Exact Inference Algorithms Bucket-elimination

Transcription

1 Exact Inference Algorithms Bucket-elimination COMPSCI 276, Spring 2013 Class 5: Rina Dechter (Reading: class notes chapter 4, Darwiche chapter 6) 1

2 Belief Updating Smoking lung Cancer Bronchitis X-ray Dyspnoea P (lung cancer=yes smoking=no, dyspnoea=yes ) =? 2

3

4 Probabilistic Inference Tasks Belief updating: E is a subset {X1,,Xn}, Y subset X-E, P(Y=y E=e) P(e)? BEL(X i ) = P(Xi = xi evidence) Finding most probable explanation (MPE) x* = argmax P(x,e) Finding maximum a-posteriory hypothesis * * (a1,...,ak) = argmax P(x, e) a X/A Finding maximum-expected-utility (MEU) decision * * (d1,...,dk ) = argmax P(x, e)u(x) d X/D x A X : hypothesis variables D X : decision variables U( x) : utility function 4

5 Belief updating is NP-hard 5

6 A simple network Given: A B C D How can we compute P(D)?, P(D A=0)? P(A D=0)? Brute force O(k^4) Maybe O(4k^2) 6

7

8

9

10 Belief updating: P(X evidence)=? A P(a e=0) P(a,e=0)= B C e = 0, d, c, b P(a)P(b a)p(c a)p(d b,a)p(e b,c)= D E Moral graph P(a) e=0 d P(c a) c b P(b a)p(d b,a)p(e b,c) Variable Elimination h B ( a, d, c, e) 10

11 A B C D E 11

12 A B C D E 12

13 A B C D E Using a different 13

14 14

15

16

17

18

19 A Bayesian network ordering: C,B,E,D,G 20

20 A Bayesian network ordering: C,B,E,D,G 21

21 A different ordering 22

22 A Bayesian network processed along 2 orderings 23

23 A Bucket elimination Algorithm BE-bel (Dechter 1996) b bucket B: bucket C: bucket D: bucket E: bucket A: P(b a) P(d b,a) P(e b,c) P(c a) e=0 P(a) P(a e=0) h B (a,d,c,e) h C (a,d,e) h D (a,e) h E (a) Elimination operator W*=4 induced width (max clique size) D B B C D E A E C 24

24 BE-BEL 25

25 26

26 Student Network example Difficulty Intelligence P(J)? Grade SAT Letter Apply Job

27 E D C B A B C D E A 28

28 Complexity of elimination w * ( d ) O( n exp( w * ( d)) the induced width of moral graph along ordering d The effect of the ordering: A B C E D B C D C E B D E Moral graph A A * * w ( d 1 ) = 4 w ( d 2 ) = 2 31

29 BE-BEL More accurately: O(r exp(w*(d)) where r is the number of cpts. For Bayesian networks r=n. For Markov networks? 32

30 34

31 The impact of observations G G G G B B B B C C C C D D D D F F F F A A A A (a) (b) (c) (d) Ordered graph Induced graph Ordered conditioned graph 35

32 A B C D E Moral graph BE-BEL Use the ancestral graph only 36

33

34

35

36

37 Probabilistic Inference Tasks Belief updating: BEL(X i ) = P(Xi = x i evidence) Finding most probable explanation (MPE) x* = argmax P(x,e) Finding maximum a-posteriory hypothesis * * (a1,...,ak) = argmax P(x, e) a x X/A A X : hypothesis variables 42

38 43 Finding Algorithm BE-mpe max P(x) MPE x = ), ( ), ( ) ( ) ( ) ( max by replaced is,,,, c b e P b a d P a b P a c P a P MPE : b c d e a = max A D E C B

39 Finding x Algorithm elim-mpe (Dechter 1996) MPE = bucket B: bucket C: bucket D: bucket E: bucket A: max b P(b a) P(d b,a) P(e b,c) P(c a) e=0 P(a) MPE MPE = is replaced by max : max P( a) P( c a) P( b a) P( d a, e, d, c, b h B (a,d,c,e) h C (a,d,e) h D (a,e) h E (a) max P(x) a, b) P( e Elimination operator W*=4 induced width (max clique size) D b, c) B B C D E A A E C 44

40 Generating the MPE-tuple 5. b' = arg max P(b a' ) b P(d' b,a' ) P(e' b,c' ) B: P(b a) P(d b,a) P(e b,c) 4. c' = h arg max P(c a' B c (a',d',c,e' ) ) C: P(c a) h B (a,d,c,e) 3. d' = arg max h d C (a',d,e' ) D: h C (a,d,e) 2. e' = 0 E: e=0 h D (a, e) 1. a' = arg max P(a) h a E (a) A: P(a) h E (a) Return (a',b',c',d',e' ) 45

41 47

42 49 Finding MAP Algorithm BE-map ), ( ), ( ) ( ) ( ) ( max, c b e P b a d P a b P a c P a P MPE e,d,b c a = : max and A D E C B

43 50

44 Algorithm BE-MAP Variable ordering: Restricted: Max buckets should Be processed after sum buckets 52

45 BE for Markov networks queries 53

46 O(nexp(w*+1)) and O(n exp(w*)), respectively More accurately: O(r exp(w*(d)) where r is the number of cpts. For Bayesian networks r=n. For Markov networks? 54

47 Finding small induced-width NP-complete A tree has induced-width of? Greedy algorithms: Min width Min induced-width Max-cardinality Fill-in (thought as the best) See anytime min-width (Gogate and Dechter) 56

48 Type of graphs 57

49 The induced width 58

50 Min-width ordering Proposition: algorithm min-width finds a min-width ordering of a graph 59

51 Greedy orderings heuristics Theorem: A graph is a tree iff it has both width and induced-width of 1. 60

52 Different Induced-graphs 61

53 Induced-width for chordal graphs Definition: A graph is chordal if every cycle of length at least 4 has a chord Finding w* over chordal graph is easy using the maxcardinality ordering: order vertices from 1 to n, always assigning the next number to the node connected to a largest set of previously numbered nodes. Lets d be such an ordering A graph along max-cardinality order has no fill-in edges iff it is chordal. On chordal graphs width=induced-width. 62

54 Max-cardinality ordering What is the complexity of min-fill? Min-induced-width? 63

55 K-trees 64

56 Which greedy algorithm is best? 65

57 Recent work in my group Vibhav Gogate and Rina Dechter. "A Complete Anytime Algorithm for Treewidth". In UAI Andrew E. Gelfand, Kalev Kask, and Rina Dechter. "Stopping Rules for Randomized Greedy Triangulation Schemes" in Proceedings of AAAI Potential project 66