Maximizing a Monotone Submodular Function Subject to a Covering

MaximizingaMonotoneSubmodular
FunctionSubjecttoaCoveringanda
PackingConstraint
JoachimSpoerhase
UniversityofWroclaw,Poland
[email protected]
HighlightsofAlgorithms
10th June,2017
JointworkwithSumedhaUniyal
Messagetothe
Audience
SubmodularFunctions
DurationofyourPresentation(minutes)
Messagetothe
Audience
SubmodularFunctions
5
DurationofyourPresentation(minutes)
Messagetothe
Audience
SubmodularFunctions
5
10
DurationofyourPresentation(minutes)
Messagetothe
Audience
SubmodularFunctions
5
10
20
DurationofyourPresentation(minutes)
Messagetothe
Audience
SubmodularFunctions
5
10
20
DurationofyourPresentation(minutes)
Capturesmanycombinatorialoptimizationproblemslike
- Max-k-Coverage
- Max-Cut
- MaxFacilityLocationโ€ฆ
๐ด
๐‘’
๐ต
๐“ค
Asetfunctiononallsubsetsof๐’ฐ, ๐‘“: 2|๐’ฐ| โ†’ โ„A suchthatโˆ€๐ด โІ ๐ต โІ ๐’ฐ, ๐‘’ โˆˆ ๐’ฐ\B,
๐‘“ ๐ด โˆช ๐‘’ โˆ’ ๐‘“ ๐ด โ‰ฅ ๐‘“ ๐ต โˆช ๐‘’ โˆ’ ๐‘“(๐ต) (Submodular)
๐‘“(๐ต) โ‰ฅ ๐‘“(๐ด) (Monotone)
CoveringandPackingConstraints
W=50k๏ผ„
3 30k๏ผ„
6
4
25k๏ผ„
4
2
20k๏ผ„
2
1 50k๏ผ„
10
Foreachelemente โˆˆ ๐’ฐ,wearegivesomeweightaprofitfunction๐‘: ๐’ฐ โ†’ โ„•,a
weightfunction๐‘ค: ๐’ฐ โ†’ โ„• ,aprofitrequirement๐‘ƒ andabudget๐ต.
Feasibility NP-Hard
Goal:max๐’‡(๐‘จ) forsome๐ด โІ ๐’ฐ suchthat
(Subset-Sum Problem)
๐‘(๐ด) โ‰ฅ ๐‘ƒand๐‘ค(๐ด) โ‰ค ๐‘Š
Asetfunctiononallsubsetsof๐’ฐ, ๐‘“: 2|๐’ฐ| โ†’ โ„A suchthatโˆ€๐ด โІ ๐ต โІ ๐’ฐ, ๐‘’ โˆˆ ๐’ฐ\B,
๐‘“ ๐ด โˆช ๐‘’ โˆ’ ๐‘“ ๐ด โ‰ฅ ๐‘“ ๐ต โˆช ๐‘’ โˆ’ ๐‘“(๐ต) (Submodular)
๐‘“(๐ต) โ‰ฅ ๐‘“(๐ด) (Monotone)
CoveringandPackingConstraints
W=50k๏ผ„
3 30k๏ผ„
6
2
20k๏ผ„
2
Foreachelemente โˆˆ ๐’ฐ,wearegivesomeweightaprofitfunction๐‘: ๐’ฐ โ†’ โ„•,a
weightfunction๐‘ค: ๐’ฐ โ†’ โ„• ,aprofitrequirement๐‘ƒ andabudget๐ต.
Feasibility NP-Hard
Goal:max๐’‡(๐‘จ) forsome๐ด โІ ๐’ฐ suchthat
(Subset-Sum Problem)
๐‘(๐ด) โ‰ฅ ๐‘ƒand๐‘ค(๐ด) โ‰ค ๐‘Š
Asetfunctiononallsubsetsof๐’ฐ, ๐‘“: 2|๐’ฐ| โ†’ โ„A suchthatโˆ€๐ด โІ ๐ต โІ ๐’ฐ, ๐‘’ โˆˆ ๐’ฐ\B,
๐‘“ ๐ด โˆช ๐‘’ โˆ’ ๐‘“ ๐ด โ‰ฅ ๐‘“ ๐ต โˆช ๐‘’ โˆ’ ๐‘“(๐ต) (Submodular)
๐‘“(๐ต) โ‰ฅ ๐‘“(๐ด) (Monotone)
StateoftheArtResults
โ€ข
MonotoneSubmodularMaximization
โ€“ UnderCardinalityConstraint
โ€ข Greedy1-1/eโ€“ apx [Nemhauser etal.,Math.Program.โ€™78]
โ€ข Hardness:1-1/e [Nemhauser etal.,Math.O.R..,1978],[Feige,STOCโ€™96]
โ€“ UnderPackingConstraint
โ€ข Greedy1-1/eโ€“ apx [Sviridenko,O.R.Lett.โ€™04]
โ€“ O(1)PackingConstraints
โ€ข Multilinearrelaxation:1-1/eโ€“ ฮต โ€“ apx [Kulik etal.,SODAโ€™09]
โ€“ MatroidConstraints
โ€ข Multilinearrelaxation:0.309โ€“ apx [Vondrák,STOCโ€™08],[Calinescu etal.,IPCOโ€™07]
โ€ข
Capacitatedkโ€“ MedianProblem
โ€“ Hardness:1.736 (followsfromtheuncapacitated case)
โ€“ Approximationfactor:O(logn)[Folklore]
โ€“ O(1)approximationalgorithmsundercapacity/cardinalityviolations.
9
Algorithms
โ€ข Agreedybaseddynamicprogram.
โ€ข Outputsanapproximatesolution whichisbestofallpossiblegreedy
Wโ€˜
chains.
Greedy chain may get stuck
Pโ€˜
Bestgreedy chain
Weights/profits as in
optimum solution
Tableentry:๐‘‡[๐‘™, ๐‘ƒโ€ฒ, ๐‘Šโ€ฒ] stores๐‘™-elementapproximatesolutionwith
profit๐‘ƒโ€ฒandweight๐‘Šโ€ฒ.
Tocompute๐‘‡[๐‘™, ๐‘ƒโ€ฒ, ๐‘Šโ€ฒ] pickbestwaytoextendanentry
๐‘‡[๐‘™ โˆ’ 1, ๐‘ƒโ€ฒโ€ฒ, ๐‘Šโ€ฒโ€ฒ] byoneelementinafeasiblemanner.
โ€ 
โ€ฆ
Output:Bestentry๐‘‡[๐‘™, ๐‘ƒโ€ฒ, ๐‘Šโ€ฒ]with๐‘ƒ โ‰ฅ โ€ก and๐‘Šโ€™ โ‰ค ๐‘Š.
10
Algorithms
โ€ข Agreedybaseddynamicprogram.
โ€ข Outputsanapproximatesolution whichisbestofallpossiblegreedy
chains.
ForbiddenSets
Crucialtoreducetheprofitviolationfrom2 + ๐œ€ downto 1 + ๐œ€.
Worksforonecoveringandonepackingconstraint.
Idea:Foreach๐‘‡[๐‘ƒโ€ฒ, ๐‘Šโ€ฒ],forbidthesuffixset๐นโ€นโ€ฆ oforderedsetbynonล’(โ€ข)
increasingลฝ(โ€ข) suchthat
p(๐‘‡ ๐‘ƒโ€ฆ , ๐‘Š โ€ฆ โˆช ๐นโ€นโ€ฆ ) โ‰ฅ ๐‘ƒ andw(๐‘‡ ๐‘ƒ โ€ฆ , ๐‘Š โ€ฆ โˆช ๐นโ€นโ€ฆ ) โ‰ค ๐‘Š.
Tableentry:๐‘‡[๐‘™, ๐‘ƒโ€ฒ, ๐‘Šโ€ฒ] stores๐‘™-elementapproximatesolutionwith
profit๐‘ƒโ€ฒandweight๐‘Šโ€ฒ.
Tocompute๐‘‡[๐‘™, ๐‘ƒโ€ฒ, ๐‘Šโ€ฒ] pickbestwaytoextendanentry
๐‘‡[๐‘™ โˆ’ 1, ๐‘ƒโ€ฒโ€ฒ, ๐‘Šโ€ฒโ€ฒ] byoneelementinafeasiblemanner.
โ€ 
โ€ฆ
Output:Bestentry๐‘‡[๐‘™, ๐‘ƒโ€ฒ, ๐‘Šโ€ฒ]with๐‘ƒ โ‰ฅ โ€ก and๐‘Šโ€™ โ‰ค ๐‘Š.
11
MainTheorems
Theorem1.Thereisanalgorithmformaximizingamonotonesubmodular
functionsubjecttoO(1)coveringandonepackingconstraintthatoutputsfor
any๐œบ > ๐ŸŽin๐’๐‘ถ(๐Ÿ/๐œบ) timea4-approximatesolutionwithprofitatleast
๐Ÿ
โˆ’ ๐œบ ๐‘ทandwithweightatmost ๐Ÿ + ๐œบ ๐‘พ.
๐Ÿ
Theorem2.Thereisanalgorithmformaximizingamonotonesubmodular
functionsubjecttoonecoveringandonepackingconstraintthatoutputsforany
๐œบ > ๐ŸŽin๐’๐‘ถ(๐Ÿ/๐œบ) timea4-approximatesolutionwithprofitatleast ๐Ÿ โˆ’ ๐œบ ๐‘ทand
withweightatmost ๐Ÿ + ๐œบ ๐‘พ.
Corollary.Thereisa2.6- approximationalgorithmfork-medianproblemwith
non-uniformandhardcapacitiesiftheunderlyingmetricspacehasonlytwo
possibledistances.
12
Factor-revealingLP
๐ฆ๐ข๐ง ๐’‚๐’ ๐‘ ๐‘ข๐‘๐‘—๐‘’๐‘๐‘ก๐‘ก๐‘œ
(LP) ๐ฆ๐š๐ฑ โˆ‘๐’ ๐’Š ๐’š ๐‘ ๐‘ข๐‘๐‘—๐‘’๐‘๐‘ก๐‘ก๐‘œ
๐’Н๐Ÿ ๐’ ๐’Š
๐‘–
๐‘Ž¥ โ‰ฅ 1 โˆ’ ๐‘œ¥ ;
๐‘›
(DUAL)
๐‘ฅ© + ๐‘ฆ© โˆ’ ๐‘ฅ©A¥ โ‰ค 0โˆ€๐‘– โˆˆ ๐‘› โˆ’ 1 ;
๐‘–
๐‘Ž© โ‰ฅ ๐‘Ž©ª¥ + 1 โˆ’ ๐‘œ© โˆ€๐‘– โˆˆ ๐‘› \{1};
๐‘›
๐‘ฅ· + ๐‘ฆ· โ‰ค 1;
©
๐‘–
๐‘Ž© โ‰ฅ
1 โˆ’ - ๐‘œ® โˆ€๐‘– โˆˆ ๐‘› ;
๐‘›
®¯¥
โˆ‘·®¯© ๐‘ฆ® โˆ’ 1 โˆ’ ๐‘ฅ© โ‰ค 0โˆ€๐‘– โˆˆ ๐‘› ;
·
·
๐‘Ž© โ‰ฅ 0,๐‘œ© โ‰ฅ 0โˆ€๐‘– โˆˆ ๐‘› .
๐‘ฅ© โ‰ฅ 0,๐‘ฆ© โ‰ฅ 0โˆ€๐‘– โˆˆ ๐‘› .
®
©
Theorem3.Thereisanalgorithmformaximizingamonotonesubmodular
functionsubjecttoO(1)coveringandonepackingconstraintthatoutputsfor
any๐œบ > ๐ŸŽin๐’๐‘ถ(๐Ÿ/๐œบ) timeae-approximatesolutionwithprofitatleast
๐Ÿ
โˆ’ ๐œบ ๐‘ทandwithweightatmost ๐Ÿ + ๐œบ ๐‘พ.
๐Ÿ
13
OpenProblems
โ€ข Canwegetatight1 โˆ’ 1/๐‘’ approximationfor1 cardinalityand1
coveringconstraintwithoutanyviolationortheproblembecomes
harder?
โ€ข Canwegeta๐‘‚(1)approximationfor1 coveringand1 packing
constraintbyviolatingonlyoneconstraint?
โ€ข WhichotherproblemscanbenefitfromourgreedyDPtechnique?
14
Thankyou