A Logic of Partially Satisfied Constraints Nic Wilson Cork Constraint Computation Centre Computer Science, UCC.

Slides:



Advertisements
Similar presentations
Non-Binary Constraint Satisfaction Toby Walsh Cork Constraint Computation Center.
Advertisements

Completeness and Expressiveness
Computer Science CPSC 322 Lecture 25 Top Down Proof Procedure (Ch 5.2.2)
CSCI 115 Chapter 6 Order Relations and Structures.
10 October 2006 Foundations of Logic and Constraint Programming 1 Unification ­An overview Need for Unification Ranked alfabeths and terms. Substitutions.
1 Finite Constraint Domains. 2 u Constraint satisfaction problems (CSP) u A backtracking solver u Node and arc consistency u Bounds consistency u Generalized.
Relations Relations on a Set. Properties of Relations.
Approximation Algorithms Chapter 14: Rounding Applied to Set Cover.
Hoare’s Correctness Triplets Dijkstra’s Predicate Transformers
Constraint Optimization Presentation by Nathan Stender Chapter 13 of Constraint Processing by Rina Dechter 3/25/20131Constraint Optimization.
Copyright © Cengage Learning. All rights reserved.
(CSC 102) Discrete Structures Lecture 14.
Bounds on Code Length Theorem: Let l ∗ 1, l ∗ 2,..., l ∗ m be optimal codeword lengths for a source distribution p and a D-ary alphabet, and let L ∗ be.
Database Management COP4540, SCS, FIU Functional Dependencies (Chapter 14)
Algebra Problems… Solutions Algebra Problems… Solutions © 2007 Herbert I. Gross Set 22 By Herbert I. Gross and Richard A. Medeiros next.
CMPT 354, Simon Fraser University, Fall 2008, Martin Ester 52 Database Systems I Relational Algebra.
A Semantic Characterization of Unbounded-Nondeterministic Abstract State Machines Andreas Glausch and Wolfgang Reisig 1.
1 2. Constraint Databases Next level of data abstraction: Constraint level – finitely represents by constraints the logical level.
Discrete Mathematics Lecture 4 Harper Langston New York University.
Approximation Algorithms
Winter 2004/5Pls – inductive – Catriel Beeri1 Inductive Definitions (our meta-language for specifications)  Examples  Syntax  Semantics  Proof Trees.
1 CMSC424, Spring 2005 CMSC424: Database Design Lecture 9.
Theory and Applications
Dr. Alexandra I. Cristea CS 319: Theory of Databases: C3.
Partial Orderings: Selected Exercises
Proof by Deduction. Deductions and Formal Proofs A deduction is a sequence of logic statements, each of which is known or assumed to be true A formal.
1 Relational Algebra and Calculus Yanlei Diao UMass Amherst Feb 1, 2007 Slides Courtesy of R. Ramakrishnan and J. Gehrke.
Partially Ordered Sets (POSets)
Chapter 6. Order Relations and Structure
The Game of Algebra or The Other Side of Arithmetic The Game of Algebra or The Other Side of Arithmetic © 2007 Herbert I. Gross by Herbert I. Gross & Richard.
1 Constraint Programming: An Introduction Adapted by Cristian OLIVA from Peter Stuckey (1998) Ho Chi Minh City.
Chapter 10 Functional Dependencies and Normalization for Relational Databases.
Database Systems Normal Forms. Decomposition Suppose we have a relation R[U] with a schema U={A 1,…,A n } – A decomposition of U is a set of schemas.
Constraint Satisfaction Problems (CSPs) CPSC 322 – CSP 1 Poole & Mackworth textbook: Sections § Lecturer: Alan Mackworth September 28, 2012.
The Integers. The Division Algorithms A high-school question: Compute 58/17. We can write 58 as 58 = 3 (17) + 7 This forms illustrates the answer: “3.
Pattern-directed inference systems
Advanced Topics in Propositional Logic Chapter 17 Language, Proof and Logic.
Slide 1 Propositional Definite Clause Logic: Syntax, Semantics and Bottom-up Proofs Jim Little UBC CS 322 – CSP October 20, 2014.
Computing & Information Sciences Kansas State University Tuesday, 27 Feb 2007CIS 560: Database System Concepts Lecture 18 of 42 Tuesday, 27 February 2007.
CP Summer School Modelling for Constraint Programming Barbara Smith 2. Implied Constraints, Optimization, Dominance Rules.
1 Lecture 6: Schema refinement: Functional dependencies
Copyright © Curt Hill Quantifiers. Copyright © Curt Hill Introduction What we have seen is called propositional logic It includes.
CS201: Data Structures and Discrete Mathematics I
CS 267: Automated Verification Lecture 3: Fixpoints and Temporal Properties Instructor: Tevfik Bultan.
Relations and their Properties
Copyright © Cengage Learning. All rights reserved.
Week 6 - Friday.  What did we talk about last time?  Solving recurrence relations.
Information and Coding Theory Cyclic codes Juris Viksna, 2015.
CS6133 Software Specification and Verification
Computer Science CPSC 322 Lecture 22 Logical Consequences, Proof Procedures (Ch 5.2.2)
Functional Dependencies CIS 4301 Lecture Notes Lecture 8 - 2/7/2006.
Chapter 8: Relations. 8.1 Relations and Their Properties Binary relations: Let A and B be any two sets. A binary relation R from A to B, written R : A.
Logic: Proof procedures, soundness and correctness CPSC 322 – Logic 2 Textbook §5.2 March 7, 2011.
Set Theory Concepts Set – A collection of “elements” (objects, members) denoted by upper case letters A, B, etc. elements are lower case brackets are used.
Chapter 8 Relational Database Design. 2 Relational Database Design: Goals n Reduce data redundancy (undesirable replication of data values) n Minimize.
Linear Programming Chap 2. The Geometry of LP  In the text, polyhedron is defined as P = { x  R n : Ax  b }. So some of our earlier results should.
1 IAS, Princeton ASCR, Prague. The Problem How to solve it by hand ? Use the polynomial-ring axioms ! associativity, commutativity, distributivity, 0/1-elements.
Theory of Computational Complexity Probability and Computing Chapter Hikaru Inada Iwama and Ito lab M1.
Chapter 5 Relations and Operations
Partial Orderings: Selected Exercises
Chapter 6 Order Relations and Structures
Richard Dedekind ( ) The grandfather of mathematical structuralism
Fixpoints and Reachability
Copyright © Cengage Learning. All rights reserved.
Unit-III Algebraic Structures
Handout 4 Functional Dependencies
Chap 3. The simplex method
Concurrent Models of Computation
Lecture 20: Dataflow Analysis Frameworks 11 Mar 02
Computer Science cpsc322, Lecture 20
Presentation transcript:

A Logic of Partially Satisfied Constraints Nic Wilson Cork Constraint Computation Centre Computer Science, UCC

Soft Constraints Soft Constraints formalisms tend to involve some form of degrees: degrees to which tuples satisfy a constraint (or set of constraints). These degrees can represent, for example, (a) costs, of not satisfying constraints in an over-constrained problem (b) extent of satisfaction of a vague constraint – constraints can be partially satisfied by tuples (c) degree of uncertainty, given uncertain constraint – we’re unsure about the identity of the constraint

Set of variables V. Each variable v has domain v. For U  V, U is set of possible assignments to U, i.e.,  v  U v A complete tuple x is an element of V. A constraint c has an associated set of variables V c  V. c is defined to be a subset of V c (so a set of partial tuples). Associated with constraint c is the constraint c  V = {x  V : x  Vc  c} on variables V: the set of complete tuples which project to an element of c. c may be considered as a compact representation of c  V. For example, let V = {V 1, V 2, V 3, V 4, V 5 }, V 1 = {a 1, b 1 }, V 2 = {a 2, b 2 } etc. Let U = {V 1, V 2 }. Then U = { (a 1, a 2 ), (a 1, b 2 ), (b 1, a 2 ), (b 1, b 2 ) }. Define constraint c by: V c = {V 1, V 2 }, c = { (a 1, a 2 ), (b 1, b 2 ) }. x = (a 1, a 2, b 3, a 4, b 5 ) is a complete tuple. x  c  V since x  Vc = (a 1, a 2 ) is in c. Finite CSPs as a Logic

The constraints are for a particular purpose. There is a set M of complete tuples that are adequate for this purpose. M is unknown: constraints give us information about it. Each constraint c is viewed as a restriction on complete tuples: it says that if a tuple is not in c, then it is inadequate for the purpose (so constraints are taken to give negative information). c tells us that M  c  V : any complete tuple outside c  V is inadequate. Define a model M to be a subset of V. Each M is a possible candidate for M. M |= c (M is a model of c) iff M  c  V Let K be a set of constraints, and d be a constraint. We define K |= d (K entails d) iff every model of (every element of) K is a model of d. In other words, if K tells us that every tuple outside d is inadequate for our purpose. Semantics

Combination of constraints. Constraints c and d. c  d is the constraint c  U  d  U on variables U = V c  V d. So y  U is in c  d iff y  Vc  c and y  Vd  d. c  d is essentially the intersection, as c  U is essentially the same constraint as c. This operation is commutative and associative. Projection of constraints For constraint c, and U  V c, c  U, the projection of c to U, is defined to be {y  U : y  c}. Identity (trivial) constraint Let 1 V be the constraint V. (Every tuple is allowed.) This gives us no information, and is satisfied by all models.

Axiom: 1 V Inference Rules: From c and d deduce c  d When T  V c : From c deduce c  T When V c = V d and c  d: From c deduce d The proof theory is sound and complete: K |= d if and only if d can be derived from K  {1 V } using the above rules. Proof Theory

Derived inference rules Arc consistency can be considered as derived inference rules: combination of projections to a single variable.

Derived Inference Rule: Deletion of a variable V 1 : Combine all constraints involving that variable and project to U - {V 1 }: (  {c  K : V 1  V c })  U – {V1} where U is the set of variables involved in the combination. (  K)  Vd can be computed by repeated deletion of variables; the deletion inference rule gives a sound and complete proof procedure (Dechter and Pearl’s adaptive consistency algorithm; special case of Shenoy’s fusion algorithm, Dechter’s bucket elimination).

A constraint c, as defined above, can also be viewed as a function from V c to {0,1}: assigning 1’s to partial tuples in the constraint. Similarly, models can be viewed as functions from V to {0,1}. Then we have: M |= c iff M  c  V, i.e., iff for all tuples x  V, M(x)  c  V (x), i.e., M(x)  c(x  Vc ). c  d is the constraint on V c  V d given by (c  d) (y) = c(y  Vc )  d(y  Vd ) where the last  is logical AND (i.e., min). c  U, the projection of c to U, c  U (u) =  {c(y): y  U = u} where  is logical OR (i.e., max). 1 U is the constraint on variables U which is everywhere equal to 1. The only difference in the proof theory is that the subset inference rule is replaced by: When V c = V d and c  d: From c deduce d since: c  d, i.e., for all y, c(y)  d(y), iff every tuple in c is a tuple in d.

Suppose we now want to allow degrees of satisfaction of constraints. We choose a partially (or totally) ordered set (A, , 0, 1) to represent these degrees, where A contains a unique maximal element 1 and a unique minimal element 0. A (soft) constraint c is defined to be a function from V c to A, with a value of 1 meaning that the tuple completely satisfies the constraint, and a value of 0 meaning that the tuple doesn’t satisfy the constraint at all (tuple is known to be completely inadequate for our purpose). Define models to be functions from V to A. A model is intended to represent the ‘true’ degree to which a tuple is adequate for our purpose. Each constraint is viewed as restricting the set of models: it gives upper bounds on these ‘true’ degrees. M |= c iff M  c  V, i.e., iff for all tuples x  V, M(x)  c  V (x), i.e., M(x)  c(x  Vc ). As before, for set of (soft) constraints K, and (soft) constraint d, we say K |= d iff every model M of K is also model of d. Partially Satisfied Constraints

When A is a distributive lattice An example of a distributive lattice is a subset lattice: Let B be a set. Let A be a set of subsets of B, which is closed under intersection and union: i.e., if ,   A then ,   B, and   ,     A. In fact any finite distributive lattice is isomorphic to such a subset lattice.    is the unique greatest lower bound of  and .    is the unique least upper bound of  and . This enables us to define combination and projection, with the same form as before. Let c: V c  A and d: V d  A be two soft constraints. Their combination c  d is the constraint on V c  V d. given by (c  d) (y) = c(y  Vc )  d(y  Vd ) For constraint c, and U  V c, c  U, the projection of c to U, is defined by c  U (u) =  {c(y): y  U = u}. The degree of the projection is the least upper bound (meet) of c(y) over all y which project to u.

Axiom: 1 V Inference Rules: From c and d deduce c  d When U  V c : From c deduce c  U When V c = V d and c  d: From c deduce d Again we have a form of arc consistency as sound (derived) inference rules; and also the deletion inference rule leads again to a sound and complete inference procedure which is efficient if there’s a nice hypertree cover of the hypergraph {V c : c  K}. Sound and Complete Proof Theory

Idempotent c-semi-rings are distributive lattices; the combination operation is the same as described above, and the summation is the projection defined above. The deduction is the same. So we have a new semantics for deduction in idempotent semi-ring CSPs. Relationship with idempotent semi-ring CSPs

Any partial order can be embedded in a subset lattice (so that the ordering is maintained). In particular we can choose A  A, and for  A define X  to be {  A :    }. If A is chosen sufficiently large we have    iff X   X , so that ({X  :  A},  ) is a representation of the partial order (A,  ). ( We also need the following equivalence, which holds for the above representation: For all  A, ( X    {X  :  B}  X    {X  ’ :  B} ) is equivalent to  {X  :  B}   {X  ’ :  B}. ) We can thus use the above proof procedures on this embedding to get sound and complete deduction procedures for the general case. The General Partially Ordered Case

Models in the above semantics are functions from V to partially ordered set A: they assign partially ordered degrees to complete tuples. It can also be natural to define models as functions from V to some partially ordered set A that extends A. Otherwise we use the same definitions to define the entailment relation. The intuition is that the partially ordered values mentioned in the constraints are from some unknown partially ordered set. This leads to a much weaker entailment (set of constraints K will often have many less consequences d). A Different Semantics for Partially Satisfied Constraints

Summary A logical approach was taken to partially satisfied constraints, where tuples can be assigned values from an arbitrary partially ordered set. The logic has a simple semantics, and sound and complete proof theory. This gives a new semantics for idempotent semi- ring CSP deduction.

For example, if we choose A = A so that for   A, we define X  = {  A :    }. Then    implies X   X  : this follows by transitivity of  : if  X  then    so    so   X  And X   X  implies    : this follows by reflexivity of  :    so  X   X  therefore    However, we can usually find much smaller sets A than A leading to a more efficient representation of the partial order.