:   Similarly, $\beta \ominus \alpha \not\in E'$ implies that $\exists\gamma$ absorbing in $E$

879:   such that $\beta\ominus\alpha\prec \gamma\preccurlyeq \beta_0$. But, since $\beta\succcurlyeq

880:   \gamma$ by the definition of $\textit{Slice}(\beta_0)$, this again

881:   contradicts Theorem~\ref{slicetheo}.

882: \end{proof}

883: 

884: By theorem~\ref{finivs}, we can now assume, without loss of

885: generality, that the number of absorbing valuations in the valuation

886: structure is finite.

887: 

888: \subsection{A Definition of Soft Arc Consistency}

889: 

890: Before giving an arc consistency enforcing algorithm which is valid

891: over any fair valuation structure, we require a formal definition of

892: arc consistency for fair VCSPs. We first consider the usual restriction

893: to binary VCSPs.

894: 

895: \begin{definition}\label{sac}

896:   A fair binary VCSP is \emph{arc consistent} if for all $i,j \in X$

897:   such that $c_{ij} \in C^+$, for all $a \in d_i$ we have:

898: \begin{enumerate}

899: \item\label{conda} $\forall b\in d_j, c_{ij}(a,b) = \bigl(c_i(a)\oplus c_{ij}(a,b)\oplus c_j(b)\bigr) \ominus \bigl(c_i(a)\oplus c_j(b)\bigr)$.

900: \item\label{condb} $c_i(a) = \min_{b\in d_j} (c_i(a) \oplus c_{ij}(a,b))$

901: \end{enumerate}

902: \end{definition}

903: 

904: Condition~\ref{conda} states that $c_{ij}(a,b)$ has been increased to

905: the maximal element in $E$ which does not increase the valuation

906: $\bigl(c_i(a)\oplus c_{ij}(a,b)\oplus c_j(b)\bigr)$ of $(a,b)$ on $\{i,j\}$.

907: If $\oplus$ is strictly monotonic or idempotent, then this is equivalent

908: to saying that absorbing valuations have been propagated from $c_i(a)$

909: to $c_{ij}(a,b)$.  Condition~\ref{condb} says that we

910: have propagated as much weight as possible from the constraint

911: $c_{ij}$ onto $c_i$. 

912: 

913: To gain a better understanding of condition~\ref{conda} of

914: Definition~\ref{sac} in the most general case, consider a simple

915: valuation structure in which penalties lies in the range

916: $\{0,1,2,3,4,5\}$ and $\forall \alpha,\beta\in E, (\alpha\oplus\beta = \min(5,\alpha + \beta))$. $5$

917: is absorbing and verifies $5\ominus\alpha = 5$ for all $\alpha\preccurlyeq5$.

918: Figure~\ref{figconda}(a) shows a 2-variable VCSP over this valuation

919: structure.  Figure~\ref{figconda}(b) shows the result of enforcing

920: condition~\ref{conda} of Definition~\ref{sac}: $c_{12}(a,a)$ and

921: $c_{12}(b,a)$ can both be increased to $5$ without changing the

922: valuations of the solutions $(a,a)$ and $(b,a)$.

923: Figure~\ref{figconda}(c) shows the result of then enforcing

924: condition~\ref{condb}: penalties are projected down from constraints

925: to domains, as we have seen in the example of Figure~\ref{maxcsp1}.

926: 

927: \begin{figure}

928: \begin{center}

929: \includegraphics[width=0.7\textwidth]{figconda}

930: \caption{(a) An example of a VCSP and how conditions~\ref{conda} and~\ref{condb} are enforced\label{figconda}}.

931: \end{center}

932: \end{figure}

933: 

934: Definition~\ref{sac} can be generalised to non binary VCSP. We call

935: this generalised arc consistency, to be consistent with the

936: terminology employed in the CSP literature~\cite{Mohr88}.

937: 

938: \begin{definition}\label{defgac}

939:   A fair VCSP is \emph{generalised arc consistent} if for all $c_P\in

940:   C^+$,  we have:

941: \begin{enumerate}

942: \item\label{gconda} $\forall t \in \ell(P), c_P(t) = (c_P(t) \oplus \beta) \ominus \beta$, where $\beta = \mathop\bigoplus_{j\in P} c_j(t_{\proj \{j\}})$.

943: \item\label{gcondb} $\forall i\in P, \forall a \in d_i, c_i(a) = \min_{t\in \ell(P-\{i\})} (c_i(a) \oplus c_P(t,a))$

944: \end{enumerate}

945: \end{definition}

946: Having given the necessary definitions, we can now define a

947: generalised arc consistency enforcing algorithm. 

948: 

949: \subsection{Enforcing generalised arc consistency in fair VCSPs}

950: 

951: Arc consistency is established by repeated calls to two subroutines

952: denoted by \ProjAC\ and \ExtAC\ (see Algorithms~\ref{algoproj}

953: and~\ref{algoext} respectively), called \emph{arc consistency

954:   operations}. The data structure $Q$ is a queue containing elements

955: $(t,P,\alpha)$, where $\alpha = c_P(t)$. The subroutine \ProjAC{$c_P,i,a$} is

956: a simple modification of the basic equivalence-preserving

957: transformation \Proj\ that memorises VCSP modifications in the queue

958: for further propagation. This simple modification obviously does not

959: alter the fact that it is an equivalence-preserving transformation.

960: 

961: \begin{algorithm}[htbp]

962:   \Procedure{\ProjAC{$c_P,i,a$}}{

963:     $\beta  \gets \min_{t \in \ell(P-\{i\})}(c_P(t,a))$\;

964:     \If{($c_i(a)\oplus \beta \succ c_i(a)$)}{

965:       $c_i(a) \gets c_i(a) \oplus \beta$\;

966:       Add $(a,\{i\},c_i(a))$ to $Q$\;

967:       \ForEach{($t\in \ell(P-\{i\})$)}

968:       {$c_P(t,a) \gets c_P(t,a) \ominus \beta$\;}

969:       }

970:     }

971:   \caption{Projection for Soft AC enforcing\label{algoproj}}

972: \end{algorithm}

973: 

974: \begin{lemma}\label{lemmaproj}

975:   For any fair VCSP $V=\langle X,D,C,S\rangle$, if $c_P\in C, i\in P, a\in d_i$ then

976:   the result of applying \ProjAC{$c_P,i,a$} to a fair VCSP $V$ is an

977:   equivalent VCSP in which

978: \[c_i(a) = \min_{t \in \ell(P-\{i\})} (c_i(a) \oplus c_P(t,a))\]

979: \end{lemma}

980: \begin{proof}

981:   This property follows from the fact that either $c_i(a)\oplus \beta$ is not

982:   strictly greater than $c_i(a)$ and in this case $c_i(a) = c_i(a)\oplus

983:   \min_{t \in \ell(P-\{i\})}(c_P(t,a)) = \min_{t \in \ell(P-\{i\})} (c_i(a)

984:   \oplus c_P(t,a))$ or else let $t_{\min}$ be the tuple such that

985:   $c_P(t_{\min},a) = \min_{t \in \ell(P-\{i\})}(c_P(t,a)) = \beta$ before

986:   execution of \ProjAC. Let $\delta$ be the original value of $c_i(a)$.

987:   After execution, $c_i(a) = \delta \oplus \beta = c_i(a) \oplus c_P(t_{\min},a)$.

988:   Therefore we have $c_i(a) \succcurlyeq \min_{t \in \ell(P-\{i\})}

989:   (c_i(a) \oplus c_P(t,a))$. The equality follows from monotonicity.

990: \end{proof}

991: 

992: The subroutine \ExtAC{$i,a,c_P$} is a modified version of the

993: equivalence-preserving transformation \Ext\ that propagates an

994: increase in the valuation $c_i(a)$ to all the valuations $c_P(t,a)$

995: for $t \in \ell(P-\{i\})$ when this can be done without any compensation

996: at the unary level.  It also memorises the new valuation $\gamma$ of each

997: tuple of $c_P$ in $Q$ for further propagation. In this case, $\gamma$ is

998: always an absorbing valuation (by Theorem~\ref{theoabs}).

999:  

1000: \begin{algorithm}[htbp]

1001:   \Procedure{\ExtAC{$i,a,c_P$}}{

1002:     \ForEach{($t \in \ell(P-\{i\})$)}{

1003:       $\beta \gets \mathop\bigoplus_{j\in P}c_j((t,a)_{\proj \{j\}})$\;

1004:       $\gamma \gets (c_P(t,a)\oplus\beta) \ominus \beta$\;

1005:       \If{($\gamma \succ c_P(t,a)$)}{

1006:         $c_P(t,a) \gets \gamma$\;

1007:         Add $((t,a),P,c_P(t,a))$ to $Q$\;

1008:         }

1009:       }

1010:     }

1011:   \caption{Extension for Soft AC enforcing\label{algoext}}

1012: \end{algorithm}

1013: 

1014: \begin{lemma}\label{lemmaext}

1015:   For any fair VCSP $V=\langle X,D,C,S\rangle$, if $c_P\in C, i\in P, a\in d_i$

1016:   then the result of applying \ExtAC{$i,a,c_P$} to a fair VCSP $V$ is

1017:   an equivalent VCSP in which

1018: \[\forall t\in \ell(P-\{i\}), (c_P(t,a) = (c_P(t,a)\oplus \beta)\ominus \beta),\textrm{where~}\beta =\mathop\bigoplus_{j\in P} c_j((t,a)_{\proj \{j\}})\]

1019: \end{lemma}

1020: \begin{proof}

1021:   To demonstrate equivalence, it is sufficient to prove that $\forall

1022:   t\in\ell(P-\{i\})$, $\beta \oplus c_P(t,a)$ is an invariant of

1023:   \ExtAC{$i,a,c_P$}, where $\beta =\mathop\bigoplus_{j\in P} c_j((t,a)_{\proj

1024:     \{j\}})$. But this is certainly the case because, if $c_P(t,a) =

1025:   \alpha$ before execution of \ExtAC{$i,a,c_P$}, then $\beta \oplus c_P(t,a) = \beta

1026:   \oplus ((\alpha \oplus \beta)\ominus \beta) = \beta \oplus \alpha$ after $c_P(t,a)$ is updated by

1027:   \ExtAC{$i,a,c_P$}.

1028:   

1029:   We know that $\alpha\preccurlyeq (\alpha\oplus \beta)\ominus\beta$, since

1030:   $\alpha$ is clearly a difference of $\alpha\oplus\beta$ and $\beta$. If $\alpha\prec

1031:   (\alpha\oplus\beta) \ominus \beta$, then $c_P(t,a)$ is assigned $(\alpha\oplus\beta) \ominus \beta$.

1032:   Hence, after $c_P(t,a)$ is updated by \ExtAC{$i,a,c_P$}, $\beta \oplus

1033:   c_P(t,a) = \beta \oplus \alpha$ and $(c_P(t,a) \oplus \beta) \ominus \beta = (\alpha \oplus \beta) \ominus \beta

1034:   = c_P(t,a)$.

1035: \end{proof}

1036: 

1037: We are now in a position to give an algorithm (Algorithm~\ref{refGAC})

1038: for generalised arc consistency in fair VCSPs.

1039: 

1040: \begin{algorithm}

1041:   \Procedure{\GAC{}}{

1042:     \{ \emph{Initialisation phase} \}\;

1043:     \ForEach{$i\in X$}

1044:     {

1045:       \ForEach{$a\in d_i$}

1046:       {

1047:         \ForEach{$c_P\in C^+$ s.t. $i\in P$}

1048:         {

1049:           \ExtAC{$i,a,c_P$}\;

1050:           \ProjAC{$c_P,i,a$}\;

1051:           }

1052:         }

1053:       }

1054:     \{ \emph{Propagation phase} \}\;

1055:     \While{$Q\neq\varnothing$}

1056:     {

1057:       Extract the first element $(t,P,\alpha)$ from $Q$\;

1058:       \If{$c_P(t) = \alpha$}

1059:       {

1060:         \eIf{$P$ is a singleton $\{i\}$}

1061:         {\lForEach{$c_P \in C^+$ s.t. $i\in P$}{\ExtAC{$i,t_{\proj \{i\}},c_P$}\;}}

1062:         {\lForEach{$i\in P$} {\ProjAC{$c_P,i,t_{\proj\{i\}}$}\;}}

1063:         }

1064:       }

1065:     }

1066:   \caption{Generalised Arc Consistency enforcing for fair VCSPs\label{refGAC}}

1067: \end{algorithm}

1068: 

1069: \begin{theorem}

1070: When \GAC\ terminates, the resulting VCSP is generalised arc consistent.

1071: \end{theorem}

1072: \begin{proof}

1073:   Consider $(c_P,i,a)$ such that $c_P \in C, i\in P, a\in d_i$. We know

1074:   that \ProjAC{$c_P,i,a$} is called at least once during \GAC, since it

1075:   is called in the initialisation phase. After the last call to

1076:   \ProjAC{$c_P,i,a$}

1077: \[c_i(a) = \min_{t \in \ell(P-\{i\})} (c_i(a)\oplus c_P(t,a))\]

1078: by Lemma~\ref{lemmaproj}. This can only later become invalid by an

1079: increase in some $c_P(t)$ by \ExtAC, which would necessarily be

1080: accompanied by the addition of $(t,P,c_P(t))$ to $Q$ and would hence

1081: entail another call of \ProjAC{$c_P,i,a$}. This contradiction

1082: demonstrates that Condition~\ref{gcondb} in Definition~\ref{defgac} of

1083: generalised arc consistency holds when \GAC\ terminates.

1084: 

1085: Consider $t\in\ell(P)$ where $c_P \in C^+$. We know that

1086: \ExtAC{$i,t_{\proj \{i\}},c_P$} is called during the initialisation

1087: phase for each $i \in P$. After the last such call of

1088: \ExtAC{$i,t_{\proj\{i\}},c_P$} for any $i\in P$,

1089: \[\forall t\in \ell(P), c_P(t) = (c_P(t) \oplus \beta) \ominus \beta, \textrm{where~}\beta =\mathop\bigoplus_{j\in P} c_j(t_{\proj \{j\}})\]

1090: by Lemma~\ref{lemmaext}. This could only later become invalid by an

1091: update of $c_P(t)$ or some $c_j(t_{\proj \{j\}})$ by

1092: \ProjAC{$c_N,j,t_{\proj \{j\}}$} for some $c_N\in C$ such that $j\in P\cap

1093: N$. But then a call of \ExtAC{$j,t_{\proj \{j\}},c_P$} would ensue. This

1094: contradiction shows that Condition~\ref{gcondb} of

1095: Definition~\ref{defgac} of generalised arc consistency also holds when

1096: \GAC\ terminates.

1097: \end{proof}

1098: 

1099: \begin{theorem}

1100: \GAC\ has polynomial time complexity.

1101: \end{theorem}

1102: \begin{proof}

1103:   Let $n_1$ be the number of elements $(t,P,\alpha)$ extracted from $Q$

1104:   during \GAC\ such that $P$ is a singleton, and let $n_2$ be the

1105:   number of elements $(t,P,\alpha)$ extracted from $Q$ during \GAC\ such

1106:   that $|P|\geq 2$. By Theorem~\ref{finivs}, we can assume that the

1107:   valuation structure contains at most $2ed^r+2$ absorbing valuations.

1108:   

1109:   Let $c_P\in C^+$ and $t\in \ell(P)$. By Theorem~\ref{theoabs}, \ExtAC\ 

1110:   can only increase $c_P(t)$ to an absorbing valuation strictly

1111:   greater than its previous valuation. \ProjAC\ can decrease $c_P(t)$

1112:   but, by the Slice Independence Theorem, only from a non-absorbing

1113:   valuation $\gamma$ to a valuation larger than or equal to $\delta = \gamma \ominus

1114:   \gamma$, the maximal absorbing valuation less than or equal to $\gamma$.

1115:   Thus the sequence of absorbing valuations taken on by $c_P(t)$

1116:   during \GAC\ are strictly increasing. Thus for each of the $2ed^r+2$

1117:   attainable absorbing valuations $\alpha$, $(t,P,\alpha)$ is added to $Q$ at

1118:   most once.  Thus $n_2 \leq ed^r(2ed^r+2)$.

1119:   

1120:   Now $n_1$ cannot exceed the number of calls of \ProjAC\ during \GAC\ 

1121:   since tuples $(t,P,\alpha)$ such that $P$ is a singleton are only added

1122:   to $Q$ by \ProjAC. The number of calls of \ProjAC\ is clearly bounded

1123:   above by $erd+n_2r$. Thus $n_1+n_2\leq erd+(r+1)n_2 = O(e^2d^{2r})$.

1124:   Thus the total number of iterations of the while loop in \GAC\ is a

1125:   polynomial function of $e$ and $d$. The results follows immediately.

1126: \end{proof}

1127: 

1128: \begin{definition} 

1129:   An arc consistent closure of a VCSP $V$ is a VCSP which is arc

1130:   consistent and which can be obtained from $V$ by a finite sequence

1131:   of applications of arc consistency operations \ExtAC\ and \ProjAC.

1132: \end{definition}

1133: 

1134: Note that confluence of arc consistency enforcing is lost and

1135: therefore the arc consistent closure of a problem is not necessarily

1136: unique as it is in classical CSPs. Figure~\ref{uniq}(a) shows a

1137: 2-variable VCSP on the valuation structure $\langle\setN\cup\{\infty\}, +,\geq\rangle$.

1138: Each edge has a weight of $1$.  Figures~\ref{uniq}(b)

1139: and~\ref{uniq}(c) show two different arc consistency closures of this

1140: VCSP.

1141: 

1142: 

1143: \begin{figure}[bthp]

1144: \begin{center}

1145:   \includegraphics[scale=0.5]{uniq}

1146: \end{center}

1147: \caption{A \textsc{Max-CSP} and two different equivalent arc consistent

1148:   closures\label{uniq}}

1149: \end{figure}

1150: 

1151: \subsection{Maximum arc consistency}

1152: 

1153: One of the practical use of arc consistency in VCSP is the computation

1154: of lower bounds on the valuation of an optimal solution. Obviously,

1155: given a VCSP $V$, the following valuation $f_{\min}(V)$ is always a

1156: lower bound on the cost of an optimal solution:

1157: \[  f_{\min}(V)= \mathop\bigoplus_{c_P \in C}[\min_{t\in \ell(P)} c_P(t)]\]

1158: From this point of view, the closure in ~\ref{uniq}(b) is preferable

1159: to the closure in Figure~\ref{uniq}(c) since it makes explicit the

1160: fact that $1$ is a lower bound on the valuation of all solutions.

1161: 

1162: \begin{definition}

1163:   A VCSP $V$ is said to be \emph{maximally arc consistent} if it is

1164:   arc consistent and if the associated lower bound $f_{\min}(V)$ is

1165:   maximum over all arc consistent closures of $V$.

1166: \end{definition}

1167: 

1168: The problem of enforcing maximal arc consistency is certainly

1169: practically important and some closely related problems are known to

1170: be \textsf{NP}-hard~\cite{Schiex98a,Schiex95a}. Consider the following

1171: problem:

1172: 

1173: \begin{problem}[Max-AC]

1174:   Given a VCSP $V=(X,D,C,S)$ and a valuation $\kappa\in S$, does there

1175:   exist an arc consistency closure $V'$ of $V$ such that $f_{\min} (V')

1176:   \succcurlyeq \kappa$ ?

1177: \end{problem}

1178: 

1179: \begin{theorem}

1180:   The decision problem \textsc{Max-AC} is \textsf{NP}-complete.

1181: \end{theorem}

1182: \begin{proof}

1183:   \textsc{Max-AC} is clearly in \textsf{NP}. It is therefore

1184:   sufficient to give a polynomial reduction from \textsc{3-Sat} to

1185:   \textsc{Max-AC}.

1186:   

1187:   Let $I_{\textsc{\scriptsize 3-Sat}}$ be an instance of \textsc{3-Sat},

1188:   consisting of $n$ variables and $d$ clauses, each clause being the

1189:   disjunction of exactly $3$ literals.

1190:   

1191:   We assume that each boolean variable $u$ and its negation $\lnot u$

1192:   occur exactly the same number of times in $I_{\textsc{\scriptsize 3-Sat}}$. Note

1193:   that if this is not initially the case, it can easily be achieved by

1194:   adding the required number of tautological clauses of the form $(u

1195:   \lor u \lor \lnot u)$ or $(u \lor \lnot u \lor \lnot u)$. We will now construct an instance

1196:   $V$ of \textsc{Max-CSP} on $4d$ variables such that $V$ has an arc

1197:   consistency closure $V'$ with $f_{\min}(V') \succcurlyeq

1198:   \frac{5d}{2}$ iff $I_{\textsc{\scriptsize 3-Sat}}$ is satisfiable. Note that $d$

1199:   is necessarily even by our assumption that each variable $u$ and its

1200:   negation occur the same number of times.

1201:   

1202: \begin{figure}

1203:   \begin{center}

1204:     \includegraphics[width=0.9\textwidth]{gadget-var}

1205:     \caption{The gadget $G(u)$ representing the $m_u$ occurrences of $u$ and the $m_u$ occurrences of $\lnot u$ in $I_{\textsc{\scriptsize 3-Sat}}$\label{gadget-var}}

1206:   \end{center}

1207: \end{figure}

1208: 

1209: Suppose that the boolean variable $u$ (and its negation $\lnot u$) occur

1210: in exactly $m(u)$ clauses in $I_{\textsc{\scriptsize 3-Sat}}$. Then we add the

1211: gadget $G(u)$ shown in Figure~\ref{gadget-var} to $V$, containing the

1212: $2m(u)$ variables $i(u,r)$, $i(\lnot u,r)$ ($r=1,\ldots,m(u)$), each with

1213: domain-size $3$. Each edge joining value $a$ at variable $i$ and value

1214: $b$ at variable $j$ represents a penalty of $1$, \textit{i.e.}

1215: $c_{ij}(a,b) =1$. For each clause $c$ in $I_{\textsc{\scriptsize 3-Sat}}$, we add

1216: the gadget $H(c)$ shown in Figure~\ref{gadget-cl}, involving a new

1217: variable $i(c)$ connected to three existing variables. In the example

1218: shown, $c\equiv (u\lor v\lor \lnot w)$ where $c$ contains the $p^{th}$ occurrence

1219: of the boolean variable $u$ in $I_{\textsc{\scriptsize 3-Sat}}$, the $q^{th}$

1220: occurrence of $v$ and the $r^{th}$ occurrence of $\lnot w$. In the gadget

1221: $H(c)$ shown in Figure~\ref{gadget-cl}, $i(c)$ is connected to $i(\lnot

1222: u,p)$, $i(\lnot v,q)$ and $i(w,r)$.

1223: 

1224: \begin{figure}[bhtp]

1225:   \begin{center}

1226:     \includegraphics[width=0.4\textwidth]{gadget-cl}

1227:     \caption{The gadget $H(c)$ for  $c\equiv (u\lor v\lor \lnot w)$.\label{gadget-cl}}

1228:   \end{center}

1229: \end{figure}

1230: 

1231: Let $M_i = \max_{a\in d_i}(c'_i(a))$, where $c'$ represents the

1232: constraint functions in $V'$. Each variable $i(c)$ can contribute at

1233: most $1$ to $f_{\min}(V')$, \textit{i.e.} $M_{i(c)} \leq 1$.  By

1234: construction of $G(u)$, if any variable (\textit{e.g.} $i(u,r)$)

1235: contributes $1$ to $f_{\min}(V')$, then its adjacent variables ($i(\lnot

1236: u,r-1)$ and $i(\lnot u ,r)$) cannot contribute to $f_{\min}(V')$. This

1237: means that the maximum value of $f_{\min}(V')$ is $d+\frac{3d}{2} =

1238: \frac{5d}{2}$, since the total number of variables in the gadgets

1239: $G(u)$ is $3d$. Indeed, $f_{\min}(V') = \frac{5d}{2}$ iff for all

1240: clauses $c$, $M_{i(c)} = 1$ and for all boolean variables $u$,

1241: 

1242:     \begin{eqnarray}\label{eqn-gadget}

1243:       (\forall r\in\{1,\ldots,m(u)\}, M_{i(u,r)} =1 \land M_{i(\lnot u,r)} = 0) & \lor\nonumber\\

1244:       (\forall r\in\{1,\ldots,m(u)\}, M_{i(\lnot u,r)} =1 \land M_{i(u,r)} = 0)

1245:     \end{eqnarray}

1246: 

1247: Consider the clause $c\equiv (u\lor v\lor \lnot w)$, whose gadget $H(c)$ is shown in Figure~\ref{gadget-cl}. By construction of $H(c)$, $M_{i(c)}=1$ implies that

1248: 

1249: \[(M_{i(\lnot u,p)} = 0) \lor (M_{i(\lnot v,q)} = 0) \lor (M_{i(w,r)} = 0) \]

1250: 

1251: which, in turn, implies from~(\ref{eqn-gadget}), that

1252: 

1253: \begin{equation}\label{eqn-gadget2}

1254:   (M_{i(u,1)} = 1) \lor (M_{i(v,1)} = 1) \lor (M_{i(w,1)} = 0) 

1255: \end{equation}

1256: 

1257: Suppose that $V$ has an arc consistency closure $V'$ with

1258: $f_{\min}(V') \succcurlyeq \frac{5d}{2}$. For each variable $u$ in

1259: $I_{\textsc{\scriptsize 3-Sat}}$, set $u = \textit{true}$ iff $M_{i(u,1)} = 1$.

1260: For each clause $c$, for example $c\equiv (u\lor v\lor \lnot w)$, we know

1261: from~(\ref{eqn-gadget2}) that either $u =\textit{true}$,

1262: $v=\textit{true}$ or $w=\textit{false}$. Hence $I_{\textsc{\scriptsize 3-Sat}}$ is

1263: satisfied.

1264: 

1265: Suppose that $\omega$ is a model of $I_{\textsc{\scriptsize 3-Sat}}$. In each

1266: gadget $G(u)$ in $V$ and for each $r\in\{1,\ldots,m(u)\}$, if

1267: $u=\textit{true}$ in $\omega$ then project penalties onto $i(u,r)$

1268: from its constraints with the adjacent variables $i(\lnot u, r-1)$ and

1269: $i(\lnot u,r)$; if $u=\textit{false}$ in $\omega$, then project penalties

1270: onto $i(\lnot u,r)$ from its constraints with the adjacent variables

1271: $i(u,r)$ and $i(u,r+1)$. Consider a gadget $H(c)$ in $V$, such as the

1272: gadget illustrated in Figure~\ref{gadget-cl} for $c\equiv(u\lor v\lor\lnot w)$. If

1273: $u=\textit{true}$ in $\omega$, then project penalties onto $i(c)$ from

1274: its constraint with $i(\lnot u,p)$; if $u=\textit{false}$ in $\omega$,

1275: then project penalties onto $i(\lnot u,p)$ from its constraint with

1276: $i(c)$. Similarly, if $\lnot w =\textit{true}$ in $\omega$, then project

1277: penalties onto $i(c)$ from its constraint with $i(w,r)$; if $\lnot w

1278: =\textit{false}$ in $\omega$, then project penalties onto $i(w,r)$

1279: from its constraint with $i(c)$.  Let $V'$ be the resulting arc

1280: consistent VCSP. Since each clause is satisfied by $\omega$,

1281: $M_{i(c)}=1$. Furthermore, if $u=\textit{true}$ in $\omega$, then

1282: $M_{i(u,r)} =1$ for $r\in\{1,\ldots,m(u)\}$ and if $u=\textit{false}$ in

1283: $\omega$, then $M_{i(\lnot u,r)} =1 $ for $r\in\{1,\ldots,m(u)\}$. Thus

1284: $f_{\min}(V') = d+\frac{3d}{2} = \frac{5d}{2}$.

1285: \end{proof}

1286: 

1287: \subsection{The case of  strictly monotonic VCSPs}

1288: 

1289: For strictly monotonic VCSPs, the previous algorithm can be improved

1290: using an alternative equivalent definition of arc consistency based on

1291: the notion of the underlying CSP.

1292: 

1293: \begin{definition}

1294:   The \emph{underlying CSP} of a VCSP $V$ has the same variables and

1295:   domains as $V$ together with, for each constraint $c_P\in C$, a crisp

1296:   constraint $c'_P$ satisfying $\forall t\in \ell(P) (t\in c'_P \Leftrightarrow c_P(t) \prec \top)$

1297:   (i.e., $t$ is not a totally forbidden labelling).

1298: \end{definition} 

1299: 

1300: In strictly monotonic VCSPs, the only absorbing elements are $\top$ and

1301: $\bot$. This allows us to give an equivalent but simpler definition of

1302: generalised arc consistency:

1303: \begin{theorem}\label{defsmgac} 

1304:   If $\oplus$ is a strictly monotonic operator, then a VCSP is generalised

1305:   arc consistent iff:

1306:   \begin{enumerate}

1307:   \item\label{smconda} its underlying CSP is generalised arc consistent

1308:   \item\label{smcondb} $\forall c_P \in C^+, \forall i\in P, \forall a\in d_i$, if $c_i(a) \prec \top$ then $\exists t\in\ell(P-\{i\}) (c_P(t,a) = \bot)$.

1309:   \end{enumerate}

1310: \end{theorem}

1311: \begin{proof}

1312:   ($\Rightarrow$) Suppose that a strictly monotonic VCSP is generalised arc

1313:   consistent but its underlying CSP is not. Then $\exists c_P \in C, i\in P,

1314:   a \in d_i$ such that $(c_i(a) \prec \top) \land (\forall t\in \ell(P-\{i\}) (c_P(t,a)

1315:   = \top \lor \exists j\in P-\{i\} (c_j(t_{\proj\{j\}}) = \top)))$. But, by

1316:   Condition~\ref{gconda} of Definition~\ref{defgac},

1317:   $(c_j(t_{\proj\{j\}}) = \top)$ implies $c_P(t,a) = \top$. Hence, $c_i(a)

1318:   = \top$, by Condition~\ref{gcondb} of Definition~\ref{defgac}, which

1319:   is a contradiction. Suppose on the other hand that the VCSP is

1320:   generalised arc consistent but Condition~\ref{smcondb} of

1321:   Theorem~\ref{defsmgac} is not satisfied. Then $\exists c_P \in C^+, i\in P,

1322:   a\in d_i$ such that $c_i(a) \prec \top$ and $\forall t\in\ell(P-\{i\}), (c_P(t,a)

1323:   \succ \bot)$. But by Condition~\ref{gcondb} of Definition~\ref{defgac},

1324:   for some $t\in \ell(P-\{i\}), c_i(a) = c_i(a) \oplus c_P(t,a) \succ c_i(a)$ by

1325:   strict monotonicity, which is impossible.\smallskip

1326:   

1327:   ($\Leftarrow$) Suppose that a strictly monotonic VCSP satisfies

1328:   Condition~\ref{smconda} and~\ref{smcondb} of Theorem~\ref{defsmgac}.

1329:   Condition~\ref{smcondb} clearly implies Condition~\ref{gcondb} of

1330:   Definition~\ref{defgac}. Suppose that Condition~\ref{gconda} of

1331:   Definition~\ref{defgac} is not satisfied. Then $\exists c_P\in C, t\in

1332:   \ell(P)$ such that $c_P(t) \neq \delta = (c_p(t) \oplus \beta) \ominus \beta$ where $\beta =

1333:   \mathop\bigoplus_{j\in P} c_j(t_{\proj \{j\}})$. By

1334:   Theorem~\ref{theoabs}, $\delta$ is absorbing, which is only possible

1335:   if $c_j(t_{\proj \{j\}})=\top$ for some $j\in P$. But the generalised

1336:   arc consistency of the underlying CSP implies $c_P(t) = \top$, which

1337:   provides the necessary contradiction.

1338: \end{proof}

1339: 

1340: A possible way to enforce soft arc consistency on a strictly monotonic

1341: VCSP is therefore to first enforce classical arc consistency on the

1342: underlying CSP, assign a valuation of $\top$ to all deleted values in

1343: the original VCSP and then enforce the second property by applying

1344: \Proj{$c_P,i,a$} once for all $c_P \in C^+$, all $i \in P$ and all

1345: $a\in d_i$ (see Algorithm~\ref{algoSACSM}).

1346: 

1347: \begin{algorithm}

1348:   Enforce arc consistency in the underlying CSP\;

1349:   \ForEach{$c_P\in C^+$}  {

1350:     \ForEach{$i \in P$}{

1351:       \lForEach{$a \in d_i$}{\Proj{$c_P,i,a$}\;}}

1352:     }

1353:   \caption{Enforcing soft arc consistency on strictly monotonic VCSPs\label{algoSACSM}}

1354: \end{algorithm}

1355: 

1356: The only additional result needed to prove that this algorithm works

1357: is the following one:

1358: \begin{theorem}

1359:   Let $V$ be a strictly monotonic VCSP whose underlying CSP is arc

1360:   consistent. Then, $\forall c_P\in C^+, \forall i\in P, \forall a\in d_i$, the

1361:   application of the equivalence-preserving transformation

1362:   \Proj{$c_P,i,a$} yields a VCSP whose underlying CSP is unchanged

1363:   (and therefore arc consistent).

1364: \end{theorem}

1365: 

1366: \begin{proof}

1367:   \Proj{$c_P,i,a$} cannot increase a valuation $c_i(a) \neq \top$ to $\top$

1368:   since arc consistency on the underlying CSP would have deleted

1369:   $(i,a)$ in this case and therefore we would have set $c_i(a) =\top$.

1370:   It cannot decrease the valuation of any tuple such that $c_P(t,a) =

1371:   \top$ since $\forall \beta\in E, \top \ominus \beta = \top$.

1372: \end{proof}

1373: 

1374: \subsubsection{Improving space complexity}

1375: 

1376: If an optimal $O(ed^r)$ arc consistency enforcing algorithm such as

1377: generalised AC7~\cite{Bessiere97} is used to enforce arc consistency

1378: on the underlying CSP of a fair VCSP, Algorithm~\ref{algoSACSM}

1379: establishes arc consistency in $O(ed^r)$ too.  However, the space

1380: complexity of this algorithm is dominated by the space complexity of

1381: the modified constraints which requires $O(ed^r)$ valuations. This is

1382: extremely expensive, especially for constraints defined using a cost

1383: function.  This space requirement can be reduced using a simple data

1384: structure for representing modifications of costs induced by basic

1385: equivalence-preserving transformations such as \Proj\ and \Ext.

1386: 

1387: Let us denote by $c^{or}_P \in C$ the original definition of a

1388: constraint in a fair VCSP by any possible way: explicitly by a table

1389: of valuations or implicitly by a cost function from $\ell(P) \to E$. For

1390: each constraint $c_P$, for each variable $i\in P$, we use $2$ tables of

1391: $d$ valuations noted $\Delta^+_{Pi}$ and $\Delta^-_{Pi}$, initialised to $\bot$.

1392: Let $a\in d_i$:

1393: \begin{itemize}

1394: \item $\Delta^-_{Pi}[a]$ contains the combination of all the valuations

1395:   that are projected from $c_P$ onto $(i,a)$;

1396: \item $\Delta^+_{Pi}[a]$ contains the aggregation of all the valuations

1397:   that are extended from $(i,a)$ to $c_P$.

1398: \end{itemize}

1399: At any time, the valuation of a tuple $t\in\ell(P)$ in the modified

1400: constraint $c_P$ can simply be obtained by~:

1401: 

1402: \[ c_P(t) = c^{or}_P(t) \oplus (\mathop\bigoplus_{i\in P} \Delta^+_{Pi}[t_{\proj \{i\}}])  \ominus (\mathop\bigoplus_{i\in P} \Delta^-_{Pi}[t_{\proj \{i\}}])  \]

1403: 

1404: By definition of projection, $(\mathop\bigoplus_{i\in P} \Delta^-_{Pi}[t_{\proj

1405:   \{i\}}]) \preccurlyeq (c^{or}_P(t) \oplus (\mathop\bigoplus_{i\in P}

1406: \Delta^+_{Pi}[t_{\proj \{i\}}]))$ and the difference always exists in a

1407: fair VCSP. The space complexity is now reduced to $O(edr)$ instead of

1408: $O(ed^r)$ and our two basic equivalence-preserving transformations

1409: \Proj\ and \Ext\ become space tractable even for large arity

1410: constraints defined using cost functions.

1411: 

1412: \begin{algorithm}[htbp]

1413:   \Procedure{\Proj{$c_{ij},i,a$}}{

1414:     $\displaystyle\alpha  \gets \min_{t \in \ell(P-\{i\})}(c^{or}_P(t,a) \oplus (\mathop\bigoplus_{j\in P} \Delta^+_{Pj}[(t,a)_{\proj \{j\}}])  \ominus (\mathop\bigoplus_{j\in P} \Delta^-_{Pj}[(t,a)_{\proj \{j\}}]))$\;

1415:     $c_i(a) \gets c_i(a) \oplus \alpha$\;

1416:     $\Delta^-_{Pi}[a] \gets \Delta^-_{Pi}[a] \oplus \alpha$\;

1417:     }

1418:   \BlankLine

1419:   \BlankLine

1420:   \Procedure{\Ext{$i,a,c_P$}}{

1421:     $\alpha  \gets c_i(a)$\;

1422:     $\Delta^+_{Pi}[a] \gets \Delta^+_{Pi}[a] \oplus \alpha$\;

1423:     $c_i(a) \gets c_i(a) \ominus \alpha$\;

1424:     }

1425:   \caption{Projection and Extension for Soft AC enforcing\label{Proj2}}

1426: \end{algorithm}

1427: 

1428: Algorithm~\ref{Proj2} describes the procedures that implement these

1429: transformations with these data structures.  The time complexity of

1430: \Ext\ is reduced to $O(1)$ but since computing $c_P(t)$ requires

1431: $O(r)$ $\oplus$ operations, \Proj\ is $O(rd^{r-1})$ instead of

1432: $O(d^{r-1})$. This makes generalised arc consistency enforcing on

1433: strictly monotonic VCSPs $O(red^r)$ in time. For binary constraints, we

1434: recover the usual $O(ed^2)$ time and $O(ed)$ space complexities for

1435: arc consistency enforcing.

1436: 

1437: \section{Directional arc consistency}

1438: 

1439: In CSPs, directional arc consistency is a weak version of arc

1440: consistency. In VCSPs, the order imposed on variables by directional

1441: arc consistency (first defined in~\cite{CooperFCSP} for strictly

1442: monotonic operators) makes it possible to use the unlimited version of

1443: \Ext (instead of \ExtAC) together with a terminating algorithm. For

1444: this reason, soft directional arc consistency may be stronger than arc

1445: consistency.

1446: 

1447: \begin{figure}[bhtp]

1448:   \begin{center}

1449:     \includegraphics[width=0.85\textwidth]{extp2}

1450:   \end{center}

1451:   \caption{Enforcing directional arc consistency\label{extp2}}

1452: \end{figure}

1453: 

1454: Consider the Max-CSP in figure~\ref{extp2}(a).  It includes one binary

1455: constraint that forbids pair $((1,b)(2,b))$ and two unary constraints

1456: that forbid values $(1,a)$ and $(2,a)$. This VCSP is already arc

1457: consistent and the corresponding lower bound $f_{\min}(V)$ is equal to

1458: $0$.  However, we can apply \Ext\ on value $(1,a)$, getting the

1459: equivalent VCSP~\ref{extp2}(b) which is not arc consistent.  We can

1460: then apply \Proj\ on value $(2,b)$ and obtain the CSP~\ref{extp2}(c)

1461: with a corresponding lower bound $f_{\min}(V) = 1$.

1462: 

1463: This improved lower bound has been obtained because we have decided to

1464: pool all the unary valuations on one of the variables. This can be

1465: done successively on all variables using any given variable order. In

1466: the context of branch and bound (or other tree-based search), weights

1467: can for example be propagated towards those variables which occur

1468: earlier in the instantiation order.

1469: \begin{definition}

1470:   A binary VCSP is \emph{directional arc consistent} according to an

1471:   order $<$ on variables if $\forall c_{ij} \in C^+$ such that $i < j$, $\forall

1472:   a\in d_i$

1473:   \[c_i(a) = \min_{b\in d_j} (c_i(a) \oplus c_{ij}(a,b) \oplus c_j(b))\]

1474: \end{definition}

1475: 

1476: Provided that the VCSP is fair, directional arc consistency can be

1477: established in polynomial time by the procedure \DAC\ in

1478: Algorithm~\ref{algodac}, where \Proj\ and \Ext\ are as given in

1479: Algorithm~\ref{Proj2}.

1480: 

1481: \begin{algorithm}

1482:   \Procedure{\DAC{}}{

1483:     \For{($i \gets (n-1)$ downto $1$)}

1484:     {

1485:       \ForEach{($j \in X$ s.t. $j> i$ and $c_{ij}\in C$)} {

1486:         \{Start of propagation of $c_{ij}$\}\;

1487:         \lForEach{($b\in d_j$)} {\Ext{$j,b,c_{ij}$}\;}

1488:         \lForEach{($a\in d_i$)}{\Proj{$c_{ij},i,a$}\;}

1489:         \{End of propagation of $c_{ij}$\}\;

1490:         \lnl{assert}      \{Assert $A(i,j): \forall a\in d_i, c_i(a) =\min_{b\in d_j} (c_i(a)

1491:         \oplus c_{ij}(a,b)\oplus c_j(b))$\}\;

1492:         }

1493:       }

1494:     }

1495: \caption{Enforcing directional arc consistency on fair binary VCSPs\label{algodac}}

1496: \end{algorithm}

1497: 

1498: Since \DAC\ only applies equivalence-preserving transformations, it

1499: yields an equivalent VCSP. The following lemma is needed to prove that

1500: the VCSP obtained is directional arc consistent.

1501: 

1502: \begin{lemma}\label{lemmadac}

1503:   If $\oplus$ is fair, then

1504:   $$(\alpha = \alpha \oplus \beta \oplus \gamma) \land (\alpha' \succcurlyeq \alpha) \land (\gamma'\preccurlyeq \gamma) \Rightarrow (\alpha' = \alpha' \oplus \beta \oplus \gamma') $$

1505: \end{lemma}

1506: \begin{proof}

1507:   Suppose that $(\alpha = \alpha \oplus \beta \oplus \gamma) \land (\alpha' \succcurlyeq \alpha) \land

1508:     (\gamma'\preccurlyeq \gamma)$. Then $\alpha' \preccurlyeq \alpha'\oplus \beta \oplus \gamma'

1509:     \preccurlyeq \alpha'\oplus \beta \oplus \gamma \preccurlyeq (\alpha' \ominus \alpha) \oplus (\alpha \oplus \beta \oplus

1510:     \gamma) = (\alpha' \ominus \alpha) \oplus \alpha = \alpha'$. It follows that $\alpha' = \alpha' \oplus \beta

1511:     \oplus \gamma'$.

1512: \end{proof}

1513: 

1514: \begin{theorem}\label{theodac}

1515:   If the binary VCSP is fair, then directional arc consistency can be

1516:   established in $O(ed^2)$ time and $O(ed)$ space complexity.

1517: \end{theorem}

1518: \begin{proof}

1519:   The assertion $A(i,j)$ is clearly true during execution of \DAC\ at

1520:   line~1 of Algorithm~\ref{algodac} when constraint

1521:   $c_{ij}$ has just been propagated.

1522:   

1523:   It suffices to show that $A(i,j)$ cannot be invalidated by later

1524:   propagations of constraints $c_{i'j'}$ where $i'<j'$ and $(i'<i) \lor

1525:   (i' = i \land j\neq j')$. Such operations may increase $c_i(a)$ and may

1526:   decrease $c_j(b)$ but cannot modify $c_{ij}(a,b)$. From

1527:   Lemma~\ref{lemmadac}, $c_i(a) = c_i(a) \oplus c_{ij}(a,b) \oplus c_j(b)$

1528:   remains true and hence assertion $A(i,j)$ cannot be invalidated by

1529:   later propagations and \DAC\ yields a directional arc consistent

1530:   VCSP.

1531:   

1532:   As for time complexity, procedures \Ext\ and \Proj\ are called

1533:   $O(ed)$ times and are both $O(d)$ for binary VCSP. \DAC\ is

1534:   therefore $O(ed^2)$ in time. The $O(ed)$ space complexity can be

1535:   attained using the implementations of \Proj\ and \Ext\ presented in

1536:   Algorithm~\ref{Proj2}.

1537: \end{proof}

1538: 

1539: When restricted to strictly monotonic VCSP, Theorem~\ref{theodac} can

1540: be related to the result, proved in~\cite{CooperFCSP}, that

1541: \emph{full directional arc consistency}, a stronger version of

1542: directional arc consistency, can be established in $O(ed^2)$ time and

1543: space complexity. A VCSP is full directional arc consistent if and

1544: only if it is simultaneously arc consistent and directional arc

1545: consistent.

1546: 

1547: \begin{theorem}

1548:   Suppose that the constraint graph of a binary fair VCSP $V$ is a

1549:   tree $T$ and that $V$ is directional arc consistent according to some

1550:   topological ordering of the tree ($i$ is the father of $j$ in $T \Rightarrow

1551:   (i < j)$). Then, for all $a\in d_1$, $c_1(a)$ is the optimal

1552:   valuation over all solutions to $V$ in which variable $1$ is

1553:   assigned value $a \in d_1$.

1554: \end{theorem}

1555: 

1556: \begin{proof}

1557:   Let $S$ be the set of all the sons of variable $1$ in $T$ and for

1558:   each $j\in S$, let $T_j$ be the set of all variables in the subtree

1559:   rooted in $j$.  By induction, we assume that $\forall j \in S,\forall b\in d_j$,

1560:   the valuation $c_j(b)$ is the optimal valuation over all solutions

1561:   to the subproblem on $T_j$. Let $t_{bj}$ be one corresponding

1562:   optimal tuple over $T_j$.

1563:   

1564:   Let $a\in d_1$ and for each $j\in S$, let $b_{ja}$ be a value in $d_j$

1565:   that minimises $c_1(a)\oplus c_{1j}(a,b_{ja}) \oplus c_j(b_{ja})$. Since $\forall

1566:   i,j \in S, i\neq j$, $T_i \cap T_j = \varnothing$, we can build a tuple $t$ over

1567:   $X$ by concatenation of each $t_{b_{ja}j}$ for all $j \in S$ and by

1568:   assigning value $a$ to variable $1$.  The valuation of the tuple

1569:   $t$, is $\mathcal{V}_V(t) = c_1(a) \oplus \mathop\bigoplus_{j\in S}\bigl(c_j(b_{ja}) \oplus c_{1j}(a,b_{ja})\bigr) =c_i(a)$ since the VCSP is

1570:   directional arc consistent.

1571:   

1572:   Suppose there exists $t'$ such that $t_{\proj \{1\}} = a$ and

1573:   $\mathcal{V}_V(t') \prec \mathcal{V}_V(t)$. Since the problem is

1574:   tree-structured, the valuation of $t'$ can be written as 

1575:   \[\begin{array}{rclr}

1576:     \mathcal{V}_V(t')\!\!\! & =\!\!\! & c_1(a)  \oplus \mathop\bigoplus_{j\in S} \bigl(c_{1j}(a,t'_{\proj \{j\}}) \oplus \mathcal{V}_{V(T_j)}(t'_{\proj T_j})\bigr)\!\!\!\!\!\! & \textrm{(tree structure)}\\

1577:                               & \succcurlyeq\!\!\! & c_1(a) \oplus  \mathop\bigoplus_{j\in S}\bigl(c_{1j}(a,t'_{\proj \{j\}}) \oplus c_j(t'_{\proj \{j\}})\bigr) & \textrm{(induction, monotonicity)}\\

1578:                               & \succcurlyeq\!\!\! & c_1(a)  \oplus \mathop\bigoplus_{j\in S}\bigl(c_j(b_{ja}) \oplus c_{1j}(a,b_{ja})\bigr)& \textrm{(definition of $b_{ja}$)}\\

1579:                               & =\!\!\! & c_1(a) = \mathcal{V}_V(t)\\

1580:                             \end{array}\]

1581: which shows that no such $t'$ exists.

1582: \end{proof}

1583: 

1584: 

1585: %\section{Full directional arc consistency}

1586: 

1587: %As the VCSP in Figure~\ref{extp2}(c) shows, a VCSP which has been made

1588: %directional arc consistent is not necessarily arc consistent. In this

1589: %example, arc consistency can be recovered by projecting back the

1590: %constraint on the other side (Figure~\ref{extp2}(d)). This leads to the

1591: %following notion:

1592: %\begin{definition}

1593: %  A fair binary VCSP is \emph{fully directional arc consistent} for

1594: %  the order $>$ of the variables if it is directional arc consistent

1595: %  for $>$ and arc consistent.

1596: %\end{definition}

1597: 

1598: %\begin{definition}

1599: %  A full directional arc consistent closure of a fair binary VCSP $P$

1600: %  according to an order $>$ is a VCSP which is fully directional arc

1601: %  consistent and which can be obtained from $P$ by a finite sequence

1602: %  of applications of arc consistency operations \Ext\ and \Proj.

1603: %\end{definition}

1604: 

1605: %By definition, a fully directional AC closure is an AC closure.

1606: %Algorithm~\ref{FDAC} shows that fully directional arc consistent

1607: %closures always exist. We assume that the order $1<\cdots<n$ is used but

1608: %any other order can be used.

1609: 

1610: %\begin{algorithm}

1611: %  Enforce classical arc consistency on the underlying CSP\;

1612: %  \For{$j = n$ down to $2$}{

1613: %    \ForEach{$c_{ij} \in C$ s.t. $i < j$} {

1614: %      \lnl{proj1}    \lForEach{$b\in d_j$}{\Proj{$c_{ij},j,b$}\;}

1615: %      }

1616: %    \ForEach{$c_{ij} \in C$ s.t. $i < j$} {

1617: %      \lnl{ext}    \lForEach{$b\in d_j$}{\Ext{$j,b,c_{ij}$}\;}

1618: %      \lnl{proj2}    \lForEach{$a\in d_i$}{\Proj{$c_{ij},i,a$}\;}

1619: %      \lnl{proj3}    \lForEach{$b\in d_j$}{\Proj{$c_{ij},j,b$}\;}

1620: %      }

1621: %    }

1622: %\caption{Enforcing full directional arc consistency\label{FDAC}}

1623: %\end{algorithm}

1624: 

1625: 

1626: \section{Arc irreducibility}

1627: 

1628: We have seen that the cost of generalising arc consistency from crisp

1629: to valued constraint satisfaction problems is the loss of uniqueness

1630: of the arc consistent closure. As Figure~\ref{uniq} showed, two

1631: different arc consistent closures may also induce different lower

1632: bounds via $f_{\min}$.

1633: 

1634: If a VCSP $V$ is equivalent to another VCSP $V'$ which is better than

1635: $V$ (according to some formally defined criterion $\mathcal{C}$) then

1636: we say that $V$ is reducible for this criterion. Reducing a VCSP to an

1637: equivalent irreducible problem is, in general, an \textsf{NP}-hard

1638: problem. To see this, consider a CSP. If it is inconsistent

1639: (\emph{i.e.}, has no solution), then its best equivalent problem, for

1640: any reasonable criterion $\mathcal{C}$, is a CSP in which this fact is

1641: made explicit, for example, by having $\forall a\in d_i, c_i(a) = \top$ for

1642: some variable $i\in X$. Since testing consistency of a CSP is an

1643: \textsf{NP}-complete problem, we can deduce that testing global

1644: irreducibility is \textsf{NP}-hard.

1645: 

1646: Fortunately, irreducibility has local versions which are analogous to

1647: local consistency. For example, a binary VCSP $V$ is arc-irreducible

1648: if, for all pairs of variables $i,j\in X$, $V$ cannot be improved by

1649: replacing $c_i,c_j,c_{ij}$ by an equivalent set of constraints

1650: $c_i',c_j',c_{ij}'$. However, before giving a formal definition of

1651: arc-irreducibility, we have to consider which criteria we could use to

1652: compare equivalent VCSPs.

1653: 

1654: \begin{definition}

1655:   A \emph{problem evaluation function} $f$ is a function which, for

1656:   each VCSP $V$, assigns a value to $V$ in a totally ordered range.

1657:   When comparing two equivalent VCSPs, $V_1$ and $V_2$, $V_1$ is

1658:   considered as a better expression of the problem if $f(V_1) >

1659:   f(V_2)$.

1660: \end{definition}

1661: 

1662: \begin{example} The function $f_{\min}$ previously defined as

1663: \[f_{\min}(V) = \mathop\bigoplus_{c_P\in C} ( \min_{t\in \ell(P)i} (c_P(t))) \]

1664: is a problem evaluation function.

1665: \end{example}

1666: 

1667: The main result presented in this section concerns $f_{\min}$, but the

1668: definitions are valid for any problem evaluation function $f$.

1669: 

1670: \begin{definition}

1671:   A binary VCSP $V$ is \emph{arc-irreducible} with respect to the

1672:   problem evaluation function $f$ (or $(2,f)$-irreducible), if $\forall J

1673:   \subseteq X, |J| =2$, for all VCSP $V'$ derived from $V$ by an

1674:   equivalence-preserving transformation on $J$, $f(V) \geq f(V')$.

1675: \end{definition}

1676: Note that for certain choices of the problem evaluation function $f$,

1677: an arc-irreducible VCSP is not necessarily arc-consistent. For

1678: example, if $\forall i,j\in X, \exists a\in d_i, b\in d_j$ such that $c_i(a) =

1679: c_j(b) = c_{ij}(a,b) = \bot$, then the VCSP is

1680: $(2,f_{\min})$-irreducible but it is not necessarily arc-consistent.

1681: Conversely, the VCSP in Figure~\ref{uniq}(c) is arc consistent, but

1682: not $(2,f_{\min})$-irreducible.  The following theorem shows that

1683: there is an important relationship between directional arc consistency

1684: and $(2,f_{\min})$-irreducibility.

1685: 

1686: \begin{theorem}\label{dac2ir}

1687:   A fair binary VCSP $V=\langle X,D,C,S\rangle$ which is directional arc consistent is

1688:   $(2,f_{\min})$ -irreducible.

1689: \end{theorem}

1690: \begin{proof}

1691:   Suppose that $V$ is directional arc consistent and let $i,j\in X$ be

1692:   such that $i<j$. Then, by definition, $\forall a\in d_i,  \exists b\in d_j$ such that

1693:   \[c_i(a) = \min_{b\in d_j} (c_i(a) \oplus c_{ij}(a,b) \oplus c_j(b))\]

1694:   Thus the minimum valuation of a solution to the subproblem of $V$ on

1695:   $\{i,j\}$ is

1696:   \[\min_{a\in d_i}(c_i(a))\]

1697:   Suppose that this minimum is attained for $a=u\in d_i$ and that $v\in

1698:   d_j$ is such that

1699:   \begin{equation}\label{eqdac}

1700:     c_i(u) = (c_i(u) \oplus c_{ij}(u,v) \oplus c_j(v))

1701:   \end{equation}

1702:   Now consider a VCSP $V'$ obtained by an equivalence-preserving

1703:   transformation of $V$ on $\{i,j\}$ which replaces $c_i, c_j,c_{ij}$

1704:   by $c'_i, c'_j,c'_{ij}$. Then

1705: \[f_{\min}(V') = \Bigl(f_{\min}(V) \ominus \bigl(\mathop\bigoplus_{P\subseteq \{i,j\}} [\min_{t\in \ell(P)} c_P(t)]\bigr)\Bigr) \oplus \bigl(\mathop\bigoplus_{P\subseteq \{i,j\}} [\min_{t\in \ell(P)} c'_P(t)] \bigr)\]

1706: since $V$ and $V'$ only differ on $\{i,j\}$. Thus

1707: \[\begin{array}{lcl}

1708: f_{\min}(V')\!\!\!\! & \preccurlyeq\!\!\! & \bigl(f_{\min}(V) \ominus c_i(u)\bigr) \oplus \bigl( c'_i(u)\oplus c'_{ij}(u,v) \oplus c'_j(v)\bigr) \\

1709:               & \preccurlyeq\!\!\! & \bigl(f_{\min}(V) \ominus  \bigl(c_i(u) \oplus c_{ij}(u,v) \oplus c_j(v)\bigr)\bigr) \oplus\bigl(c'_i(u) \oplus c'_{ij}(u,v) \oplus c'_j(v)\bigr)

1710: \end{array}\]

1711: by equation~\ref{eqdac}. Thus $f_{\min}(V') \preccurlyeq f_{\min}(V)$ due to the

1712: equivalence of the subproblems of $V$ and $V'$ on $\{i,j\}$. Thus, $V$

1713: is $(2,f_{\min})$-irreducible.

1714: \end{proof}

1715: 

1716: This result must be considered with care. Given any VCSP $V$,

1717: Theorem~~\ref{dac2ir}. states that any directional arc consistent closure

1718: $V_{dac}$ of $V$ is always $(2,f_{\min})$-irreducible. However, an arc

1719: consistent closure $V_{ac}$ may exist such that $f_{\min}(V_{ac}) \succ

1720: f_{\min}(V_{dac})$.

1721: 

1722: \begin{corollary}

1723:   Arc-irreducibility with respect to $f_{\min}$ can be established

1724:   in $O(ed^2)$ time complexity and $O(ed)$ space complexity on fair

1725:   binary VCSPs.

1726: \end{corollary}

1727: \begin{proof}

1728:   This follows directly from Theorems~\ref{theodac} and~\ref{dac2ir}.

1729: \end{proof}

1730: 

1731: 

1732: %\begin{definition}

1733: %  A VCSP $V$ is \emph{star arc-irreducible} with respect to the

1734: %  problem evaluation function $f$ (or $(2,f)^*$-irreducible), if for

1735: %  all VCSP $V'$ derived from $V$ by a finite series of

1736: %  equivalence-preserving arc transformations.

1737: %  $f(V) \geq f(V')$.

1738: %\end{definition}

1739: 

1740: \section*{Conclusion}

1741: 

1742: The concept of arc consistency plays an essential role in constraint

1743: satisfaction as a problem simplification operation and as a

1744: tree-pruning technique during search through the detection of local

1745: inconsistencies among the uninstantiated variables. We have shown that

1746: it is possible to generalise arc consistency to any instance of the

1747: valued CSP framework provided the operator for aggregating penalties

1748: has an inverse.

1749: 

1750: A polynomial-time algorithm for establishing soft arc consistency

1751: exists. Its space and time complexity is identical to that of

1752: establishing arc consistency in CSPs whenever the aggregation operator

1753: of the VCSP is strictly monotonic, which is the case in

1754: \textsc{Max}-CSP, for example. Contrarily to classical CSP arc

1755: consistency, it does not define a unique arc consistency closure.

1756: This algorithm nevertheless provides an efficient technique for

1757: generating lower bounds on the value of a solution which can be used

1758: during branch-and-bound search as in~\cite{Schiex98,Schiex99}. The

1759: problem of finding the maximal lower bound is however

1760: \textsf{NP}-hard.

1761: 

1762: We have also defined a directional version of soft arc consistency

1763: which is potentially stronger since it allows non-local propagation of

1764: penalties. Directional soft arc consistency implies a form of local

1765: optimality in the expression of the VCSP, called arc

1766: irreducibility. Furthermore, the complexity of establishing

1767: directional arc consistency is identical to that of establishing arc

1768: consistency in CSPs.

1769: 

1770: %\bibliography{../../bib/CSP,../../bib/pos,../../bib/book}

1771: %\bibliographystyle{acm}

1772: \begin{thebibliography}{10}

1773: 

1774: \bibitem{Bennaceur98}

1775: {\sc Affane, M.~S., and Bennaceur, H.}

1776: \newblock A weighted arc consistency technique for {Max-CSP}.

1777: \newblock In {\em Proc.\ of the 13$^{th}$ ECAI\/} (Brighton, United Kingdom,

1778:   1998), pp.~209--213.

1779: 

1780: \bibitem{Bessiere97}

1781: {\sc Bessi{\`e}re, C., and R{\'e}gin, J.-C.}

1782: \newblock Arc consistency for general constraints networks: preliminary

1783:   results.

1784: \newblock In {\em Proc.\ of the 15$^{th}$ IJCAI\/} (Nagoya, Aichi, Japan,

1785:   1997), pp.~398--404.

1786: 

1787: \bibitem{Bistarelli99}

1788: {\sc Bistarelli, S., Fargier, H., Montanari, U., Rossi, F., Schiex, T., and

1789:   Verfaillie, G.}

1790: \newblock Semiring-based {CSP}s and valued {CSP}s: Frameworks, properties and

1791:   comparison.

1792: \newblock {\em Constraints 4\/} (1999), 199--240.

1793: 

1794: \bibitem{Bistarelli95}

1795: {\sc Bistarelli, S., Montanari, U., and Rossi, F.}

1796: \newblock Constraint solving over semirings.

1797: \newblock In {\em Proc.\ of the 14$^{th}$ IJCAI\/} (Montr{\'e}al, Canada, Aug.

1798:   1995).

1799: 

1800: \bibitem{Bistarelli97}

1801: {\sc Bistarelli, S., Montanari, U., and Rossi, F.}

1802: \newblock Semiring based constraint solving and optimization.

1803: \newblock {\em Journal of the ACM 44}, 2 (1997), 201--236.

1804: 

1805: \bibitem{CooperFCSP}

1806: {\sc Cooper, M.~C.}

1807: \newblock Reduction operations in fuzzy or valued constraint satisfaction.

1808: \newblock {\em Fuzzy Sets and Systems\/} (2002).

1809: \newblock To appear.

1810: 

1811: \bibitem{DPnormes}

1812: {\sc Dubois, D., and Prade, H.}

1813: \newblock A class of fuzzy measures based on triangular norms. a general

1814:   framework for the combination of uncertain information.

1815: \newblock {\em Int. Journal of Intelligent Systems 8}, 1 (1982), 43--61.

1816: 

1817: \bibitem{Schiex93c}

1818: {\sc Fargier, H., Lang, J., and Schiex, T.}

1819: \newblock Selecting preferred solutions in {F}uzzy {C}onstraint {S}atisfaction

1820:   {P}roblems.

1821: \newblock In {\em Proc. of the 1$^{st}$ European Congress on Fuzzy and

1822:   Intelligent Technologies\/} (1993).

1823: 

1824: \bibitem{Freuder92}

1825: {\sc Freuder, E., and Wallace, R.}

1826: \newblock Partial constraint satisfaction.

1827: \newblock {\em Artificial Intelligence 58\/} (Dec. 1992), 21--70.

1828: 

1829: \bibitem{Horsch00}

1830: {\sc Horsch, M.~C., and Havens, W.~S.}

1831: \newblock An empirical study of probabilistic arc consistency as a variable

1832:   ordering.

1833: \newblock In {\em Principles and Practice of Constraint Programming - CP

1834:   2000\/} (Singapore, Sept. 2000), vol.~1894 of {\em LNCS}, pp.~525--530.

1835: 

1836: \bibitem{Koster99b}

1837: {\sc Koster, A.~M.}

1838: \newblock {\em Frequency assignment: Models and Algorithms}.

1839: \newblock PhD thesis, University of Maastricht, The Netherlands, Nov. 1999.

1840: \newblock Available at \textsf{www.zib.de/koster/thesis.html}.

1841: 

1842: \bibitem{Schiex99}

1843: {\sc Larrosa, J., Meseguer, P., and Schiex, T.}

1844: \newblock Maintaining reversible {DAC} for {M}ax-{CSP}.

1845: \newblock {\em Artificial Intelligence 107}, 1 (Jan. 1999), 149--163.

1846: 

1847: \bibitem{Schiex98}

1848: {\sc Larrosa, J., Meseguer, P., Schiex, T., and Verfaillie, G.}

1849: \newblock Reversible {DAC} and other improvements for solving max-{CSP}.

1850: \newblock In {\em Proc.\ of AAAI'98\/} (Madison, WI, July 1998).

1851: 

1852: \bibitem{Mohr88}

1853: {\sc Mohr, R., and Masini, G.}

1854: \newblock Good old discrete relaxation.

1855: \newblock In {\em Proc. of the 8$^{th}$ ECAI\/} (Munchen FRG, 1988),

1856:   pp.~651--656.

1857: 

1858: \bibitem{Rosenfeld76}

1859: {\sc Rosenfeld, A., Hummel, R., and Zucker, S.}

1860: \newblock Scene labeling by relaxation operations.

1861: \newblock {\em {IEEE} Trans. on Systems, Man, and Cybernetics 6}, 6 (1976),

1862:   173--184.

1863: 

1864: \bibitem{Schiex92}

1865: {\sc Schiex, T.}

1866: \newblock Possibilistic constraint satisfaction problems or ``{H}ow to handle

1867:   soft constraints~?''.

1868: \newblock In {\em Proc. of the 8$^{th}$ Int. Conf. on Uncertainty in Artificial

1869:   Intelligence\/} (Stanford, CA, July 1992).

1870: 

1871: \bibitem{Schiex98a}

1872: {\sc Schiex, T.}

1873: \newblock Maximizing the reversible {DAC} lower bound in {M}ax-{CSP} is

1874:   \textsf{NP}-hard.

1875: \newblock Tech. Rep. 1998/02, INRA, July 1998.

1876: 

1877: \bibitem{Schiex00b}

1878: {\sc Schiex, T.}

1879: \newblock Arc consistency for soft constraints.

1880: \newblock In {\em Principles and Practice of Constraint Programming - CP

1881:   2000\/} (Singapore, Sept. 2000), vol.~1894 of {\em LNCS}, pp.~411--424.

1882: 

1883: \bibitem{Schiex95a}

1884: {\sc Schiex, T., Fargier, H., and Verfaillie, G.}

1885: \newblock Valued constraint satisfaction problems: hard and easy problems.

1886: \newblock In {\em Proc.\ of the 14$^{th}$ IJCAI\/} (Montr{\'e}al, Canada, Aug.

1887:   1995), pp.~631--637.

1888: 

1889: \bibitem{Shapiro81}

1890: {\sc Shapiro, L., and Haralick, R.}

1891: \newblock Structural descriptions and inexact matching.

1892: \newblock {\em {IEEE} Transactions on Pattern Analysis and Machine Intelligence

1893:   3\/} (1981), 504--519.

1894: 

1895: \bibitem{Wallace94}

1896: {\sc Wallace, R.}

1897: \newblock Directed arc consistency preprocessing.

1898: \newblock In {\em Selected papers from the ECAI-94 Workshop on Constraint

1899:   Processing}, M.~Meyer, Ed., no.~923 in LNCS. Springer, Berlin, 1995,

1900:   pp.~121--137.

1901: 

1902: \end{thebibliography}

1903: 

1904: \end{document}

1905: