1 / 24

Russian Doll Search with Tree Decomposition

Russian Doll Search with Tree Decomposition. Martí Sànchez, David Allouche, Simon de Givry , and Thomas Schiex INRA Toulouse, FRANCE. Real applications with a structure. Radio Link Frequency Assignment. Earth Observation Satellite Management. CELAR SCEN-07r ( Constraints 4(1), 1999).

peony
Download Presentation

Russian Doll Search with Tree Decomposition

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. Russian Doll Searchwith Tree Decomposition Martí Sànchez, David Allouche, Simon de Givry, and Thomas Schiex INRA Toulouse, FRANCE IJCAI 2009

  2. Real applications with a structure RadioLinkFrequencyAssignment EarthObservationSatelliteManagement CELAR SCEN-07r(Constraints 4(1), 1999) SPOT5 #509 (Constraints 4(3), 1999) MendelianErrorDetection TagSNPSelection langladeM7 sheep pedigree (Constraints 13(1), 2008) HapMap chr01 r2≥0.8 #14481 IJCAI 2009

  3. Framework: Weighted CSP • (X,D,F,k) • X={X1,..., Xn}n variables • D={D1,..., Dn}n finitedomains of maximum size d • F={f1,…,fm}, a set ofmcost functions • fS with scope S, assigns a positive costto any assignment of S • f∅problem lower bound (obtained by enforcing local consistency by Equivalence Preserving Transformations) • k, maximum cost and global upper bound • Find an assignmentAof all the variables which minimizes ∑fSFfS ( A[S] ) NP-hard problem • Improve bounds computation compared to BTD (Jégou&Terrioux, ECAI 2004; de Givry et al., AAAI 2006) IJCAI 2009

  4. Soft 2-Coloring example A B C D E H I F J K G IJCAI 2009

  5. Soft 2-Coloring example A B C Optimal solution of cost 5 D E H I F J K G IJCAI 2009

  6. C1 C1 {D} {D} C2 C4 {E,F} {D,I,J} C3 C5 C2 C4 C3 C5 Tree Decomposition A B C D E H The set of clusters covers the set of variables and the set of cost functions I F J K G Separator = intersection between two connected clusters IJCAI 2009

  7. {D} {D} C1 C1 C2 C4 {E,F} {D,I,J} C3 C5 C2 C4 C3 C5 Search with Tree Decomposition A B C D E H I F J The assignment of a separator disconnects the problem into two independent subproblems G K IJCAI 2009

  8. {D} {D} C1 C2 C2 C4 {E,F} {D,I,J} C3 C5 H C4 I C3 J K C5 Search with Tree Decomposition A B C D H I E J F K G D 0 1 P2 / {(D=green)} P4 / {(D=green)} P4 / {(D=red)} P2 / {(D=red)} E 0 1 1 0 0 1 0 1 F 0 1 0 1 0 1 0 1 0 1 0 1 G 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 IJCAI 2009

  9. C2 C4 C3 C5 Search with Tree Decomposition A B C D H I E J F K G D 0 P2 / {(D=red)} E 0 1 F 0 1 0 1 G 0 1 0 1 0 1 0 1 IJCAI 2009

  10. C2 C4 C3 C5 Search with Tree Decomposition A Record the optimum of P2 / {(D=red)} LBP2 / {(D=red)} = 1 OPTP2 / {(D=red)} = true B C D H I E J F K G Assume a global upper bound k = 5. It may be useless to compute the optimum of P2 / {(D=red)},only a lower bound is needed! D 0 P2 / {(D=red)} E 0 1 F 0 1 0 1 G 0 1 0 1 0 1 0 1 IJCAI 2009

  11. C1 C2 C2 C4 C3 C5 C4 C3 C5 UBP2 / {(D=red)} = k – 3 – max ( fC4 + fC5, LBP4 / {(D=red)} ) Maintaining local consistency Recorded during search Backtrack bounded by Tree Decomposition BTD A (Jégou & Terrioux, ECAI 2004) (de Givry et al., AAAI 2006) B C D H I E J F K G Assume a global upper bound k = 5. It may be useless to compute the optimum of P2,only a lower bound is needed! D 0 P2 / {(D=red)} E 0 1 F 0 1 0 1 Add a local upper bound: UBP2 / {(D=red)} = k – 3 – LBP4 / {(D=red)} G 0 1 0 1 0 1 0 1 IJCAI 2009

  12. C1 C2 C4 C3 C5 Russian Doll Search with Tree Decomposition UBP2 / {(D=red)} = k – 3 – max ( fC4 + fC5, LBP4 / {(D=red)}, LBP4RDS ) RDS-BTD lower bound is a maximum between • Local Consistency • Recorded lower bounds • RDS lower bounds • PiRDS is a relaxation of Pi / A byremoving separator variablesand cost functions on them • For all i, solves PiRDSin a bottom-up orderusing BTD exploiting RDS lb Found during search Obtained in preprocessing The optimum of PiRDS is a lower bound of Pi / A for any separator assignment A IJCAI 2009

  13. RDS RDS RDS RDS RDS P1=P1 P2 P4 P3 P5 A A A B C B C B C LBP3 / {(E=red),(F=green)}=1, … D D D E E E H I H H I I F F F J J J K G K K G G A A B C B C D E D H I E H I F J F K G J K G Russian Doll Search with Tree Decomposition LBP5RDS=0 LBP3RDS=0 LBP2RDS=1 LBP1RDS=5 LBP4RDS=1 Same complexities as BTD: Time: O(exp(w+1)) Space: O(exp(s)) w = tree-width = largest cluster size –1, s = largest separator size IJCAI 2009

  14. Related Work C1 A C2 C1 C3 C2 C3 B C C4 C4 C5 D C5 C8 E K H C6 C6 F C7 C7 C9 J I G C10 C8 C11 C9 RDS (Verfaillie et al, AAAI 1996) C10 C11 IJCAI 2009

  15. Related Work C1 A C1 C2 C2 C3 B C C3 C4 C4 D C5 C5 C8 C8 E K H C6 C9 C6 F C9 J C7 C10 C11 I G C10 C7 C11 Pseudo-Tree RDS (Larrosa et al, ECAI 2002) AND/OR graph search (Marinescu & Dechter, AAAI 2006) IJCAI 2009

  16. C1 C1 {D} {D} C2 C4 {E,F} {D,I,J} C3 C5 C2 C1 C4 {D} {D} C3 C2 C4 U C5 C5 {E,F} C4 U C5 C3 Exploiting small separators only (Jégou et al., CP 2007) A B C D p = 2 E H I Merge clusters i and father(i) if |sepi,father(i)| > p F J K G  Gives more freedom for the dynamic variable ordering heuristic IJCAI 2009

  17. Exploiting small separators only A B C C1 D E K H p = 0 F J I G DFBB IJCAI 2009

  18. Tag SNP Selection Set covering problem with additional coverage quadratic criteria IJCAI 2009

  19. BTD RDS-BTD 4 DFBB Tag SNP Selection 25 largest instances from HapMap human chr01 (r2 ≥ 0.5) 171 ≤ n/2 ≤ 777 30 ≤ d ≤ 266 6% ≤ m ≤ 37% 14% ≤ w ≤ 23% Solved by toulbar2 v.0.8 with initialupper bound found by FESTA and with MCStree decomposition Time limit: 2 hours ∞ p IJCAI 2009

  20. Radio Link Frequency Assignment CELAR SCEN-07r (Constraints 4(1), 1999) IJCAI 2009

  21. Radio Link Frequency Assignment • Maximum Cardinality Search tree decomposition heuristic • Root selection: largest (SCEN-06) / most costly (SCEN-07r) cluster • Last-conflict variable ordering (Lecoutre et al., 20006) and dichotomic binary branching • Closing 1 open problem by exploiting tree decompostion RECORDS : SCEN-07r p = 3 BTD  90528 recorded LB (20 %) RDS-BTD 29453 recorded LB (6.4 %) RECORDS : SCEN-06 p = ∞ BTD  5633 recorded LB RDS-BTD 7145 recorded LB IJCAI 2009

  22. Conclusion and perspectives • RDS-BTD generalizes DFBB, RDS, PT-RDS, BTD,…exploiting better bounds • Needs only linear extra space • But there are not obtained in polynomial time • Restriction to small separators gives more robustness • Give more freedom to dynamic variable ordering inside clusters • Perspectives • Produce any-time lower bounds (Iterative Deepening) • Dedicated heuristics to find small separators, p auto-tuning toulbar2 C++ solver version 0.8 ( http://mulcyber.toulouse.inra.fr/projects/toulbar2 ) IJCAI 2009

  23. Cost projection from a unary cost function to f 2 2 1 1 1 0 1 0 1 1 0 0 0 0 1 0 Cost extension from a unary to a binary cost function Cost projection from a binary to a unary cost function f=0 f=0 f=0 A B A B A B 1 2 0 0 1 0 0 f=1 A B Enforcing local consistency byEquivalencePreservingTransformations IJCAI 2009

  24. C A E B D a b b a c d Weighted CSP model 1.0 0.8 0.9 1.0 0.8 A1 T|F B1 T|F a A2 a B2 IJCAI 2009

More Related