Upload
others
View
5
Download
0
Embed Size (px)
Citation preview
Johannes K. Fichte & Markus Hecher: PACE 2019
PACE 2019:The 4th Iteration
IPEC 2019, TU Munich, Germany
1
Johannes K. Fichte, TU DresdenMarkus Hecher, TU Wien & Univ. of Potsdam
Johannes K. Fichte & Markus Hecher: PACE 2019
History & Mission of PACEBart M.P. Jansen
2
Johannes K. Fichte & Markus Hecher: PACE 2019
History● Conceived in fall 2015
parameterized algorithmics should have a greater impact on practice● First iteration: 2015/16
○ Track A: Treewidth○ Track B: Feedback Vertex Set
● Second iteration: 2016/17○ Track A: Treewidth○ Track B: Minimum Fill-In
● Third iteration: 2017/18○ Track 1: Steiner tree exact with few terminals○ Track 2: Steiner tree exact with low treewidth○ Track 3: Steiner tree heuristic
● Currently: Fourth iteration3
Johannes K. Fichte & Markus Hecher: PACE 2019
MissionInvestigate applicability of algorithmic ideas
from parameterized complexity
1. Bridge gap between theory and practice2. Inspire new theoretical developments3. Investigate theoretical algorithms in practice4. Produce accessible implementations & benchmarks5. Encourage dissemination in scientific papers
4
Johannes K. Fichte & Markus Hecher: PACE 2019
Outcome● Previous iterations inspired long list of follow-up works● Follow-up Applications and Frameworks● Increased awareness of Parameterized Complexity
○ SAT community was particularly interested (this year)
● Usage of Benchmark instances○ Almost 12,000 publicly available, citable instances published (this year)
● Scientific Papers
5
Johannes K. Fichte & Markus Hecher: PACE 2019
Outcome● Previous iterations inspired long list of follow-up works● Follow-up Applications and Frameworks● Increased awareness of Parameterized Complexity
○ SAT community was particularly interested (this year)
● Usage of Benchmark instances○ Almost 12,000 publicly available, citable instances published (this year)
● Scientific Papers
6
Johannes K. Fichte & Markus Hecher: PACE 2019
Outcome● Previous iterations inspired long list of follow-up works● Follow-up Applications and Frameworks● Increased awareness of Parameterized Complexity
○ SAT community was particularly interested (this year)
● Usage of Benchmark instances○ Almost 12,000 publicly available, citable instances published (this year)
● Scientific Papers
7
Johannes K. Fichte & Markus Hecher: PACE 2019
Outcome● Previous iterations inspired long list of follow-up works● Follow-up Applications and Frameworks● Increased awareness of Parameterized Complexity
○ SAT community was particularly interested (this year)
● Usage of Benchmark instances○ Almost 12,000 publicly available, citable instances published (this year)
● Scientific Papers
8
Johannes K. Fichte & Markus Hecher: PACE 2019
Outcome● Previous iterations inspired long list of follow-up works● Follow-up Applications and Frameworks● Increased awareness of Parameterized Complexity
○ SAT community was particularly interested (this year)
● Usage of Benchmark instances○ Almost 12,000 publicly available, citable instances published (this year)
● Scientific Papers
9
Johannes K. Fichte & Markus Hecher: PACE 2019
Outcome● Previous iterations inspired long list of follow-up works● Follow-up Applications and Frameworks● Increased awareness of Parameterized Complexity
○ SAT community was particularly interested (this year)
● Usage of Benchmark instances○ Almost 12,000 publicly available, citable instances published (this year)
● Scientific Papers
10
Johannes K. Fichte & Markus Hecher: PACE 2019
Outcome● Previous iterations inspired long list of follow-up works● Follow-up Applications and Frameworks● Increased awareness of Parameterized Complexity
○ SAT community was particularly interested (this year)
● Usage of Benchmark instances○ Almost 12,000 publicly available, citable instances published (this year)
● Scientific Papers
11
Johannes K. Fichte & Markus Hecher: PACE 2019
Outcome● Previous iterations inspired long list of follow-up works● Follow-up Applications and Frameworks● Increased awareness of Parameterized Complexity
○ SAT community was particularly interested (this year)
● Usage of Benchmark instances○ Almost 12,000 publicly available, citable instances published (this year)
● Scientific Papers
12
Johannes K. Fichte & Markus Hecher: PACE 2019
Outcome● Previous iterations inspired long list of follow-up works● Follow-up Applications and Frameworks● Increased awareness of Parameterized Complexity
○ SAT community was particularly interested (this year)
● Usage of Benchmark instances○ Almost 12,000 publicly available, citable instances published (this year)
● Scientific Papers
13
Johannes K. Fichte & Markus Hecher: PACE 2019
Outcome● Previous iterations inspired long list of follow-up works● Follow-up Applications and Frameworks● Increased awareness of Parameterized Complexity
○ SAT community was particularly interested (this year)
● Usage of Benchmark instances○ Almost 12,000 publicly available, citable instances published (this year)
● Scientific Papers
14
Johannes K. Fichte & Markus Hecher: PACE 2019
Outcome● Previous iterations inspired long list of follow-up works● Follow-up Applications and Frameworks● Increased awareness of Parameterized Complexity
○ SAT community was particularly interested (this year)
● Usage of Benchmark instances○ Almost 12,000 publicly available, citable instances published (this year)
● Scientific Papers
15
Johannes K. Fichte & Markus Hecher: PACE 2019
PACE 2019
16
Johannes K. Fichte & Markus Hecher: PACE 2019
Program and Steering Committee
17
● Program Committee
Johannes K. Fichte TU DresdenMarkus Hecher TU Wien, University of Potsdam
Intern:Muhammad A. Dzulfikar University of Indonesia @TU Dresden
● Steering Committee
Édouard Bonnet Middlesex UniversityHolger Dell IT University of CopenhagenBart M. P. Jansen Eindhoven University of TechnologyThore Husfeldt IT University of Copenhagen and Lund UniversityPetteri Kaski Aalto UniversityChristian Komusiewicz Philipps-Universität MarburgFrances A. Rosamond University of BergenFlorian Sikora LAMSADE, Université Paris Dauphine
Johannes K. Fichte & Markus Hecher: PACE 2019
● … for computing resources
We would like to thank our Sponsors...
18
● … for prizes
Johannes K. Fichte & Markus Hecher: PACE 2019
Thanks go to
19
● All the participants of PACE!
● Intern Muhammad A. Dzulfikar from the University of Indonesia○ Performing instance selection○ Support for validating results○ ….
● Jan Badura at optil.io, who quickly implemented our requests○ Using results of several runs for the final results○ Customizing our judges for optil.io○ …
Johannes K. Fichte & Markus Hecher: PACE 2019
Tracks of PACE 2019
20
Johannes K. Fichte & Markus Hecher: PACE 2019
Track 1: Vertex Cover● Among the famous 21 first NP-complete problems by Karp● One of the famous, if not the most famous, graph problems● Great tradition in parameterized complexity
○ Well studied problem variants○ Different parameters○ Kernelizations○ Applications○ ...
u
● Track 1a: Compute a Minimum Vertex Cover (Exact)
21
Johannes K. Fichte & Markus Hecher: PACE 2019
Instance Selection for Vertex Cover● 9,591 instances among 8 different origins
○ PACE 2016○ TransitGraphs, Road-graphs○ SNAP○ frb○ ASP Horn backdoors, SAT Horn backdoors○ SAT2VC
● Classification by “Difficulty” (via Gurobi, numVC, Glucose) in intervals →
22
Johannes K. Fichte & Markus Hecher: PACE 2019
Track 2: Hypertree Decompositions● Motivation: Success of PACE 2016 & 2017 (Treewidth)● Applications for (Hyper-)tree Decompositions
○ Databases○ Constraint Programming
● Track 2a (EXACT): Compute a hypertree decomposition of minimum width
● Track 2b (HEUR): Heuristically compute a hypertree decomposition of small width (HEUR)
23
Johannes K. Fichte & Markus Hecher: PACE 2019
Hypertree Decompositions
● Given: Hypergraph H (higher arity edges)
● A Hypertree Decomposition D of H is, roughly speaking,○ A Tree Decomposition of H○ + a bag covering function (edge cover) over hyperedges○ + a certain monotonicity property (Descendent Condition) for the edge cover
● width(D) is size of the largest edge cover● htw(H) smallest width over all Hypertree Decompositions of H
24
Johannes K. Fichte & Markus Hecher: PACE 2019
Instance Selection for Hypertree Decompositions● 2,191 instances from hyperbench, originating from the area of CSP
○ DaimlerChrysler○ Grid2D○ MaxSAT, csp_application, csp_random, csp_other○ CQ
● Classification of “Difficulty” by means of ○ htdecomp, kdetdecomp○ Frasmt using the more generalized (fractional / generalized) hypertree decompositions
25
hyperbench.dbai.tuwien.ac.at
Johannes K. Fichte & Markus Hecher: PACE 2019
PACE 2019: Submission Requirements
26
Johannes K. Fichte & Markus Hecher: PACE 2019
Submission Requirements
27
1. Solvers + Dependencies have to be open source2. Source code of solver is maintained on public repository
+ long term data library3. A dedicated solver description is required4. Solvers for Tracks 1a and 2a are provably optimal
1. Submission on optil.io2. 30 minutes per instance3. 8 GB RAM per instance
Submission Limits
Johannes K. Fichte & Markus Hecher: PACE 2019
Participants of PACE 2019
28
Johannes K. Fichte & Markus Hecher: PACE 2019
Participants
29
● 18 teams and 33 participants from 10 countries
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of PACE 2019
30
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 1: Vertex Cover● Track 1a: Minimum Vertex Cover
31
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 1: Vertex Cover● Track 1a: Minimum Vertex Cover
32
Johannes K. Fichte & Markus Hecher: PACE 2019 33
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 1: Vertex Cover● Track 1a: Minimum Vertex Cover
34
Johannes K. Fichte & Markus Hecher: PACE 2019 35
225 €
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 1: Vertex Cover● Track 1a: Minimum Vertex Cover
36
Johannes K. Fichte & Markus Hecher: PACE 2019 37
250 €
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 1: Vertex Cover● Track 1a: Minimum Vertex Cover
38
Johannes K. Fichte & Markus Hecher: PACE 2019 39
275 €
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 1: Vertex Cover● Track 1a: Minimum Vertex Cover
40
Johannes K. Fichte & Markus Hecher: PACE 2019 41
300 €
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 1: Vertex Cover● Track 1a: Minimum Vertex Cover
42
Johannes K. Fichte & Markus Hecher: PACE 2019 43
450 €
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 1: Vertex Cover● Track 1a: Minimum Vertex Cover
44
Johannes K. Fichte & Markus Hecher: PACE 2019 45
450 €
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 1: Vertex Cover● Track 1a: Minimum Vertex Cover
47
Johannes K. Fichte & Markus Hecher: PACE 2019 48
750 €
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 1: Vertex Cover● Track 1a: Minimum Vertex Cover
49
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 2: Hypertree Decompositions
50
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 2: Hypertree Decompositions● Track 2a: MinHypertreeWidth
51
Johannes K. Fichte & Markus Hecher: PACE 2019 52
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 2: Hypertree Decompositions● Track 2a: MinHypertreeWidth
54
Johannes K. Fichte & Markus Hecher: PACE 2019 55
225 €
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 2: Hypertree Decompositions● Track 2a: MinHypertreeWidth
56
Johannes K. Fichte & Markus Hecher: PACE 2019 57
350 €
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 2: Hypertree Decompositions● Track 2a: MinHypertreeWidth
● Track 2b: HeurHypertreeWidth
58
Johannes K. Fichte & Markus Hecher: PACE 2019 59
150 €
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 2: Hypertree Decompositions● Track 2a: MinHypertreeWidth
● Track 2b: HeurHypertreeWidth
60
Johannes K. Fichte & Markus Hecher: PACE 2019 61
225 €
Johannes K. Fichte & Markus Hecher: PACE 2019
Results of Track 2: Hypertree Decompositions● Track 2a: MinHypertreeWidth
● Track 2b: HeurHypertreeWidth
62
Johannes K. Fichte & Markus Hecher: PACE 2019 63
350 €
Johannes K. Fichte & Markus Hecher: PACE 2019
Conclusion & Future of PACE
65
Johannes K. Fichte & Markus Hecher: PACE 2019
Lessons Learned● Solver description and public library essential
○ Maybe set solver description (abstract) deadline even at the beginning
● Selecting instances of moderate difficulty (vertex cover)○ Possible, but not easy○ Requires to collect numerous instances
● Problems and instances should be ready by the end of September○ Some students lost attention due to late problem announcement
● Cluster resources essential○ optil.io troubles during final submission phase○ optil.io runtime difference was in some cases up to 7 minutes between same instance + solver
● Score hard to comprehend for submitters○ limitation of optil.io: normalized scoreboard 0..1 depending
on performance of individual solver and instance66
Johannes K. Fichte & Markus Hecher: PACE 2019
Future Editions● Next years PACE Deadline will be later● Process of selecting new PCs ongoing
Problems?
● Subscribe to the newsletterand stay tuned!
pacechallenge.org
Hope we see you at PACE 2020.67
Johannes K. Fichte & Markus Hecher: PACE 2019
The Winning Team of Track 1a (Vertex Cover)
68
WeGotYouCovered
IPEC’19 · September 11, 2019Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash
KIT – University of the State of Baden-Wuerttemberg andNational Laboratory of the Helmholtz Association
INSTITUTE OF THEORETICAL INFORMATICS · ALGORITHMICS GROUP
www.kit.edu
The Winning Solver from the PACE 2019 Implementation Challenge, Vertex Cover Track
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
Vertex Cover and Complementary Problems
Input graph
Vertex cover Independent Set Clique
1
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
Techniques
Kernelization
Iterated Local Search
Branch-and-Bound
Branch-and-Reduce
· · ·3
7
Branch
ReduceBacktrack
Reduce
(1,2)-Swap
2
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
Kernelization [AI2016]
Reduce
Solver
Expand
Technique from FPT algorithmsApplies rich set of reduction rulesSignificantly reduces graph size
3
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
Iterated Local Search [ARW2012]
(1,2)-Swap
Originally developed for independent sets
Can often find (near-)optimal solutions
Perturbation to escape local optima
4
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
Branch and Reduce [AI2016]
Reduce graph after each branch
Additional branching rules to reduce graph size
Prune search based on lower bounds
Branch
ReduceBacktrack
5
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
Branch and Bound [LJM2017]
· · ·3
7
Originally developed for maximum cliques
Incremental MaxSAT reasoning to prune search
Combination of static and dynamic vertex ordering
6
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
Algorithm Overview
1.
Input Graph
Kernel
Branch-and-Reduceshort burst short burst
Branch-and-Reducelong run
Kernelization
2. Initial Solution
Branch-and-Boundlong run
Branch-and-Bound
6. 3. 4. 5.
Iterated Local Search
7
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
Algorithm Overview
1.
Input Graph
Kernel
Branch-and-Reduceshort burst short burst
Branch-and-Reducelong run
Kernelization
2. Initial Solution
Branch-and-Boundlong run
Branch-and-Bound
6. 3. 4. 5.
Iterated Local Search
7
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
Algorithm Overview
1.
Input Graph
Kernel
Branch-and-Reduceshort burst short burst
Branch-and-Reducelong run
Kernelization
2. Initial Solution
Branch-and-Boundlong run
Branch-and-Bound
6. 3. 4. 5.
Iterated Local Search
7
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
Algorithm Overview
1.
Input Graph
Kernel
Branch-and-Reduceshort burst short burst
Branch-and-Reducelong run
Kernelization
2. Initial Solution
Branch-and-Boundlong run
Branch-and-Bound
6. 3. 4. 5.
Iterated Local Search
7
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
Algorithm Overview
1.
Input Graph
Kernel
Branch-and-Reduceshort burst short burst
Branch-and-Reducelong run
Kernelization
2. Initial Solution
Branch-and-Boundlong run
Branch-and-Bound
6. 3. 4. 5.
Iterated Local Search
7
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
Algorithm Overview
1.
Input Graph
Kernel
Branch-and-Reduceshort burst short burst
Branch-and-Reducelong run
Kernelization
2. Initial Solution
Branch-and-Boundlong run
Branch-and-Bound
6. 3. 4. 5.
Iterated Local Search
7
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
Instances Solved Over Time
1 10 100 1000Time t (s)
0
40
80
120
160
200
Inst
ance
sso
lved
BnBKern. + BnB
ILS + BnRBnR
WeGotYouCovered
8
Demian Hespe, Sebastian Lamm, Christian Schulz, Darren Strash – WeGotYouCovered Institute of Theoretical InformaticsAlgorithmics Group
References
Akiba, Takuya, and Yoichi Iwata. “Branch-and-reduce exponential/FPTalgorithms in practice: A case study of vertex cover.” TheoreticalComputer Science 609 (2016): 211-225.
Li, Chu-Min, Hua Jiang, and Felip Manya. “On minimization of thenumber of branches in branch-and-bound algorithms for the maximumclique problem.” Computers & Operations Research 84 (2017): 1-15.
Andrade, Diogo V., Mauricio G. C. Resende, and Renato F. Werneck.“Fast local search for the maximum independent set problem.” Journalof Heuristics 18.4 (2012): 525-547.
Hespe, Demian, Sebastian Lamm, Christian Schulz and Darren Strash“WeGotYouCovered: The Winning Solver from the PACE 2019Implementation Challenge, Vertex Cover Track.” arXiv preprintarXiv:1908.06795 (2019).
9
Johannes K. Fichte & Markus Hecher: PACE 2019 69
Johannes K. Fichte & Markus Hecher: PACE 2019
See you outside at the
70
Thanks again for participating
Poster Session & PACE2020