Presentation is loading. Please wait.

Presentation is loading. Please wait.

Adnan Darwiche Computer Science Department UCLA

Similar presentations


Presentation on theme: "Adnan Darwiche Computer Science Department UCLA"— Presentation transcript:

1 Adnan Darwiche Computer Science Department UCLA
Searching while Keeping a Trace The Evolution from SAT to Knowledge Compilation Adnan Darwiche Computer Science Department UCLA A. Darwiche

2 Searching while Keeping a Trace The evolution from SAT to Knowledge Compilation
Satisfiability (SAT) Knowledge compilation The connection: The trace of search Implications Open questions A. Darwiche

3 Satisfiability (SAT) Is a set of Boolean constraints satisfiable?
Input to SAT is typically a CNF SAT is mostly solved by DPLL search A & okX => B A & okX => B B & okY => C B & okY => C A. Darwiche

4 Satisfiability (SAT) SAT Solvers: Significant growth in last decade; many solvers publicly available (source code); millions of variables & constraints not uncommon. Applications: Verification, planning, diagnosis, CAD, non-propositional reasoning (e.g., SMTs), … A. Darwiche

5 Knowledge Compilation
A & okX => B A & okX => B B & okY => C B & okY => C Compiled Structure Compiler Evaluator (Polytime) Queries A. Darwiche

6 Knowledge Compilation
A & okX => B A & okX => B B & okY => C B & okY => C ? Compiler Evaluator (Polytime) Queries A. Darwiche

7 Knowledge Compilation
A & okX => B A & okX => B B & okY => C B & okY => C ..... Prime Implicates OBDD Compiler Evaluator (Polytime) Queries A. Darwiche

8 Knowledge Compilation Map
What’s the space of possible target compilation languages? Can it be synthesized in a semantically systematic way? How do the languages compare? Succinctness (relative size) Operations they support in polytime A. Darwiche

9 Applications Diagnosis Planning Probabilistic reasoning
Is this a normal behavior? What are the possible faults? Planning Can this goal be achieved? Generate a set of plans Probabilistic reasoning What is the probability of X given Y Non-monotonic reasoning (penalty logics) Does X follow preferentially from Y Formal verification / CAD: Is it possible that the design will exhibit behavior X? Are two designs equivalent? A. Darwiche

10 Knowledge Compilation Map
For a given application: identify needed operations Choose most succinct language that supports desired operations Compile knowledge base into chosen language A. Darwiche

11 A Knowledge Compilation MAP
Succinctness Polytime Operations Consistency (CO) Validity (VA) Clausal entailment (CE) Sentential entailment (SE) Implicant testing (IP) Equivalence testing (EQ) Model Counting (CT) Model enumeration (ME) Projection (exist. quantification) Conditioning Conjoin, Disjoin, Negate Negation Normal Form or Decomposability Determinism Smoothness Flatness Decision Ordering and and or or or or and and and and and and and and A B  B A C  D D  C A. Darwiche

12 Negation Normal Form A B  B A C  D D  C and or
rooted DAG (Circuit) A. Darwiche

13 Negation Normal Form Decomposability Determinism Smoothness Flatness
and or Decomposability Determinism Smoothness Flatness Decision Ordering A. Darwiche

14 Decomposability or and and or or or or and and and and and and and and
C,D or or or or and and and and and and and and A B  B A C  D D  C A. Darwiche

15 Determinism or and and or or or or and and and and and and and and A
B  B A C  D D  C A. Darwiche

16 Decision or a b X and and X a X b A. Darwiche

17 Binary Decision Diagrams (BDDs)
or and X1  X1 X2  X2 X3  X3 true false X1 X2 X3 1 Decision + decomposability = FBDD A. Darwiche

18 Binary Decision Diagrams (BDDs)
or and X1  X1 X2  X2 X3  X3 true false X1 X2 X3 1 Decision + decomposability + ordering = OBDD A. Darwiche

19 NNF Subsets (2002) NNF d-NNF s-NNF DNNF f-NNF BDD d-DNNF FBDD DNF CNF
CO, CE, ME d-NNF s-NNF DNNF f-NNF VA,IP,CT BDD d-DNNF EQ? FBDD DNF CNF EQ? sd-DNNF CO,CE,ME VA,IP,SE,EQ SE,EQ SE,EQ VA,IP, SE,EQ OBDD MODS IP PI A. Darwiche

20 Space Efficiency (succinctness)
Tractability & Succinctness NNF Tractable Operations OBDD FBDD d-DNNF DNNF decomposability Diagnosis, Non-mon Probabilistic reasoning determinism decision ordering Space Efficiency (succinctness) A. Darwiche

21 Inference by Compiling to d-DNNF
Deterministic Conformant Planning Blai Bonet and Hector Geffner (KR 2006) Probabilistic Conformant Planning Jinbo Huang (AIPS 2006) Model-based diagnosis Paul Elliott and Brian Williams (AAAI 2006) Anthony Barrett (IJCAI 2005) Databases (query re-write) Yolife Arvelo, Blai Bonet and Maria Esther Vidal (AAAI 2006) Inference in Bayesian Networks (2006 competition) Mark Chavira, Adnan Darwiche (IJCAI 2005) Inference in Probabilistic Relational Models Mark Chavira, Adnan Darwiche and Manfred Jaeger (IJAR 2006) c2d compiler: A. Darwiche

22 SAT by DPLL Search Unit resolution Conflict-directed backtracking
x  y ¬x  ¬z ¬w  z  ¬v v  w  z SAT? x  y ¬x  ¬z ¬w  z v = true SAT? x  y ¬x  ¬z w  z v = false SAT? Unit resolution Conflict-directed backtracking Clause learning Branching heuristics Restarts Terminating condition for recursion: empty set (satisfied), or empty clause (contradiction) A. Darwiche

23 Recent Trend: Exhaustive DPLL
Count number of models: Model counters, e.g., relsat, cachet Generate all/subset of models: Image computation in model checking SMTs (non-propositional reasoning) Variations on DPLL Search A. Darwiche

24 The Language of Search Exhaustive DPLL Knowledge Compiler Record Trace
Variations Languages A. Darwiche

25 Trace of DPLL X  Y X  Y  Z X  Y  Z X Y 1 Z 1 unsat unsat sat
Y 1 Z unsat 1 unsat sat A. Darwiche

26 Exhaustive DPLL Run to Exhaustion X  Y X  Y  Z X  Y  Z X 1 Y
1 Y Y 1 1 Z Z unsat sat 1 1 unsat sat unsat sat A. Darwiche

27 Trace of DPLL: a Formula or X and and or X X or Y Y and and and and Z
unsat sat Y Y Z Z Y Y 1 unsat sat unsat sat A. Darwiche

28 Trace of DPLL: a Formula
Equivalent to original CNF Tractable (e.g., count models) and and or X X or and and and and Y Y Z Z Y Y 1 A. Darwiche

29 Dealing with Redundancy
Level One: Do not record redundant portions of trace Level Two: Try not to solve equivalent subproblems X Y Y Z Z unsat sat unsat sat unsat sat A. Darwiche

30 Dealing with Redundancy
X Y Y Z Z unsat unsat unsat sat A. Darwiche

31 Dealing with Redundancy
Simply point to existing node Do not create X Y Y Z Z unsat sat A. Darwiche

32 This is an OBDD! X Y Y Z 1 A. Darwiche

33 NNF + decision, decomposability, ordering
This is an OBDD! or X and and Y Y or X X or and and and and Z Z 1 Y Y 1 NNF + decision, decomposability, ordering A. Darwiche

34 A Non-traditional OBDD Compiler
X Exhaustive DPLL, Fixed variable order, Unique nodes X  Y X  Y  Z X  Y  Z Compile Y Y Z 1 New complexity guarantees A. Darwiche

35 FBDD X X  Y X  Y  Z X  Y  Z Y Z Compile Z Y 1
Exhaustive DPLL, Dynamic variable order, Unique nodes X  Y X  Y  Z X  Y  Z Y Z Compile Z Y 1 NNF + decision, decomposability A. Darwiche

36 FBDD vs OBDD FBDD more succinct than OBDD (dynamic var ordering in sat) Top-down vs bottom-up algorithms OBDD: equivalence test (canonical) FBDD: probabilistic equivalence test Both allow model counting A. Darwiche

37 Dealing with Redundancy
Level One: Unique nodes (done) Level Two: Avoid redundant compilation (searches) A. Darwiche

38 Redundant Compilation
x5  x6 x4  x5  x6 x1  x3  x4  x5 x2  x3 x1  x2  x3 X1 1 X2 X2 1 X3 X3 1 1 x5  x6 x4  x5  x6 x5  x6 x4  x5  x6 Formula Caching: complexity guarantees A. Darwiche

39 Formula Caching Majercik and Litmman, 1998 Darwiche, 2002
Bacchus et al, 2003, 2004 Huang, 2004 Sang, Kautz, Beam, 2004, 2005 Thurley, 2006 A. Darwiche

40 Beyond BDDs… Plain DPLL FBDD Fixed Variable Ordering OBDD A. Darwiche

41 Decomposition (Component Analysis)
Solve disjoint subproblems independently d-DNNF Combine as AND node A. Darwiche

42 and and A B D B D C E 1 Deterministic Decomposable Negation Norm Form
(d-DNNF) A  B  C A  B  C A  D  E A  D  E A B  C D  E and and B  C D  E B D B D C E 1 A. Darwiche

43 or and and and A A and or or or or and and and and B D B D C E
Deterministic Decomposable Negation Norm Form (d-DNNF) or and and and A A and or or or or and and and and B D B D C E A. Darwiche

44 FBDD vs d-DNNF d-DNNF more succinct than FBDD (effectiveness of decomposition) Deterministic equivalence test open Probabilistic equivalence test apply Other queries same… A. Darwiche

45 The Language of Search Fixed Variable Ordering OBDD Plain DPLL FBDD
Allowing Decomposition d-DNNF Other languages: deterministic DNF A. Darwiche

46 Relation to AND/OR Search (CP)
AND/OR graphs are deterministic and decomposable AND/OR search algorithms are doing enough work to compile networks into (multi-valued equivalent of) d-DNNF Capable of more than answering a single query (model counting, belief revision, etc) A. Darwiche

47 Implications SAT techniques harnessed for knowledge compilation
c2d compiler based on Rsat Solver (SAT-race 06) Language properties (succinctness/tractability) help characterize power and limitations of search A. Darwiche

48 Understanding DPLL Take any program X that runs exhaustive DPLL-style search Examine traces, if traces  L, then X can answer all queries tractable for L X is hopeless on any input having no polynomial-size representation in L A. Darwiche

49 Power of DPLL Traces of several model counters (Relsat, Cachet, e.g.) are in d-DNNF Are doing enough work to compile formulas into d-DNNF solve tasks beyond model counting (e.g., minimum cardinality, probabilistic equivalent testing) A. Darwiche

50 Limitation of DPLL: General determinism
or Decision nodes (d-DNNF’) and X X or Deterministic nodes (d-DNNF) A B C and or A B A B A. Darwiche

51 Beyond DPLL: Decomposability (D) without determinism (d)
or DNNF: CO, CE, ME, exist quant and X3 or and X1 X2 A. Darwiche

52 Summary Overview of recent results in knowledge compilation
Overview of recent trends in exhaustive DPLL search A connection between SAT and knowledge compilation (search trace): SAT techniques harnessed for compilation into various languages Language properties (succinctness, tractability) characterize power and limitations of search algorithms A. Darwiche

53 A. Darwiche

54 Knowledge Compilation Map, with Pierre Marquis DPLL with a Trace,
Language of Search, with Jinbo Huang A. Darwiche

55

56 Multi-Linear Functions  Arithmetic Circuits
* + Factoring A B A. Darwiche

57 Propositional theory: Multi-linear function:
MLFsACs CNFsd-DNNF Propositional theory: c ^ (a   b) a c + a b c + c Multi-linear function: Encode Compile * + c c Decode * * + b b b 1 a a a 1 Smooth d-DNNF Arithmetic Circuit A. Darwiche


Download ppt "Adnan Darwiche Computer Science Department UCLA"

Similar presentations


Ads by Google