Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Automated Reasoning Building BlocksPart I
Christoph Weidenbach
Max Planck Institute for Informatics
September 21, 2015
Motivation Propositional Reasoning
Outline
Motivation
Propositional Reasoning
September 21, 2015 2/54
Motivation Propositional Reasoning
Automated Reasoning Building Blocks
LearnFresh
DoIndexing
ConsiderModels
CompactDatastructs
FlexibleModels
Don’tGuess
FindInvariants
AlwaysLearn
EliminateRedundancy
ConsiderOrderings
BeLazy
ConsiderTheories
UseToolbox
BeSmall
September 21, 2015 3/54
Motivation Propositional Reasoning
Propositional Clause Logic
SyntaxClauses have the form P ∨ ¬R ∨Q ∨Q where P, Q, R ∈ Σ
Clauses denoted by C, D, empty clause denoted by ⊥Clause sets N, M are interpreted as conjunctions of clauses
Semantics(Partial) Valuations A : Σ→ {0, 1}Clause Set N satisfiable if A(N) = 1 for some A, A |= NClause Set N unsatisfiable if A(N) = 0 for all AClause Set N valid if A(N) = 1 for all A, |= N
N = {P ∨Q, ¬P ∨Q, P ∨ ¬Q, ¬P ∨ ¬Q}
September 21, 2015 4/54
Motivation Propositional Reasoning
Analytic Models
Beth 55, Smullyan 68, Fitting 90Tableau Procedure: Close Branch & Backtrack, Split Disjunction
N = {P ∨Q, R ∨ S, ¬P ∨Q, ¬Q}
⇒TAB ([],>) ⇒TAB ([Q],>)
⇒TAB ([PP∨Q],>) ⇒TAB ([Q ¬Q],⊥)
⇒TAB ([PP∨Q RR∨S],>) ⇒TAB ([],⊥)
⇒TAB ([PP∨Q RR∨S ¬P¬P∨Q],⊥)
⇒TAB ([PP∨Q RR∨S Q],>)
⇒TAB ([PP∨Q RR∨S Q ¬Q],⊥)
⇒TAB ([PP∨Q S],>) . . .
Only Guessing &No ModelConsideration
September 21, 2015 5/54
Motivation Propositional Reasoning
Consider Models
Davis & Logman & Loveland 1962DPLL Procedure: Find Conflict & Backtrack, Propagate, Guess
N = {P ∨Q, R ∨ S, ¬P ∨Q, ¬Q}
⇒DPLL ([],>)
⇒DPLL ([¬Q¬Q],>)
⇒DPLL ([¬Q¬QPP∨Q],>)
⇒DPLL ([¬Q¬QPP∨Q],¬P ∨Q)
⇒DPLL ([],⊥)
Propagation &No Guessing
September 21, 2015 6/54
Motivation Propositional Reasoning
DPLL Minimal Proof Length
N = {P ∨Q, ¬P ∨Q, P ∨ ¬Q, ¬P ∨ ¬Q}
⇒DPLL ([],>)
⇒DPLL ([P1],>)
⇒DPLL ([P1Q¬P∨Q],>)
⇒DPLL ([P1Q¬P∨Q],¬P ∨ ¬Q)
⇒DPLL ([¬P],>)
⇒DPLL ([¬P, QP∨Q],>)
⇒DPLL ([¬P, QP∨Q], P ∨ ¬Q)
⇒DPLL ([],⊥)
No LearningO(2n)
September 21, 2015 7/54
Motivation Propositional Reasoning
Always Learn
Robinson 1965Resolution: C ∨ P, ¬P ∨ D ⇒RES C ∨ D
N = {P ∨Q, ¬P ∨Q, P ∨ ¬Q, ¬P ∨ ¬Q}
¬P ∨Q, ¬P ∨ ¬Q ⇒RES ¬PP ∨Q, P ∨ ¬Q ⇒RES PP, ¬P ⇒RES ⊥
O(n)No ModelConsideration
September 21, 2015 8/54
Motivation Propositional Reasoning
Eliminate Redundancy
N = {P ∨Q, ¬P ∨ ¬Q}P ∨Q, ¬P ∨ ¬Q ⇒RES Q ∨ ¬QMay we eliminate Q ∨ ¬Q?
N = {P ∨Q ∨ R, ¬P ∨Q, P ∨ R}¬P ∨Q, P ∨ R ⇒RES Q ∨ RMay we eliminate P ∨Q ∨ R?
Number of clauses generated in a typical resolution run:42, 420, 42000, 42000000, . . .
September 21, 2015 9/54
Motivation Propositional Reasoning
Redundancy: A First-Class Citizen
Boyer 1971Lock Resolution: C ∨ Pk ,¬Pl ∨ D ⇒LRES C ∨ Dif k , l are maximal indexes, respectively
N = {P5 ∨Q4, ¬P4 ∨Q5, P4 ∨ ¬Q5, ¬P5 ∨ ¬Q4}⇒LRES Q4 ∨ ¬Q4
⇒LRES ¬P4 ∨ P4
Redundancy needs to be considered in the context of a calculus.
September 21, 2015 10/54
Motivation Propositional Reasoning
Learn Fresh, Eliminate Redundancy
Bachmair & Ganzinger 1990Superposition: C ∨ P,¬P ∨ D ⇒LRES C ∨ Dif P, ¬P are maximal in their respective clauses
Ordering≺ is a total strict ordering on Σ: P ≺ Q≺ on literals: P ≺ ¬P ≺ Q ≺ ¬Q≺ on clauses: multiset extension: {P, Q} ≺ {P, Q, Q} ≺ {¬Q}N≺C = {D ∈ N | D ≺ C}
September 21, 2015 11/54
Motivation Propositional Reasoning
Superposition Redundancy
Definition (Redundancy)
A clause C is redundant with respect to a clause set N if N≺C |= C.
P ≺ Q ≺ RN = {P ∨Q, R ∨ ¬P} clauses P ∨Q ∨ ¬R, Q ∨ R redundant
Superposition Static Model NINC :=
⋃D≺C δD
δD :=
{{P} if D = D′ ∨ P, P strictly maximal ND 6|= D∅ otherwise
NI :=⋃
C∈N δC
September 21, 2015 12/54
Motivation Propositional Reasoning
Superposition Results
Theorem (Completeness, Models, Redundancy)If all superposition inferences in N up to redundancy are performedand ⊥ /∈ N then N is satisfiable and NI |= N.It is sufficient to consider inferences between a minimal false clause¬P ∨ C, NI 6|= ¬P ∨ C and its productive counterpart P ∨ D.The result C ∨ D of the superposition inference is not redundant.
Model Properties
- fixed by ordering: P ∨Q, P ∨ ¬Q- minimal: P ∨Q, ¬P ∨ R, where R ≺ Q ≺ P then NI = {P}
Static Ordering &Model
September 21, 2015 13/54
Motivation Propositional Reasoning
Always Learn Fresh, Flexible Models
Silva, Sakallah, Bayardo, Schrag, Et Al 2000-CDCL: Find Conflict & Backtrack & Learn, Propagate, Guess
N = {P ∨Q, ¬P ∨Q, P ∨ ¬Q, ¬P ∨ ¬Q}
⇒CDCL ([], ∅,>)
⇒CDCL ([P1], ∅,>)
⇒CDCL ([P1Q¬P∨Q], ∅,>)
⇒CDCL ([P1Q¬P∨Q], ∅,¬P ∨ ¬Q)
⇒CDCL ([¬P], {¬P},>)
⇒CDCL ([¬P, QP∨Q], {¬P},>)
⇒CDCL ([¬P, QP∨Q], {¬P}, P ∨ ¬Q)
⇒CDCL ([], {¬P, ⊥},⊥)
O(n)No Redundancy
September 21, 2015 14/54
Motivation Propositional Reasoning
Consider Theories
while Program Analysis
1 n = 0; ¬P1(n, x , y) ∨ P2(0, x , y)
2 while (x > 0) { ¬x > 0 ∨ ¬P2(n, x , y) ∨ P3(n, x , y)
¬x ≤ 0 ∨ ¬P2(n, x , y) ∨ P6(n, x , y)
3 n = n + y ; ¬P3(n, x , y) ∨ P4(n + y , x , y)
4 x = x − 1; ¬P4(n, x , y) ∨ P5(n, x − 1, y)
5 } ¬P5(n, x , y) ∨ P2(n, x , y)
6 return n;
September 21, 2015 15/54
Motivation Propositional Reasoning
Axiomatize Theories
Program Formalization
¬x > 0 ∨ ¬P2(n, x , y) ∨ P3(n, x , y)¬x ≤ 0 ∨ ¬P2(n, x , y) ∨ P6(n, x , y)¬P3(n, x , y) ∨ P4(n + y , x , y)¬P4(n, x , y) ∨ P5(n, x − 1, y)
Theory Axioms
x = 0 ∨ x > 0 ∧ s(x) > xs(x) + y = s(x + y) ∧ 0 + y = y ∧ x + y = y + x. . .
Incomplete & No Decision Procedure
September 21, 2015 16/54
Motivation Propositional Reasoning
Combine Theories
Program Formalization
¬x > 0 ∨ ¬P2(n, x , y) ∨ P3(n, x , y)¬x ≤ 0 ∨ ¬P2(n, x , y) ∨ P6(n, x , y)¬P3(n, x , y) ∨ P4(n + y , x , y)¬P4(n, x , y) ∨ P5(n, x − 1, y)
Combinationx > 0 ‖¬P2(n, x , y) ∨ P3(n, x , y)x ≤ 0 ‖¬P2(n, x , y) ∨ P6(n, x , y)z = n + y ‖¬P3(n, x , y) ∨ P4(z, x , y)z = x − 1 ‖¬P4(n, x , y) ∨ P5(n, z, y)
Expressive Logic & Difficult AutomationSeptember 21, 2015 17/54
Motivation Propositional Reasoning
Summary
Open QuestionsCan redundancy and completeness be combined?Can redundancy and ordering restrictions be combined?Can model building and inferences be combined?How does reasoning in combination of theories work?
AnswersSuperposition [BachmairGanzinger90]CDCL [BayardoSchrag96, SilvaSakallah96, NieuwenhuisEtAl06]CDCL(T)/SMT [NieuwenhuisEtAl06]HierarchicSuperposition [BachmairEtAl94, KruglovW12, FietzkeW12]
September 21, 2015 18/54
Motivation Propositional Reasoning
Disclaimer
Simplified PresentationMany Technical Details OmittedAlmost no EqualitySee References for Further Reading
GoalGet the Ideas/Intuition
September 21, 2015 19/54
Motivation Propositional Reasoning
References Tableau
E.W. Beth.Semantic entailment and formal derivability.Mededelingen van de Koninklijke Nederlandse Akademie vanWetenschappen, Afdeling Letterkunde, 18(13):309–342, 1955.
Raymond M. Smullyan.First-Order Logic.Ergebnisse der Mathematik und ihrer Grenzgebiete. Springer,1968.
Melvin Fitting.First-Order Logic and Automated Theorem Proving.Texts and Monographs in Computer Science. Springer, 1990.
September 21, 2015 20/54
Motivation Propositional Reasoning
References Superposition
Leo Bachmair and Harald Ganzinger.On restrictions of ordered paramodulation with simplification.In CADE-10, LNCS 449, pages 427–441. Springer, 1990.
R.S. Boyer.Locking: A Restriction of Resolution.PhD thesis, University of Texas at Austin, August 1971.
John Alan Robinson.A machine-oriented logic based on the resolution principle.Journal of the ACM, 12(1):23–41, January 1965.
September 21, 2015 21/54
Motivation Propositional Reasoning
References CDCL
Roberto J. Bayardo Jr. and Robert Schrag.Using CSP look-back techniques to solve exceptionally hard SATinstances.In Eugene C. Freuder, editor, CP 1996, Cambridge,Massachusetts, USA, August 19-22, LNCS 1118, pages 46–60.Springer, 1996.
João P. Marques Silva and Karem A. Sakallah.Grasp - a new search algorithm for satisfiability.In ICCAD 1996, pages 220–227. IEEE Press, 1996.
Robert Nieuwenhuis, Albert Oliveras, and Cesare Tinelli.Solving sat and SAT modulo theories: From an abstractDavis–Putnam–Logemann–Loveland procedure to DPLL(T).Journal of the ACM, 53:937–977, November 2006.
September 21, 2015 22/54
Motivation Propositional Reasoning
References Combination
Leo Bachmair, Harald Ganzinger, and Uwe Waldmann.Refutational theorem proving for hierarchic first-order theories.AAECC, 5(3/4):193–212, 1994.
Evgeny Kruglov and Christoph Weidenbach.Superposition decides the first-order logic fragment over groundtheories.MCS, 6(4):427–456, 2012.
Arnaud Fietzke and Christoph Weidenbach.Superposition as a decision procedure for timed automata.MCS, 6(4):409–425, 2012.
September 21, 2015 23/54
Motivation Propositional Reasoning
Propositional Resolution
Resolution [Robinson65]Resolution(N ] {C ∨ P, D ∨ ¬P}) ⇒RES (N ∪ {C ∨ P, D ∨ ¬P, C ∨ D})
Factoring(N ] {C ∨ L ∨ L}) ⇒RES (N ∪ {C ∨ L ∨ L} ∪ {C ∨ L})
Theorem (Resolution is Sound and Complete)N is unsatisfiable iff N ⇒∗RES N ′ ∪ {⊥}
September 21, 2015 24/54
Motivation Propositional Reasoning
Soundness
Resolution(N ] {C ∨ P, D ∨ ¬P}) ⇒RES (N ∪ {C ∨ P, D ∨ ¬P, C ∨ D})
if A((C ∨ P) ∧ (D ∨ ¬P)) = 1 then A(C ∨ D) = 1
Factoring(N ] {C ∨ L ∨ L}) ⇒RES (N ∪ {C ∨ L ∨ L} ∪ {C ∨ L})
if A(C ∨ L ∨ L) = 1 then A(C ∨ L) = 1
September 21, 2015 25/54
Motivation Propositional Reasoning
Completeness: Semantic Trees
N = {P ∨Q, ¬P ∨ ¬Q, P ∨ ¬Q, ¬P ∨Q ∨ R, ¬P ∨Q ∨ ¬R}
¬P ∨ ¬Q
Q
¬P ∨Q ∨ ¬R
R
¬P ∨Q ∨ R
¬R
¬Q
P
P ∨ ¬Q
Q
P ∨Q
¬Q
¬P
⇒RES P ∨ P⇒RES P
September 21, 2015 26/54
Motivation Propositional Reasoning
Semantic Tree Redundancy
N = {P ∨Q, ¬P ∨ ¬Q, P ∨ ¬Q, ¬P ∨Q ∨ R, ¬P ∨Q ∨ ¬R, P}
¬P ∨ ¬Q
Q
¬P ∨Q ∨ ¬R
R
¬P ∨Q ∨ R
¬R
¬Q
P
P
¬P
⇒RES P ∨ P⇒RES Premoved P ∨ P
September 21, 2015 27/54
Motivation Propositional Reasoning
Eliminate Redundancy
Well-Founded Semantic Tree OrderingA clause C is redundant if D ⊆ C for some DAny clause C not occurring in any semantic tree is redundant
Subsumption(N ] {C, D}) ⇒RES (N ∪ {C})provided C ⊂ D
Condensation(N ] {C ∨ L ∨ L}) ⇒RES (N ∪ {C ∨ L})
Tautology Deletion(N ] {C ∨ P ∨ ¬P}) ⇒RES (N)
Can redundancy and completeness be combined?September 21, 2015 28/54
Motivation Propositional Reasoning
TerminationN0 ⇒RES N1 ⇒RES . . .
How Many Different Clause Sets?at most 3n different clauses, n = |Σ|, modulo condensationat most 2(3n) different clause sets
ExamplesN = {P ∨ C1, P ∨ C2, . . .} obviously satisfiableN = {P ∨Q, P ∨ ¬Q}
under A(P) = 1 why consider P ∨ ¬Q for resolution?
Can model building and inferences be combined?
September 21, 2015 29/54
Motivation Propositional Reasoning
Propositional Superposition
Ordering≺ is a total strict ordering on Σ: P ≺ Q≺ on literals: P ≺ ¬P ≺ Q ≺ ¬Q≺ on clauses: multiset extension: {P, Q} ≺ {P, Q, Q} ≺ {¬Q}N≺C = {D ∈ N | D ≺ C}
Definition (Redundancy)
A clause C is redundant with respect to a clause set N if N≺C |= C.
P ≺ Q ≺ RN = {P ∨Q, R ∨ ¬P} clauses P ∨Q ∨ ¬R, Q ∨ R redundant
September 21, 2015 30/54
Motivation Propositional Reasoning
ModelsDefinition (Selection Function)
sel maps clauses to one of its negative literals or ⊥if sel(C) = ¬P then ¬P is called selected in Cif sel(C) = ⊥ then no literal in C is selected
Partial Herbrand Model Construction
NC :=⋃
D≺C δD
δD :=
{P} if D = D′ ∨ P, P strictly maximal, no literal
selected in D and ND 6|= D∅ otherwise
NI :=⋃
C∈N δC
September 21, 2015 31/54
Motivation Propositional Reasoning
Partial Herbrand Model Properties
NC :=⋃
D≺C δD
δD :=
{P} if D = D′ ∨ P, P strictly maximal, no literal
selected in D and ND 6|= D∅ otherwise
NI :=⋃
C∈N δC
PropertiesNI is minimal with respect to set inclusionif D ≺ C and NC |= D then NI |= Dif P ∨ P ≺ C then δC 6= {P}if δC = {P} then NC ∪ δC |= C
September 21, 2015 32/54
Motivation Propositional Reasoning
Propositional Superposition
SuperpositionSuperposition Left(N ] {C ∨ P, D ∨ ¬P}) ⇒SUP (N ∪ {C ∨ P, D ∨ ¬P} ∪ {C ∨ D})where (i) P is strictly maximal in C ∨ P(ii) no literal in C ∨ P is selected(iii) ¬P is maximal and no literal selected in D ∨ ¬P, or¬P is selected in D ∨ ¬P
Factoring(N ] {C ∨ P ∨ P}) ⇒SUP (N ∪ {C ∨ P ∨ P} ∪ {C ∨ P})where (i) P is maximal in C ∨ P ∨ P(ii) no literal is selected in C ∨ P ∨ P
September 21, 2015 33/54
Motivation Propositional Reasoning
Saturation and Completeness
Definition (Saturation)A set N of clauses is called saturated up to redundancy, if any clausegenerated by Superposition Left or Factoring from non-redundantclauses in N is redundant with respect to N or contained in N.
Theorem (Superposition Completeness)If N is saturated up to redundancy and ⊥ /∈ N then N is satisfiableand NI |= N.
September 21, 2015 34/54
Motivation Propositional Reasoning
Superposition Completeness Proof
Theorem (Superposition Completeness)If N is saturated up to redundancy and ⊥ /∈ N then N is satisfiableand NI |= N.
Proof.By contradiction. I assume:(i) if N ⇒SUP N ∪ D then N≺D |= D or D ∈ N, (ii) ⊥ /∈ N and(iii) NI 6|= N.Then there is a minimal clause C ∨ L ∈ N such that NI 6|= C ∨ Land L is selected or nothing selected and L maximal. This clausemust exist because ⊥ /∈ N.C ∨ L is not redundant and by Superposition Left or Factoring wecan derive a non-redundant clause that is smaller than C ∨ L andfalse in NI , a contradiction.
September 21, 2015 35/54
Motivation Propositional Reasoning
Proof: More Details.if C ∨ L redundant, then NC∨L |= C ∨ L, a contradictionL positive, L not strictly maximal, do Factoring, a contradictionL positive, L strictly maximal, L ∈ NI , a contradictionL negative, do Superposition Left, a contradiction
September 21, 2015 36/54
Motivation Propositional Reasoning
Completeness Summary
Superposition Completenessdoes inferences with respect to a candidate model NIonly inferences on false clauses needed: see proofsupports ordering restrictions: ≺supports redundancy: NC |= C
September 21, 2015 37/54
Motivation Propositional Reasoning
Superposition Results
Theorem (Completeness, Models, Redundancy)If all superposition inferences in N up to redundancy are performedand ⊥ /∈ N then N is satisfiable and NI |= N.It is sufficient to consider inferences between a minimal false clause¬P ∨ C, NI 6|= ¬P ∨ C and its productive counterpart P ∨ D.The result C ∨ D of the superposition inference is not redundant.
Propositional SpecialitiesNI can be effectively constructedNI 6|= ¬P ∨ C easy to decideunsatisfiability is co-NP, saturation always terminates
September 21, 2015 38/54
Motivation Propositional Reasoning
Concrete Redundancy
Subsumption(N ] {C, D}) ⇒SUP (N ∪ {C})provided C ⊂ D
Condensation(N ] {C ∨ L ∨ L}) ⇒SUP (N ∪ {C ∨ L})
Tautology Deletion(N ] {C ∨ P ∨ ¬P}) ⇒SUP (N)
Subsumption Resolution(N ] {C1 ∨ L, C2 ∨ ¬L}) ⇒SUP (N ∪ {C1 ∨ L, C2})where C1 ⊆ C2
September 21, 2015 39/54
Motivation Propositional Reasoning
Extension I: Saturation
Example (N saturated vs. NI |= N)
N = {P, ¬P ∨ ¬Q} with ordering Q ≺ P.Then NI |= N but N is not saturated.
FixAlways check whether NI |= N.
September 21, 2015 40/54
Motivation Propositional Reasoning
Extension II: Flexible Model Operator
Example (Inflexible Model Operator)
N = {P ∨Q, ¬P ∨ R} with ordering R ≺ Q ≺ P.Then NI = {P} and NI 6|= NBut NI ∪ {R} |= N.
FixUse a different model operator.
September 21, 2015 41/54
Motivation Propositional Reasoning
Flexible Model Operator
H is a decision heuristic: H : Σ→ {0, 1}
NHP :=
⋃Q≺P δ
HQ
δHP :=
{P} if (D ∨ P) ∈ N, with NH
P |= ¬Dand P strictly maximal, nothing selected orH(P) = 1 no clause (D′ ∨ ¬P) ∈ N, D′ ≺ Psuch that NH
P |= ¬D′
∅ otherwise
NHΣ :=
⋃P∈Σ δ
HP
September 21, 2015 42/54
Motivation Propositional Reasoning
NHΣ Properties [W2015]
Theorem (Superposition Completeness)If N is saturated up to redundancy and ⊥ /∈ N then N is satisfiableand NH
Σ |= N.
Example (Flexible Model Operator)
N = {P ∨Q, ¬P ∨ R} with ordering R ≺ Q ≺ P, H(R) = 1Then NH
Σ = {P, R} and NHΣ |= N
September 21, 2015 43/54
Motivation Propositional Reasoning
CDCL States
(ε; N; ∅; 0;>) is the start state for some clause set N(M; N; U; k ;>) is a final state, if M |= N and all literals from N
are defined in M(M; N; U; k ;⊥) is a final state, where N has no model(M; N; U; k ;>) is an intermediate model search state if M 6|= N or
not all literals from N are defined in M(M; N; U; k ; D) is a backtracking state if D 6∈ {>,⊥}
September 21, 2015 44/54
Motivation Propositional Reasoning
CDCL Rules I
Model Extension RulesPropagate(M; N; U; k ;>) ⇒CDCL (MLC∨L; N; U; k ;>)
provided C ∨ L ∈ (N ∪ U), M |= ¬C, and L is undefined in MDecide(M; N; U; k ;>) ⇒CDCL (MLk+1; N; U; k + 1;>)
provided L is undefined in MConflict(M; N; U; k ;>) ⇒CDCL (M; N; U; k ; D)
provided D ∈ (N ∪ U) and M |= ¬D
September 21, 2015 45/54
Motivation Propositional Reasoning
CDCL Rules II
Backtracking RulesSkip(MLC∨L; N; U; k ; D) ⇒CDCL (M; N; U; k ; D)
provided D 6∈ {>,⊥} and comp(L) does not occur in DResolve(MLC∨L; N; U; k ; D ∨ comp(L)) ⇒CDCL (M; N; U; k ; D ∨ C)
provided D is of level kBacktrack(M1K i+1M2; N; U; k ; D ∨ L) ⇒CDCL (M1LD∨L; N; U ∪{D ∨ L}; i ;>)
provided L is of level k and D is of level i .
September 21, 2015 46/54
Motivation Propositional Reasoning
CDCL Properties
Theorem (CDCL Soundness)CDCL terminates reasonably in two different final states:(M; N; U; k ;>) where M |= N and (M; N; U; k ;⊥) where N isunsatisfiable.
Theorem (CDCL Strong Completeness)For any interpretation M, there is a reasonable sequence of ruleapplications generating (M ′; N; U; k ;>) as a final state, where Mand M ′ only differ in the order of literals.
Theorem (CDCL Termination)CDCL always terminates reasonably in a state (M; N; U; k ; D) withD ∈ {>,⊥}.
September 21, 2015 47/54
Motivation Propositional Reasoning
Superposition & CDCL
Theorem (NHΣ & CDCL)
If (L1 . . . Ln; N; U; k ;>) is a CDCL state, atom(Li) = Pi andP1 ≺ P2 ≺ . . . ≺ Pn and H(Pi) = 1 if Pi is a decision literal, thenNH
P1,P2,...,Pncontains exactly the atoms from L1 . . . Ln.
CDCL & RedundancyA learned CDCL clause is the result of a superposition inference andnot redundant.
September 21, 2015 48/54
Motivation Propositional Reasoning
Extension III: Ordering Change
N = { ¬P1 ∨ P2 ∨ . . . ∨ Pn, P1 ∨ P2 ∨ . . . ∨ Pn¬P2 ∨ P′
2, ¬P2 ∨ ¬P′2,
. . .¬Pn ∨ P′
n, ¬Pn ∨ ¬P′n}
O(n) refutation: Pn ≺ Pn−1 . . . ≺ P1 ≺ P′n ≺ P′
n−1 . . . ≺ P′2
O(2n) refutation: Pn � Pn−1 . . . � P1 � P′n � P′
n−1 . . . � P′2
N = { Q ∨ ¬P1 ∨ P2 ∨ . . . ∨ Pn, Q ∨ P1 ∨ P2 ∨ . . . ∨ PnQ ∨ ¬P2 ∨ P′
2, Q ∨ ¬P2 ∨ ¬P′2,
. . .Q ∨ ¬Pn ∨ P′
n, Q ∨ ¬Pn ∨ ¬P′n}
Consider N ∪ N{Pi 7→ P′i , P′
i 7→ Pi , Q 7→ ¬Q}
September 21, 2015 49/54
Motivation Propositional Reasoning
Flexible Model & Redundancy
N = {P ∨Q, R ∨ ¬P} ordering R ≺ Q ≺ P model NI = {P}
⇒SUP Q ∨ R not redundant
change ordering P ≺ Q ≺ R
now Q ∨ R redundant
- flexible ordering not compatible with redundancy
- superposition redundancy is not compatible with CDCL reasoning
- flexible models enable only weaker notions of redundancy
- still: at any point in time, learned CDCL clauses are non-redundant
September 21, 2015 50/54
Motivation Propositional Reasoning
Be Small
Implementing CDCLis subject for an independent tutoriallike resolution CDCL learns many clausesnot redundant at creation, but become redundantabout 10% are subsumed by subsequent learned clausechecking subsumption for every learned clause is too expensivegreedily throw away learned clauses by activity heuristicslowly increase number of overall kept learned clauses
September 21, 2015 51/54
Motivation Propositional Reasoning
Summary
Propositional Superpositioncombines static models, fresh learning, redundancy eliminationexplicit orderingordering changes conflict with redundancy
CDCLcombines flexible models, fresh learningimplicit orderingmodel changes conflict with redundancy
Future ResearchRelationship between flexible models and redundancy.
September 21, 2015 52/54
Motivation Propositional Reasoning
References Propositional Reasoning
Christoph Weidenbach.Automated Reasoning.Lecture Script WS14/15.http://www.mpi-inf.mpg.de/departments/automation-of-logic/teaching/.
Christoph Weidenbach.Automated reasoning building blocks.In Roland Meyer, André Platzer, and Heike Wehrheim, editors,Correct System Design, volume 9360 of LNCS, pages 172–188.Springer, 2015.
Christoph Weidenbach.Automated Reasoning.CRC. 201X. To appear.
September 21, 2015 53/54
Motivation Propositional Reasoning
References Propositional Reasoning
Armin Biere, Marijn Heule, Hans van Maaren, and Toby Walsh,editors.Handbook of Satisfiability, volume 185 of Frontiers in ArtificialIntelligence and Applications. IOS Press, 2009.
September 21, 2015 54/54