1 / 17

Generic SBDD using Computational Group Theory

Generic SBDD using Computational Group Theory . Ian Gent St Andrews Warwick Harvey Imperial College Tom Kelsey St Andrews Steve Linton St Andrews. Conclusions . Two key novelties Entirely generic implementation of SBDD

ziya
Download Presentation

Generic SBDD using Computational Group Theory

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. Generic SBDD using Computational Group Theory Ian Gent St Andrews Warwick Harvey Imperial College Tom Kelsey St Andrews Steve Linton St Andrews

  2. Conclusions • Two key novelties • Entirely generic implementation of SBDD • user defines the group, does not write a dominance checker • Algebraic algorithm for dominance checking • searches given definition of an arbitrary group • Implementation provided in GAP-ECLiPSe • Groups up to size 1036 • compared to 109 for SBDS in GAP/ECLiPSe • Future work • Exploit special features of commonly occurring groups Generic SBDD

  3. Outline of Talk • Conclusions • Symmetry Breaking, SBDD, Dominance Checking • Groups and GAP • Generic Dominance Check • Discussion • Conclusions Generic SBDD

  4. Symmetry Breaking • Many CSPs have symmetry • n-queens, colouring, golfers’ problem, … • Breaking symmetry reduces search • avoids exploring equivalent states • Three main approaches to symmetry breaking • reformulate the problem • add constraints before search • adapt search algorithm to break symmetry • SBDD takes the last of these approaches Generic SBDD

  5. SBDD • Symmetry Breaking by Dominance Detection • Fahle, Schamberger, Sellmann, 2001 • Foccaci, Milano, 2001 • prefigured by Brown, Finkelstein, Purdom, 1988 • Do not search a node if you have searched its equivalent before • check before entering a node Generic SBDD

  6. Dominance Check 1. Deduce consequences of X≠a, Y=b, Z≠c, W ≠d 2. Is X=a equivalent to anything deduced? 3. Is Y=b, Z=c equivalent to anything deduced? 4. Is Y=b, W=d equivalent to anything deduced? Generic SBDD

  7. Generic Dominance Check • Two existing approaches to dominance checking • hand code checking procedure • express check as a constraint problem • Dominance check works only for a family of instances • e.g. golfers’ problems, BIBD instances • We have written a generic dominance checker • works for any constraint problem • user has only to define the group acting on the CSP Generic SBDD

  8. Generators and Groups • Consider simple problem • A3 + B3 + C3 + D3 = E3 + F3 + G3 • A…G :: 1..50 • Given any solution, we can permute the values of ABCD and of EFG, independently. • (A B C D) is permutation cycling A through D • (A D) swaps A and D • (E F G) cycles E through G • (E G) swaps E and G • These four permutations generate the group of different ways to construct one solution from another. • there are 4!3! = 144 group elements Generic SBDD

  9. GAP and GAP-ECLiPSe • GAP is world leading computational algebra package • accepts generators and computes using the whole group • e.g. we pass four elements (A B C D), (A D), (E F G), (E G) • GAP computes with full group of size 144 • 144 is unusually small • number of generators is at worst logarithmic in size of group • GAP-ECLiPSe • define generating elements and pass to GAP • constraint problem searched in ECLiPSe • dominance check written in GAP • could be done in any pair of Algebra/Constraint packages Generic SBDD

  10. Generic Dominance Check • failsets • sets of decisions at failed nodes • {X=a}, {Y=b,Z=c}, {Y=b,W=d} • pointset • set of assigned variables at current node • {Y=b,X=b,V=c} • includes propagated values • assume X=b, V=c set by inference • It is a heuristic choice to have failsets as small as possible, pointset as big as possible. Generic SBDD

  11. Dominance Check as Search • Dominance check is an algebraic search • Aside: anyone who has written a dominance check has written a computational algebra program! • Seek group element g and failset S s.t. • S g Pointset • Search process is essentially straightforward • but we omit most algebraic details here • critical to performance by several orders of magnitude Generic SBDD

  12. Dominance Check as Search • Recursive backtracking search • work through failset • find ways of mapping first element of failset to elements of pointset: try these in turn • recurse for second element, fixing the mapping of the first element • GAP allows us to make computation efficient • work in stabilizers, subgroups fixing work done so far • use transversals, the places that elements can map to • use Schreier vectors for efficiency Generic SBDD

  13. Dominance Check as Search • is there g, S with S g Pointset ? • S in failsets: {X=a}, {Y=b,Z=c}, {Y=b,W=d} • Pointset: {Y=b,X=b,V=c} • Use the fact that Y=b shared between second two failsets • Can we map X=a into Pointset? • no! • Can we map Y=b into Pointset? • Yes! • Y=b  Y=b • Use Stabilizer of Y=b • Can we map Z=c into {X=b,V=c}? • no! • Can we map W=d into {X=b, V=c}? • Yes! • {Y=b,W=d}  {Y=b, V=c} • Dominance detected so backtrack Generic SBDD

  14. Inference from Dominance • Dominance can deduce domain removals • Where a domain value would make dominance succeed, it can be removed • Already known (Fahle et al, Brown et al) • Again generically implemented • Work heuristically • i.e. only report removals we find easily • Petrie [SymCon03] shows this can increase search compared to SBDS, which finds all removals Generic SBDD

  15. Overview of Results • Can solve problems with very large groups in very generic manner • Main data on BIBD’s • solve up to <7,21,6,2,1>, group 1021 in few secs • solve up to <13,26,6,3,1>, group 1036 in few hrs • count number of symmetrically distinct solutions • compare with generic SBDS in GAP-ECLiPSe • only practical for <6,10,5,3,2>, group 109 • Competitive CSP method for solving BIBD’s • special purpose BIBD methods much faster • Hits a limit at moderate sized groups such as 1036 Generic SBDD

  16. Problems? Solutions? • GAP search can be prohibitive • e.g. in <13,26,6,3,1> BIBD, group size 1036 • 2053 dominance checks take 59,000 cpu secs • Should be able to exploit structure of group in general ways to reduce this in many cases • Constraint programmers don’t know groups • We have supplied macros to express groups easily • Harvey, Kelsey, Petrie (SymCon03) • McDonald (SymCon03) Generic SBDD

  17. Conclusions • Two key novelties • Entirely generic implementation of SBDD • user defines the group, does not write a dominance checker • Algebraic algorithm for dominance checking • searches given definition of an arbitrary group • Implementation provided in GAP-ECLiPSe • Groups up to size 1036 • compared to 109 for SBDS in GAP/Eclipse • Future work • Exploit special features of commonly occurring groups Generic SBDD

More Related