1 / 48

Hypothetical Reasoning in Propositional Satisfiability

Hypothetical Reasoning in Propositional Satisfiability. Joao Marques-Silva Technical University of Lisbon, IST/INESC, CEL Lisbon, Portugal. SAT’02, May, 2002. Hypothetical Reasoning. Preliminary ongoing research work Not yet published

ahava
Download Presentation

Hypothetical Reasoning in Propositional Satisfiability

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Hypothetical Reasoning inPropositional Satisfiability Joao Marques-Silva Technical University of Lisbon, IST/INESC, CEL Lisbon, Portugal SAT’02, May, 2002

  2. Hypothetical Reasoning • Preliminary ongoing research work • Not yet published • Main ideas available as a (preliminary) technical report: • I. Lynce and J. Marques-Silva, “Hypothetical Reasoning in Propositional Satisfiability,” Technical Report 1/2002, INESC-ID, March 2001 • http://sat.inesc.pt/~jpms/research/tech-reports/RT-01-2002.pdf • Some of the concepts still evolving • Feedback welcome ! • Joint work with Ines Lynce

  3. Motivation • SAT solvers have been the subject of significant improvements in recent years • The utilization of SAT is increasing in industry • More challenging problem instances • Improvements to current key techniques unlikely(?) • Better (non-chronological) backtracking? • Better data structures? • Newer (more competitive) strategies? • How to improve SAT solvers? • Devise new paradigms… • Integrate already used techniques

  4. Outline • Notation & Definitions • Evolution of SAT Solvers • Overview established approaches • Next challenges in SAT • Other promising approaches • Our proposed approach • Hypothetical reasoning (HR) • The overall approach • Applying reasoning conditions • Relation with other existing techniques • Preliminary experimental results

  5. Notation & Definitions • CNF Formula, clauses, literals: • A CNF formula () is a conjunction of clauses • A clause () is a disjunction of literals • A literal (l) is a propositional variable or its complement • Assignments: • x, 0 denotes the assignment of value 0 to variable x • Can also use x = 0 to denote an assignment j = (a + c)(b + c)(d + c)(¬a + ¬b + ¬c) j = (a c)(b c)(d c)(¬a  ¬b ¬c)

  6. Notation & Definitions (Cont’d) • Unit-clause rule: • If clause  is unit (has a single free literal l), then the free literal l must be satisfied for the clause to be satisfied • Iterated application of the unit-clause rule is referred to as unit propagation (UP) or boolean constraint propagation (BCP) • BCP(x, vx): denotes the set of implied variable assignments obtained by applying BCP as the result of the triggering assignment x, vx • If BCP( x, vx) unsatisfies one or more clauses, then we say that   BCP( x, vx)

  7. Incomplete Complete Can prove unsatisfiability Cannot prove unsatisfiability Continuous formulations Genetic algorithms Simulated annealing Tabu search ... A Taxonomy of SAT Algorithms SAT Algorithms Backtrack search (DPLL) Local search (hill climbing) Resolution (original DP) Stalmarck’s method (SM) Recursive learning (RL) BDDs ...

  8. The Most Effective SAT Solvers • Backtrack search • Boolean constraint propagation • “Reasonable” branching heuristic • Clause recording • Non-chronological backtracking • Search strategies • Restarts • Random backtracking • Efficient data structures • E.g. head/tail lists; watched literals; literal sifting • Examples: BerkMin; Chaff; SATO; rel_sat; GRASP

  9. Backtrack search Unit propagation Chronological backtracking Fine-tuned branching heuristics Probing & reasoning techniques Lookahead (variable probing) Equivalency reasoning Search strategies Restarts Efficient data structures E.g. head/tail lists Examples: EQSATZ Built on top of SATZ Uses equivalency reasoninig RAND-SATZ Built on top of SATZ Branching randomization Search restarts SATZ No search restarts No equivalency reasoning Forms of look-ahead probing Other Effective SAT Solvers

  10. Other Dedicated SAT Solvers • Local search for dedicated classes of instances • Incomplete class of algorithms • Useful if instances known to be satisfiable • Solvers with domain-specific information • Incremental SAT • SAT on Boolean networks • …

  11. Challenging Problem Instances • SAT is being applied in industrial settings • Electronic design automation • Formal verification of hardware/software systems • … • SAT solvers are expected to handle problem instances: • that have hundred thousand / few million variables • that have tens of million clauses • that may be unsatisfiable • SAT solvers must be capable of proving unsatisfiability • completeness is a key issue !

  12. Challenges to SAT Solvers • Dramatic improvements to backtrack search SAT solvers unlikely • Can utilize equivalency reasoning • Hard to interact with clause recording and non-chronological backtracking • Can apply lookahead techniques • Hard to interact with clause recording and non-chronological backtracking • Can devise new search strategies • Search restarts, random backtracking, … ?

  13. Other Approaches • Resolution • Unlikely to be a practical proof procedure • Variable probing (branch-merge rule) • Clause probing (recursive learning) • Not (yet) extensively evaluated • Additional mechanisms for identifying necessary assignments and inferring new clauses • Integrated solution still lacking

  14. Resolution (original DP) • Iteratively apply resolution (consensus) to eliminate one variable each time • i.e., resolution between all pairs of clauses containing x and ¬x • formula satisfiability is preserved • Stop applying resolution when, • Either empty clause is derived  instance is unsatisfiable • Or only clauses satisfied or with pure literals are obtained  instance is satisfiable j = (a + c)(b + c)(d + c)(¬a + ¬b + ¬c) Eliminate variable c 1 = (a + ¬a + ¬b)(b + ¬a + ¬b )(d + ¬a + ¬b ) = (d + ¬a + ¬b ) Instance is SAT !

  15. Recursion can be of arbitrary depth Stalmarck’s Method (SM) in CNF • Recursive application of the branch-merge ruleto each variable with the goal of identifying common conclusions j = (a+ b)(¬a+ c) (¬b + d)(¬c + d) j = (a+ b)(¬a+ c) (¬b+ d)(¬c + d) j = (a + b)(¬a + c) (¬b + d)(¬c + d) j = (a+ b)(¬a+ c) (¬b + d)(¬c+ d) Try a = 0: (a = 0)  (b = 1)  (d = 1) C(a = 0) = {a = 0, b = 1, d = 1} Try a = 1: (a = 1)  (c = 1)  (d = 1) C(a = 1) = {a = 1, c = 1, d = 1} C(a = 0)  C(a = 1) = {d = 1} Any assignment to variable a implies d = 1. Hence, d = 1 is a necessary assignment !

  16. resolution (b + c) resolution (c + d) resolution (d) An Alternative Explanation for SM j = (a + b)(¬a + c) (¬b + d)(¬c + d) Sequence of resolution operations for finding necessary assignments Comment: SM provides a mechanism for identifying suitable resolution operations

  17. Recursion can be of arbitrary depth Recursive Learning (RL) in CNF • Recursive evaluation of clause satisfiability requirements for identifying common assignments  = (a + b)(¬a + d) (¬b + d)  = (a+ b)(¬a+ d) (¬b + d)  = (a + b)(¬a + d) (¬b + d)  = (a + b)(¬a + d) (¬b+ d) Try a = 1: (a = 1)  (d = 1) C(a = 1) = {a = 1, d = 1} Try b = 1: (b = 1)  (d = 1) C(b = 1) = {b = 1, d = 1} Every way of satisfying (a + b) implies d = 1. Hence, d = 1 is a necessary assignment ! C(a = 1)  C(b = 1) = {d = 1}

  18. resolution (b + d) resolution (d) An Alternative Explanation for RL  = (a + b)(¬a + d) (¬b + d) Sequence of resolution operations for finding necessary assignments Comment: RL provides yet another mechanism for identifying suitable resolution operations

  19. SM vs. RL • Both complete procedures for SAT • Stalmarck’s method (in CNF): • hypothetical reasoning based on variables • use variable assignment conditions to probe assignments • variable probing • Recursive learning (in CNF): • hypothetical reasoning based on clauses • use clause satisfiability conditions to probe assignments • clause probing • Both can be viewed as the process of identifying selective resolution operations • Both can be integrated into backtrack search algorithms

  20. Integrate variable probing and clause probing Complete proof procedure for SAT Devise conditions for a priori identification of new clauses That entail most of existing clause inference procedures Evolve from identification of necessary assignments to generalized clause reasoning Applications: Complete proof procedure for SAT Preprocessing engine to existing SAT solvers With polynomial effort Replace unit propagation with HR with backtrack search solvers With polynomial effort Cooperate with backtrack search solvers In parallel solutions for SAT The Objectives of HR

  21. The Organization of HR • Recursive procedure that: • Extends variable probing • To incorporate clause probing • Ensures completeness • Establishes general clause inference conditions • That cover (most) existing clause inference conditions • Readily implements a number of additional techniques • 2-var equivalence; hyper resolution (with binary clauses); equivalency reasoning; binary clause inference conditions; … • Can be integrated into backtrack search

  22. How to Implement HR ? • Independent probing, given conditions on variables and on clauses, may not be practical • O(L2+ L N) = O(L2) at each step • L: number of literals • N: number of variables • Construct assignment & trigger tables, for implementing variable and clause probing • O(L N) at each step • In practice, worst-case complexity is extremely unlikely • OBS: unrestricted clause inference conditions are computationally hard to implement

  23. Assignment Table • Captures the result of applying BCP to each variable assignment • Create a (2N x 2N) matrix: • Each row is associated with an assignment x, vx • 1-valued entries denote assignments y, vy implied by BCP due to trigger assignment x, vx, i.e. BCP(x, vx) • OBS: In practice can use a sparse matrix representation !

  24. Assignment Table (Example) • b, 0 implies (with BCP) the assignments b, 0, c, 0 and d, 0

  25. Trigger Table • Captures which variable assignments directly imply (w/ BCP) each variable assignment • Create a (2N x 2N) matrix: • Each row is associated with an assignment x, vx • 1-valued entries denote assignments y, vy that imply, with BCP, the assignment x, vx • OBS: The trigger table is the transpose of the assignment table ! • Required to create trigger table if using a sparse matrix representation of the assignment table

  26. Trigger Table (Example) • b, 1 is implied (due to BCP) by the assignments a, 0, a, 1 and b, 1

  27. For both assignments to a, a, 0 and a, 1 , we obtain b, 1. b, 1 is a necessary assignment Utilizations of Assignment Tables • Necessary assignments from variable assignment conditions — variable probing

  28. Every assignment that satisfies (b  d), also implies c, 0. c, 0 is a necessary assignment Utilizations of Assignment Tables • Necessary assignments from clause satisfiability conditions — clause probing • Assuming existence of clause (b  d)

  29. One of these assignments must hold (because of a)  create clause (b  c) Utilizations of Assignment Tables • Clause inference from variable assignment conditions

  30. Assume clause  = (a  b  c) exists. Each assignment that satisfies  implies either c, 0 or d, 0  create clause (c  d) Utilizations of Assignment Tables • Clause inference from clause satisfiability conditions

  31. The assignments a, 0 and b, 1 imply the assignments c, 0 and c, 1; are disallowed  create clause (a  b) Utilizations of Trigger Tables • Clause inference from variable assignment conditions

  32. Assume clause  = (c  d) exists. The assignments a, 0 or b, 0 unsatisfy .  create clause (a  b) Utilizations of Trigger Tables • Clause inference from clause unsatisfiability conditions

  33. Reasoning Conditions Summary • Necessary assignments: • From variable assignment conditions (variable probing) • From clause satisfiability conditions (clause probing) • Inferred clauses: • Satisfiability conditions • Variable assignments • Clause satisfiability • Unsatisfiability conditions • Variable assignments • Clause satisfiability

  34. Inference of Clauses • HR reasoning conditions can only infer binary clauses ? • No. Can infer arbitrary clauses ! • Clause satisfiability conditions: • For each clause  = (t1  t2   tm) of formula , all clauses of the form (s1  s2   sm), such that s1, s2,,sm BCP(t1, 1)  …  BCP(tm, 1), are implicates of  • Clearly subsumption can potentially be applied • Clause unsatisfiability conditions: • For each set of assignments A = {t1, 0, t2, 0, , tm, 0 }, such that   BCP(t1, 0)  BCP(t2, 0)    BCP(tm, 0), then clause  = (t1  t2   tm)is an implicate of 

  35. Reasoning Conditions Complexity • O(L N) for constructing assignment & trigger tables and implementing variable and clause probing • Why ? • BCP for filling each row is O(L) • For the 2N rows, construction of table is O(L N) • Each set intersection can trivially be accomplished in O(N)! • All intersections can be done in O((N + L) N) = O(L N) • Corresponding to variable and clause probing • Total time complexity is O(L N) • OBS: In practice worst-case complexity extremely unlikely • OBS: unrestricted clause inference conditions are computationally hard to implement; must use restrictions

  36. O(L N) Can loop O(N) times Polynomial time if depth is constant ! The HR Algorithm • Basic HR algorithm (with depth d, target variables V) • return if depth d  0 • For each variable v in set of target variables V • For each assignment to variable v • L1: Apply unit propagation (BCP) • Apply (tabular) reasoning conditions • Recur HR with depth (d-1) • Re-apply (tabular) reasoning conditions • [Optional] Loop from L1 if more assignments

  37. Relation with Other Techniques • Assignment & Trigger tables naturally capture: • Variable probing (branch-merge rule) • Lookahead techniques • Clause probing (recursive learning) • New clause inference conditions • Assignment & Trigger tables allow capturing: • Failed-literal rule • Two-variable equivalence • Closure of binary clause implication graph • Literal dropping • Equivalency-reasoning / Inference of binary clauses • Hyper resolution (with binary clause inference) • … ?

  38. Relation with Other Techniques • Failed literal rule: • If an assignment x, 0 yields an unsatisfied clause, then x, 1 is a necessary assignment • In the construction of the assignment table, • If   BCP(x, 0), then x, 1 is a necessary assignment

  39. Relation with Other Techniques • 2-variable equivalence: • First form: • If both (x  y) and (y  x) exist in formula, then x  y • Second form: • Utilize binary clause implication graph • Identify strongly connected components (SCCs) • If x, 0 and y, 0 in the same SCC, then x  y • If, from construction of the assignment table, y, 0 BCP(x, 0) and y, 1 BCP(x, 1), then x  y • Captures all SCCs in binary clause implication graph • Can identify additional 2-variable equivalences !

  40. Relation with Other Techniques • Closure of binary clause implication graph: • If l1, 1  l2, 1 and l2, 1  l3, 1, then l1, 1  l3, 1 and infer clause (l1  l3) • From construction of the assignment table, if l2, 1 BCP(l1, 1), then create clause (l1  l2) • Captures the identification of the transitive closure of the implication graph • Can identify additional implications (and clauses) !

  41. Relation with Other Techniques • Literal dropping [Dubois & Dequen, IJCAI’01]: • Given a clause  = (l1  l2   lk), if exists a proper subset of literals {s1, s2, …, sj } of , such that •   BCP(s1, 0  s2, 0    sj, 0), then create a new clause (s1  s2   sj), that subsumes  • Using the assignment table, if exists a set of assignments A = {t1, 0, t2, 0, , tm, 0 }, such that   BCP(t1, 0)  BCP(t2, 0)    BCP(tm, 0), then create the clause: •  = (t1  t2   tm) • Two techniques similar, but not comparable • Literal dropping less general (starts from existing clauses), but more accurate (considers BCP of set of assignments)

  42. Relation with Other Techniques • Equivalency reasoning [Li, AAAI’00]: • Shown to be covered with: • Unit propagation; 2-variable equivalence; conditions for inferring binary clauses • Binary clause inference conditions [MS, CP’00]: • Inference from pattern 2B/1T: • Given (l1  x)  (l2  x)  (l1  l2  y), infer (x  y) • From the assignment table: • If x, 0  y, 1, then infer the clause (x  y)

  43. Relation with Other Techniques • Binary clause inference conditions [MS, CP’00]: • Inference from pattern 0B/4T: • Given (l1  l2  x)  (l1  l2  x)  (l1  l2  y)  (l1  l2  y), infer clause (x  y) • From the assignment table: • Assume l1 = 0 (depth 1) • Can infer (x  y) (depth 2) • From (l2  x)  (l2  y), • Assume l1 = 1 (depth 1) • Can infer (x  y) (depth 2) • From (l2  x)  (l2  y), •  infer (x  y) • But HR with depth 2 required !

  44. Relation with Other Techniques • Hyper resolution (w/ binary clauses) [Bacchus, SAT’02]: • Allows inference of binary clause • Given (l1  x)  (l2  x)    (lk  x)  (l1  l2   lk  y), infer (x  y) • From the assignment table: • If x, 0  y, 1, then infer the clause (x  y)

  45. Preliminary Results • Implemented reasoning conditions on top of JQuest • Assigment tables • Trigger tables • Necessary assignments • Probing due to variables and clauses (binary and ternary) • Clause inference conditions • Simplified version: only binary clauses can be inferred • Results for reasoning conditions on example problem instances • #Vars: number of variables; #Cls: number of clauses • #NA: necessary of assignments; #IBC: inferred binary clauses

  46. Experimental Results

  47. Implementing/Completing HR • Implement (efficient) recursive wrapper • Incrementally define set of variables in recursive step • Reduce significantly the number of row updates in assignment and trigger tables • Instead of BCP-based reasoning conditions, evolve to clause-based reasoning conditions • How to use HR? • Standalone complete proof procedure ? • Integrated within backtrack search SAT solver ? • Hard to interact with clause recording and non-chronological backtracking • Used as a preprocessing engine to backtrack search SAT solvers ?

  48. Conclusions • Proposed the Hypothetical Reasoning algorithm • Integrates variable probing (branch-merge rule) and clause probing (recursive learning) • Implements a number of additional techniques • That allow inferring new clauses • That entail most existing clause inference conditions • That entail a significant number of simplification techniques • Preliminary results for practical problem instances: • By applying reasoning conditions, • a significant number of necessary assignments can be identified and a significant number of new clauses can be inferred

More Related