Graph Coloring Facets from a Constraint Programming Formulation

Transcription

1 Graph Coloring Facets from a Constraint Programming Formulation David Bergman J. N. Hooker Carnegie Mellon University INFORMS 2011

2 Motivation 0-1 variables often encode choices that can be represented with finite domain variables. x i = finite domain variable Job assigned to worker i Start time of job i City visited after city i Number of packages on truck i y ij = corresponding 0-1 variable y ij = 1 if x i = j

3 Motivation A constraint programming formulation often uses finite-domain variables. If the variables are numeric, the problem has polyhedral structure very different from the 0-1 problem. Finite-domain cuts can be mapped into the 0-1 model. This may yield stronger cuts in the 0-1 model.

4 Motivation A constraint programming formulation often uses finite-domain variables. If the variables are numeric, the problem has polyhedral structure very different from the 0-1 problem. Finite-domain cuts can be mapped into the 0-1 model. This may yield stronger cuts in the 0-1 model. We apply this idea to graph coloring. May apply to other problems with both 0-1 and CP formulations.

5 Motivation We obtain two kinds of results: If you find a structure (e.g., odd hole) that yields a known valid inequality in 0-1 space We will give you a stronger cut for free. Use whatever separation algorithm you want.

6 Motivation We obtain two kinds of results: If you find a structure (e.g., odd hole) that yields a known valid inequality in 0-1 space We will give you a stronger cut for free. Use whatever separation algorithm you want. We identify additional structures that yield valid inequalities. They are stronger than known cuts. We have separation algorithms.

7 Graph Coloring We focus on the vertex coloring problem. Given a graph, assign colors to vertices so that no two adjacent vertices receive the same color. Minimize the number of colors

8 Graph Coloring = 1 if color j is used 0-1 model min j ij j y + y w, all colors j 1j 2 j j y + y w, all colors j 1j 5 j j y + y w, all colors j 2 j 3 j j y + y + y w, all colors j y 3 j 4 j 5 j j ij y w { 0,1} j = 1, all vertices i = 1 if vertex i receives color j

9 Graph Coloring General model: min j i V y ij k y ij j w j = 1, all vertices y w, all colors j, cliques V that cover vertices ij j k { 0,1} = 1 if vertex i receives color j = 1 if color j is used i O(n 2 ) variables O(n 3 ) constraints

10 Alldiff Systems Use an all-different constraint for each clique min z x, all vertices i i ( x1 x2 ) ( x1 x5 ) ( 2 3 ) ( ) { 1,...,5 } alldiff,, all colors alldiff,, all colors alldiff x, x, all colors j alldiff x, x, x, all colors j x i z = color assigned to vertex i j j

11 Alldiff Systems General model: min z z x, all vertices i i ( ) { 1,, n} alldiff x i V, all cliques V x i i k k = color assigned to vertex i O(n) variables O(n 2 ) constraints Objective reduces symmetry

12 Alldiff Systems Applications: Scheduling, timetabling. Employee scheduling. Course timetabling. Latin squares. Alldiff for each row, column. Experimental design: orthogonal Latin squares. Sudoku puzzles. Graph coloring. Many applications.

13 Related Work Convex hull of single alldiff. Hooker (2000), Williams and Yan (2001). Convex hull of 2 alldiffs. Appa, Magos and Mourtos (2004) Convex hull of alldiff systems with inclusion property. Appa, Magos and Mourtos (2011). Same facets as individual alldiffs. Some facets of systems without inclusion property. Magos and Mourtos (2011).

14 Variable Mapping There is a linear mapping from x i to y ij : x i = Any valid linear inequality in x i -space maps to a valid linear inequality in y ij -space. Just substitute above expression for x i. j Convert any finite domain cut to a 0-1 cut. jy ij

15 Variable Mapping There is a linear mapping from x i to y ij : x i = Any valid linear inequality in x i -space maps to a valid linear inequality in y ij -space. Just substitute above expression for x i. j Convert any finite domain cut to a 0-1 cut. jy Objective function more likely to be linear in y-space. ij For coloring, it is linear in both x-space and y-space.

16 Choice of Domain We will assume each x i has domain {0,, n 1}. To simplify exposition. Most results to follow can be generalized to an arbitrary numeric domain {v 1,, v n } with each v i 0. Some results are valid for domain D = {0,δ,, (n 1)δ} with δ > 0.

17 Odd Cycles A q-cycle consists of q alldiff constraints that look like this: alldiff x 1 x 2 x 11 alldiff x 17 x 10 x 9 x 3 x 4 x 12 alldiff x 16 x 8 x 13 alldiff x 15 x 7 x 14 alldiff x 6 x 5

18 Odd Cycles Select any subset of s vertices in each overlap: s = 2 S 1 x 1 x 2 x 11 S 5 x 17 x 10 x 9 x 3 x 4 x 12 x 16 S x 2 13 x 8 x 15 x 7 S 4 x 14 x 6 x 5 S 3

19 Odd Cycles Focus on the sets S i : S 1 s = 2 q = 5 S 5 x 10 x 9 x 1 x 2 x 3 x 4 S 2 S 4 x 8 x 7 sq = 10 vertices Each color can be assigned to at most (q 1)/2 = 2 vertices. x 6 x 5 S 3 sq We need at least L = = 5 colors ( q 1) / 2

20 So i S Focus on the sets S i : x i s = 2 q = 5 q 1 q sq = 10 vertices Each color can be assigned to at most (q 1)/2 = 2 vertices. S = k S Odd Cycles q 1 q 1 + ( L 2) + sq ( L 1) ( L 1) 2 2 k S 5 x 10 x 9 S 4 x 8 x 7 x 1 S 1 x 2 x 6 x 5 x 3 x 4 S 3 sq We need at least L = = 5 colors ( q 1) / 2 S 2

21 Odd Cycles Focus on the sets S i : S 1 So i S x i s = 2 q = 5 S = q 1 q k S q 1 q 1 + ( L 2) + sq ( L 1) ( L 1) 2 2 q 1 = sq L ( L 1) = 20 4 k S 5 x 10 x 9 S 4 x 8 x 7 x 1 x 2 x 6 x 5 x 3 x 4 S 3 S 2

22 Odd Cycles So we have a valid inequality: i S q 1 xi sq L ( L 1) = x 10 x 9 0 x 1 1 x 2 x 3 x x 8 x 7 x 6 x 5 0 4

23 Odd Cycles So we have a valid inequality: i S q 1 xi sq L ( L 1) = 20 4 The inequality is facetdefining if q is odd. and if the q-cycle is the subgraph induced by vertices in the cycle. 4 3 x 10 x x 8 x 7 0 x 1 1 x 2 x 3 x 4 x x

24 Odd Cycles So we have a valid inequality: i S q 1 xi sq L ( L 1) = 20 4 The inequality is facetdefining if q is odd. and if the q-cycle is the subgraph induced by vertices in the cycle. For s = 1, we have odd hole cut: q + 3 xi 2 i S 4 3 x 10 x x 8 x 7 0 x 1 1 x 2 x 3 x 4 x x

25 Odd Cycles So we have a valid inequality: i S q 1 xi sq L ( L 1) = 20 4 We can obtain a valid bound on number of colors z by substituting z x i for x i : 4 3 x 10 x 9 x 8 x 7 0 x 1 1 x 2 x 3 x q 1 x x 5 z x 1 ( 1) 6 i + L L 2 qs 4 4 i S qs = xi This is facet defining for domain D. i S

26 z-cuts in general In fact, facet-defining x-cuts for a graph coloring problem always give rise to facet-defining z-cuts: Theorem: if ax b is facet defining for a coloring problem with domain D = {0, δ, 2δ,, (n 1)δ} for δ > 0, then aez ax + b is also facet defining, where e = (1,, 1).

27 Mapping into 0-1 Space The x-cut i S q 1 xi sq L ( L 1) = 20 4 and the z-cut 1 q 1 z xi + 1 L ( L 1) qs i S 4qs 4 3 x 10 x 9 0 x 1 1 x 2 x 3 x map into a 0-1 cut by replacing x i with jy How do they compare with classical odd hole cuts? j ij 2 1 x 8 x 7 x x

28 Comparison with Odd Hole Cuts A q-cycle defines s q = 32 odd hole cuts for each color: q 1 yij w j, all T, j 2 i T where T selects one vertex from each S k S 5 x 10 x 9 x 1 S 1 x 2 x 3 x 4 S 2 S 4 x 8 x 7 x 6 x 5 S 3

29 Comparison with Odd Hole Cuts A q-cycle defines s q = 32 odd hole cuts for each color: q 1 yij w j, all T, j 2 i T where T selects one vertex from each S k For s 2, one x-cut is stronger than all of these odd hole cuts. S 5 x 10 x 9 S 4 For s = 1, the finite domain cut is redundant of odd cycle cuts and other 0-1 constraints. x 8 x 7 x 1 S 1 x 2 x 6 x 5 x 3 x 4 S 3 S 2

30 Comparison with Odd Hole Cuts A q-cycle defines s q = 32 odd hole cuts for each color: q 1 yij w j, all T, j 2 i T where T selects one vertex from each S k For s 2, one x-cut is stronger than all of these odd hole cuts. Adding a z-cut to the x-cut tightens the bound further. S 5 x 10 x 9 S 4 x 8 x 7 x 1 S 1 x 2 x 6 x 5 x 3 x 4 S 3 S 2

31 Comparison with Odd Hole Cuts A q-cycle defines s q = 32 odd hole cuts for each color: q 1 yij w j, all T, j 2 i T where T selects one vertex from each S k For any s (including s = 1), one x-cut and one z-cut are stronger than all of these odd hole cuts. S 5 x 10 x 9 S 4 x 8 x 7 x 1 S 1 x 2 x 6 x 5 x 3 x 4 S 3 S 2

32 Comparison with Odd Hole Cuts A q-cycle defines s q = 32 odd hole cuts for each color: q 1 yij w j, all T, j 2 i T where T selects one vertex from each S k For any s (including s = 1), one x-cut and one z-cut are stronger than all of these odd hole cuts. x 10 x 9 x 8 x 7 For any separating odd hole cut, replace it with x-cut and z-cut for s = 1 to get a stronger cut. S 5 S 4 x 1 S 1 x 2 x 6 x 5 x 3 x 4 S 3 S 2

33 Computed Bounds Lower bound on number of colors in 0-1 model of 5-cycle s = All odd hole cuts x-cut only z-cut only x and z-cut only

34 Separation Heuristic Select subset of s vertices in each overlap with smallest values in current relaxation: s = 2 S 1 x 1 x 2 x 11 S 5 x 17 x 10 x 9 x 3 x 4 x 12 x 16 S x 2 13 x 8 x 15 x 7 S 4 x 14 x 6 x 5 S 3

35 A q-path looks like Odd Paths alldiff alldiff alldiff alldiff alldiff x a x 1 x 2 x 3 x 4 x b q = 5 x 5 x 6 x 7 x 7

36 Select q + 1 variables: Odd Paths x a x 1 x 2 x 3 x 4 x b q = 5 x 5 x 6 x 7 x 7

37 Odd Paths This yields a valid inequality (x-cut) x a x 1 x 2 x 3 x 4 x b q = q 1 q + 3 2( x + x ) + x = 4 2 a b i i = 1

38 Odd Paths This yields a valid inequality (x-cut): x a x 1 x 2 x 3 x 4 x b q = q 1 q + 3 2( xa + xb ) + xi = 4 i = 1 2 The inequality is facet-defining if q is odd. and if the q-path is the subgraph induced by vertices in the cycle.

39 We also have a z-cut Odd Paths x a x 1 x 2 x 3 x 4 x b q = z x + x + x + q + q ( ) 3 a b i i = 1 2 This is also facet defining.

40 Mapping into 0-1 Space When mapped into 0-1 space, the finite domain cuts are redundant of the 0-1 model. Because the 0-1 model is already totally unimodular. This is also facet defining.

41 Mapping into 0-1 Space When mapped into 0-1 space, the finite domain cuts are redundant of the 0-1 model. Because the 0-1 model is already totally unimodular. However, the finite domain cuts provide a compact relaxation. For each q-path, replace q clique constraints with one x-cut and one y-cut. Gives the same bound in a problem consisting of one path.

42 Clutters A q-clutter looks something like V 2 S 3 S 1 T 2 T 3 U S 2 V 3 S = V \ V k l k T = V \ V k U = k k V k l l k k l T 1 q = 3 V 1

43 Clutters Facet-defining inequality. Let S = S T T u = U k k = k k V 2 q = 3 S 1 T 2 T 3 U S 3 T 1 V 1 S 2 V 3 q( q 1) qs + u xi + xi b 2 ( ) A valid inequality is: i T i S U where 1 b = q q 2 qs + u qs + u + ( 1)( )( 1) Properties of 0-1 mapping?

44 Another Interpretation Combine alternate formulations to obtain a better relaxation. Often done in constraint programming to obtain better propagation. x i = job assigned to worker i y j = worker assigned to job j Write each constraint in one or both variable systems. Add channeling constraints: x = j, all j y j y = i, all i x i

45 Another Interpretation We are doing the same with relaxations. x relaxation y relaxation min z z x, all i i i q 1 xi sq L L 4 ( 1) 1 q 1 z xi + 1 L L 1 qs i 4qs ( ) x i = j jy ij min j i V k y ij = 1, all i y w, all j, V 0 y 1 j j ij j k ij w channeling constraint

46 Another Interpretation Neither relaxation alone provides a good bound. 5-cycle problem with s = 2 Bound Odd hole cuts in y-relaxation 1.5 Finite domain cuts in x-relaxation 2.6 Combined models 4.0

47 Future Work Map other known finite-domain cuts into 0-1 models. What happens? Cardinality rules. Yan and Hooker (1999). Circuit constraint (TSP). Genc-Kaya and Hooker (2010). Cumulative constraint. Polyhedral analysis for other global constraints. General cardinality, nvalues, sequence, regular.

48 General Issues When do linked multiple relaxations provide better bounds than one relaxation? Can we say anything about the properties of different variable encodings? What are variables, and why do we use them?