1. 1 Directional consistency Chapter 4 ICS-275 Spring 2007

2. ICS 275 - Constraint Networks 2 Backtrack-free search: or What level of consistency will guarantee global- consistency

3. Spring 2007 ICS 275 - Constraint Networks 3 Directional arc-consistency: another restriction on propagation D4={white,blue,black} D3={red,white,blue} D2={green,white,black} D1={red,white,black} X1=x2, x1=x3,x3=x4

4. Spring 2007 ICS 275 - Constraint Networks 4 Directional arc-consistency: another restriction on propagation D4={white,blue,black} D3={red,white,blue} D2={green,white,black} D1={red,white,black} X1=x2, x1=x3,x3=x4

5. Spring 2007 ICS 275 - Constraint Networks 5 Directional arc-consistency: another restriction on propagation D4={white,blue,black} D3={red,white,blue} D2={green,white,black} D1={red,white,black} X1=x2, x1=x3, x3=x4 After DAC: D1= {white}, D2={green,white,black}, D3={white,blue}, D4={white,blue,black}

6. Spring 2007 ICS 275 - Constraint Networks 6 Algorithm for directional arc- consistency (DAC) Complexity :

7. Spring 2007 ICS 275 - Constraint Networks 7 Example of DAC

8. Spring 2007 ICS 275 - Constraint Networks 8 Is DAC sufficient?

9. Spring 2007 ICS 275 - Constraint Networks 9 Directional arc-consistency may not be enough  Directional path-consistency

10. Spring 2007 ICS 275 - Constraint Networks 10 Directional path-consistency

11. Spring 2007 ICS 275 - Constraint Networks 11 Algorithm directional path consistency (DPC)

12. Spring 2007 ICS 275 - Constraint Networks 12 Example of DPC

13. Spring 2007 ICS 275 - Constraint Networks 13 Induced-width

14. Spring 2007 ICS 275 - Constraint Networks 14 Directional i-consistency

15. Spring 2007 ICS 275 - Constraint Networks 15 Example

16. Spring 2007 ICS 275 - Constraint Networks 16 Algorithm directional i- consistency

17. Spring 2007 ICS 275 - Constraint Networks 17 The induced-width DPC recursively connects parents in the ordered graph, yielding: Width along ordering d, w(d): max # of previous parents Induced width w*(d): The width in the ordered induced graph Induced-width w*: Smallest induced-width over all orderings Finding w* NP-complete (Arnborg, 1985) but greedy heuristics (min-fill).

18. Spring 2007 ICS 275 - Constraint Networks 18 Induced-width

19. Spring 2007 ICS 275 - Constraint Networks 19 Examples Three orderings : d 1 = (F,E,D,C,B,A), its reversed ordering d 2 = (A,B,C,D,E, F), and d 3 = (F,D,C,B,A,E). Orderings from bottom to top, The parents of A along d 1 are {B,C,E}. Width of A along d 1 is 3, of C along d 1 is 1, of A along d 3 is 2. w(d 1 ) = 3, w(d 2 ) = 2, and w(d 3 ) = 2. The width of graph G is 2.

20. Spring 2007 ICS 275 - Constraint Networks 20 Induced-width and DPC The induced graph of (G,d) is denoted (G*,d) The induced graph (G*,d) contains the graph generated by DPC along d, and the graph generated by directional i- consistency along d

21. Spring 2007 ICS 275 - Constraint Networks 21 Refined Complexity using induced-width Consequently we wish to have ordering with minimal induced-width Induced-width is equal to tree-width to be defined later. Finding min induced-width ordering is NP-complete

22. Spring 2007 ICS 275 - Constraint Networks 22 Greedy algorithms for iduced- width Min-width ordering Max-cardinality ordering Min-fill ordering Chordal graphs

23. Spring 2007 ICS 275 - Constraint Networks 23 Min-width ordering

24. Spring 2007 ICS 275 - Constraint Networks 24 Min-induced-width

25. Spring 2007 ICS 275 - Constraint Networks 25 Min-fill algorithm Prefers a node who add the least number of fill-in arcs. Empirically, fill-in is the best among the greedy algorithms (MW,MIW,MF,MC)

26. Spring 2007 ICS 275 - Constraint Networks 26 Cordal graphs and Max- cardinality ordering A graph is cordal if every cycle of length at least 4 has a chord Finding w* over chordal graph is easy using the max-cardinality ordering If G* is an induced graph it is chordal K-trees are special chordal graphs. Finding the max-clique in chordal graphs is easy (just enumerate all cliques in a max- cardinality ordering

27. Spring 2007 ICS 275 - Constraint Networks 27 Example We see again that G in Figure 4.1(a) is not chordal since the parents of A are not connected in the max-cardinality ordering in Figure 4.1(d). If we connect B and C, the resulting induced graph is chordal.

28. Spring 2007 ICS 275 - Constraint Networks 28 Max-caedinality ordering Figure 4.5 The max-cardinality (MC) ordering procedure.

29. Spring 2007 ICS 275 - Constraint Networks 29 Width vs local consistency: solving trees

30. Spring 2007 ICS 275 - Constraint Networks 30 Tree-solving

31. Spring 2007 ICS 275 - Constraint Networks 31 Width-2 and DPC

32. Spring 2007 ICS 275 - Constraint Networks 32 Width vs directional consistency (Freuder 82)

33. Spring 2007 ICS 275 - Constraint Networks 33 Width vs i-consistency DAC and width-1 DPC and width-2 DIC_i and with-(i-1)  backtrack-free representation If a problem has width 2, will DPC make it backtrack-free? Adaptive-consistency: applies i-consistency when i is adapted to the number of parents

34. Spring 2007 ICS 275 - Constraint Networks 34 Special case of induced-width 2

35. Spring 2007 ICS 275 - Constraint Networks 35 Adaptive-consistency

36. Spring 2007 ICS 275 - Constraint Networks 36 Bucket E: E  D, E  C Bucket D: D  A Bucket C: C  B Bucket B: B  A Bucket A: A  C contradiction = D = C B = A Bucket Elimination Adaptive Consistency (Dechter & Pearl, 1987) = 

37. Spring 2007 ICS 275 - Constraint Networks 37 A E D C B E A D C B || R D BE, || R E || R DB || R DCB || R ACB || R AB RARA R C BE Bucket Elimination Adaptive Consistency (Dechter & Pearl, 1987)

38. Spring 2007 ICS 275 - Constraint Networks 38 The Idea of Elimination 3 value assignment D B C R DBC eliminating E

39. Spring 2007 ICS 275 - Constraint Networks 39 Adaptive- consistency, bucket-elimination

40. Spring 2007 ICS 275 - Constraint Networks 40 Properties of bucket-elimination (adaptive consistency) Adaptive consistency generates a constraint network that is backtrack-free (can be solved without dead- ends). The time and space complexity of adaptive consistency along ordering d is respectively, or O(r k^(w*+1)) when r is the number of constraints. Therefore, problems having bounded induced width are tractable (solved in polynomial time) Special cases: trees ( w*=1 ), series-parallel networks (w*=2 ), and in general k-trees ( w*=k ).

41. Spring 2007 ICS 275 - Constraint Networks 41 Back to Induced width Finding minimum-w* ordering is NP-complete (Arnborg, 1985) Greedy ordering heuristics: min-width, min-degree, max-cardinality (Bertele and Briochi, 1972; Freuder 1982), Min-fill.

42. Spring 2007 ICS 275 - Constraint Networks 42 Solving Trees (Mackworth and Freuder, 1985) Adaptive consistency is linear for trees and equivalent to enforcing directional arc-consistency (recording only unary constraints)

43. Spring 2007 ICS 275 - Constraint Networks 43 Summary: directional i-consistency A E CD B D C B E D C B E D C B E Adaptived-arcd-path

44. Spring 2007 ICS 275 - Constraint Networks 44 Variable Elimination Eliminate variables one by one: “constraint propagation” Solution generation after elimination is backtrack-free

45. Spring 2007 ICS 275 - Constraint Networks 45 Relational consistency ( Chapter 8 ) Relational arc-consistency Relational path-consistency Relational m-consistency Relational consistency for Boolean and linear constraints: Unit-resolution is relational-arc-consistency Pair-wise resolution is relational path- consistency

46. Spring 2007 ICS 275 - Constraint Networks 46 Sudoku’s propagation http://www.websudoku.com/ What kind of propagation we do?

47. Spring 2007 ICS 275 - Constraint Networks 47 Complexity of Adaptive-consistency as a function of the hypergraph Processing a bucket is also exponential in its number of constraints. The number of constraints in a bucket is bounded by the total, r, number of functions. Can we have a better bound? Theorem: If we combine buckets into superbuckets so variables in a superbucket are covered by original constraints scopes, then “super-bucket elimination is exp in the max number of constraints in a superbucket. Hyper-induced-width: The number of constraint in a super-bucket

48. Spring 2007 ICS 275 - Constraint Networks 48 Example of CSP: Find Possible Genotypes Possible haplotypes : h 1 =A,A 1 ; h 2 = A,A 2 ; h 3 =O,A 1 ; h 4 = O,A 2 ; Possible genotypes (ordered): g ij = h i /h j (father/mother) Ind #1: g 11, g 13, g 31 Ind #2: g 44 Ind #3: g 12, g 21, g 14, g 41, g 23, g 32 Ind #4: g 22, g 24, g 42 Ind #5: g 34, g 43 12 34 5 A A 1 /A 1 O A 2 /A 2 A A 1 /A 2 A A 2 /A 2 O A 1 /A 2 A pedigree typed at the ABO and AK1 loci : Constraints: R 13 = {, }, similarly for R 23, R 35, and R 45.

49. Spring 2007 ICS 275 - Constraint Networks 49 Bucket Elimination for CSP Given an order X 1,…, X n : Distribute constraints into buckets, according to the highest variable in the constraint’s scope. Process the buckets one by one starting from X n : Join the constraints (with equal value for the eliminated variable), in the eliminated bucket and remember the value of the eliminated variable. Put the new constraint in the bucket of the highest variable in its scope.

50. Spring 2007 ICS 275 - Constraint Networks 50 Example of CSP: Find Possible Genotypes (cont. 1) Constraints : R AC = {, }; R BC = { }; R CE = {<g {14,41,23,32} g {34,43} }; R DE = { } AB C E D Constraint graph

51. Spring 2007 ICS 275 - Constraint Networks 51 Find Possible Genotypes (cont. 2) Suppose an order: A, B, C, D, E. The distribution into buckets is as follows: BABAB BCBC BDBD BEBE Initial R AC = {, } R BC = { } R CE = { } R DE = { }

52. Spring 2007 ICS 275 - Constraint Networks 52 Find Possible Genotypes (cont. 3) BABAB BCBC BDBD BEBE Initial R AC = {, } R BC = { } R CE = { } R DE = { } When processing B E : compute R CD = { }. put R CD into bucket B D. remember the value g 34 for E.

53. Spring 2007 ICS 275 - Constraint Networks 53 Find Possible Genotypes (cont. 4) BABAB BCBC BDBD BEBE After Processing B E. R AC = {, } R BC = { } R CD = { } When processing B D : compute R C = {<g {23,32,14,41 } }. put R C into bucket B C. remember the values g 24,g 42 for D.

54. Spring 2007 ICS 275 - Constraint Networks 54 Find Possible Genotypes (cont. 5) BABAB BCBC BDBD BEBE After Processing B D. R AC = {, } R BC = { } R C = {<g {23,32,14,41} }. When processing B C : compute R AB = { }. put R C into bucket B B. remember the value g 14 for C.

55. Spring 2007 ICS 275 - Constraint Networks 55 Find Possible Genotypes (cont. 6) BABAB BCBC BDBD BEBE After Processing B D. R AB = { }. When processing B B : compute R A = { }. remember the value g 44 for B.

56. Spring 2007 ICS 275 - Constraint Networks 56 Find Possible Genotypes (cont. 7) By backtracking we get the possible values: Ind #1 (A): g 11, g 13, g 31 Ind #2 (B): g 44 Ind #3 (C): g 14 Ind #4 (D): g 24, g 42 Ind #5 (E): g 34 12 34 5 A A 1 /A 1 O A 2 /A 2 A A 1 /A 2 A A 2 /A 2 O A 1 /A 2 12 34 5 A|A or A|O or O|A A 1 |A 1 O|O A 2 |A 2 A|O A 1 |A 2 A|O or O|A A 2 /A 2 O|O A 1 |A 2

57. Spring 2007 ICS 275 - Constraint Networks 57 Probability of data in one locus given an inheritance vector S 23m L 21f L 21m L 23m X 21 S 23f L 22f L 22m L 23f X 22 X 23 Model for locus 2 P(x 21, x 22, x 23 |s 23m,s 23f ) = =  P(l 21m ) P(l 21f ) P(l 22m ) P(l 22f ) P(x 21 | l 21m, l 21f ) P(x 22 | l 22m, l 22f ) P(x 23 | l 23m, l 23f ) P(l 23m | l 21m, l 21f, S 23m ) P(l 23f | l 22m, l 22f, S 23f ) l 21m,l 21f,l 22m,l 22f l 22m,l 22f The five last terms are always zero-or-one, namely, indicator functions.

58. Spring 2007 ICS 275 - Constraint Networks 58 Efficient computation S 23m L 21f L 21m L 23m X 21 S 23f L 22f L 22m L 23f X 22 X 23 Model for locus 2 Assume only individual 3 is genotyped. For the inheritance vector (0,1), the founder alleles L 21m and L 22f are not restricted by the data while (L 21f,L 22m ) have two possible joint assignments (A 1,A 2 ) or (A 2,A 1 ) only: The five last terms are always zero-or-one, namely, indicator functions. ={A 1,A 2 } =1 =0 p(x 21, x 22, x 23 |s 23m =1,s 23f =0 ) = p( A 1 )p( A 2 ) + p( A 2 )p( A 1 ) In general. Every inheritance vector defines a subgraph of the Bayesian network above. We build a founder graph

59. Spring 2007 ICS 275 - Constraint Networks 59 Efficient computation S 23m L 21f L 21m L 23m X 21 S 23f L 22f L 22m L 23f X 22 X 23 Model for locus 2 The five last terms are always zero-or-one, namely, indicator functions. ={A 1,A 2 } =1 =0 In general. Every inheritance vector defines a subgraph as indicated by the black lines above. Construct a founder graph whose vertices are the founder variables and where there is an edge between two vertices if they have a common typed descendent. The label of an edge is the constraint dictated by the common typed descendent. Now find all consistent assignments for every connected component. L 21f L 21m L 22f L 22m {A 1,A 2 }

60. Spring 2007 ICS 275 - Constraint Networks 60 A Larger Example 4 3 6 5 2 1 8 7 a,b a,c b,d 5364 2187 {a,b} 5364 {b,d} {a,c} Descent graph Founder graph (An example of a constraint satisfaction graph) Connect two nodes if they have a common typed descendant. The constraint {a,b} means the relation {(a,b)(b,a)}

61. Spring 2007 ICS 275 - Constraint Networks 61 The Constraint Satisfaction Problem 5364 2187 {a,b} 5364 {b,d} {a,c} The number of possible consistent alleles per non-isolated node is 0, 1 or 2. For example node 2 has all possible alleles, node 6 can only be b because its domain must be {a,b} and {b,d}. and node 3 can be assigned either a or b. namely, the intersection of its adjacent edges labels. For each non-singleton connected component: Start with an arbitrary node, pick one of its values. This dictates all other values in the component. Repeat with the other value if it has one. So each non-singleton component yields at most two solutions. What is the special constriant problem here?

62. Spring 2007 ICS 275 - Constraint Networks 62 Solution of the CSP 5364 2187 {a,b} 5364 {b,d} {a,c} Since each non-singleton component yields at most two solutions. The likelihood is simply the product of sums each of two terms at most. Each component contributes one term. Singleton components contribute the term 1 In our example: 1 * [ p(a)p(b)p(a) + p(b)p(a)p(b)] * p(d)p(b)p(a)p(c). Complexity. Building the founder graph: O(f 2 +n). While solving general CSPs is NP-hard. This is graph coloring where domains are often size 2.