Interpolating Between Stochastic and Worst-case Optimization · 2020. 1. 3. · Possible Responses...
Transcript of Interpolating Between Stochastic and Worst-case Optimization · 2020. 1. 3. · Possible Responses...
InterpolatingBetweenStochasticandWorst-caseOptimization
Risk– calculablegamble
Handlinginputrisk:Averagecaseanalysis
Uncertainty– incalculableunknown
'Symbol of Uncertainty', John Gilbert
Handlinginputuncertainty:Worst-casecompetitiveanalysis
Commoncomplaints
• Competitiveanalysisistoopessimistic
• Stochasticanalysisneedtoodetailedinformation
PossibleResponses• Relaxpessimism:Addrestrictionstoworst-caseinputscenariostoreduceimpact
• Temperoptimism:Makestochasticanalysisrobusttochangesininputdistribution
• “Haveitall”:Buildalgorithmsthatachievethe‘bestofbothworlds’andworkforbothkindsofinputs
• InterpolateModelANDPerformance:Buildhybridinputmodelsinterpolatingstochasticandadversarialinputstoderivenewalgorithmsthatdeterioratesmoothlyinperformance
Outline
• Motivation• Relatedwork*(incompleteandrepresentative)• AttemptatSpecificModel
RelaxPessimism:Restrictclassesofinputs• Betteralgorithmsforboundedtree-widthinputs
• Betterapproximationalgorithmsforboundedgenusgraphs
RelaxPessimism:SmoothedAnalysis• [Spielman-Teng 2001]Usesdistributionaldisturbanceovergivenworst-caseinput,tosmoothout(expected)performance
• Resultingrunningtimesarepolynomialininputandperturbationsize
• Goal:Explain“unreasonableeffectiveness”ofpopularalgorithms(likeSimplexforLP)
Temperoptimism:StochasticProgrammingVariants
Stochastic Combinatorial Optimization with controllable risk aversion levelSo, Zhang & Ye, APPROX ‘06, Math of OR
TemperOptimism:DistributionallyRobustOptimizationAddambiguitytorisk• Newsvendorproblemwithonlymeanandvariance,notdistribution[Scarf1958]
• DistributionallyRobustStochasticOptimization[Zackova,Dupacova,Bertsimas+,Sim+,Ye+,…]
TemperOptimism:CorrelationRobustness
Correlation gap: Loss in performance by ignoring correlations and assuming independence
Haveitall:UniversalApproximationsOnesinglesolutionwhoseinducedsolutionisgoodforanysubsetinput [Bartholdi-Platzman,Jia+]
Haveitall:Robust/IncrementalSolutions• Matchings[Hassin-Rubinstein,SIDMA2002]Givenaweightedgraphfindasinglematchingandanorderingofitsedgessuchthatforeveryk,theprefixofkedgesisnearoptimalmaximumweightmatchingofsizek
• Metrick-median[Mettu-Plaxton,SICOMP2003]Findanorderingoffacilitiessuchthatforeveryk,theprefixofkfacilitiesisanearoptimalk-median
Haveitall/Bestofboth:Trade-offtwoguaranteesTrytofindthebestpossibleratioswithrespecttothe“pessimistic”and“optimistic”extremes
Online Algorithms with Uncertain InformationMahdian, Nazerzadeh & Saberi, EC ’07, TALG 2012
Noteperformanceisw.r.t.thatofthegivenalgorithmsbutnotthe“optimal”solutions
Bestofboth:OnlineResourceAllocation
Bestofboth:AdWords
Maintain(1– 1/e)-worstcaseguaranteeintheworstcaseanddoquantifiablybetterforrandomarrivalmodel“Inthispaperwedesignalgorithmsthatachieveacompetitiveratiobetterthan1−1/eonaverage,whilepreservinganearlyoptimalworstcasecompetitiveratio.Ideally,wewanttoachievethebestofbothworlds,i.e,todesignanalgorithmwiththeoptimalcompetitiveratioinboththeadversarialandrandomarrivalmodels.Weachievethisforunweightedgraphs,butshowthatitisnotpossibleforweightedgraphs.”
Simultaneous Approximations for Adversarial and Stochastic Online Budgeted AllocationMirrokni, Oveis-Gharan & Zadimoghaddam, SODA ‘12
Bestofboth:Balancedguaranteesforbandits
AudienceParticipation:OtherRelatedWork?
RecallPossibleResponses• Relaxpessimism:Addrestrictionstoworst-caseinputscenariostoreduceimpact
• Temperoptimism:Makestochasticanalysisrobusttoslightchangesininputdistribution
• “Haveitall”:Buildalgorithmsthatachievethe‘bestofbothworlds’andworkforbothkindsofinputs
• InterpolateModelANDPerformance:Buildhybridinputmodelsinterpolatingstochasticandadversarialinputstoderivenewalgorithmsthatdeterioratesmoothlyinperformance
Proposal:InterpolateModelsANDPerformance• Modelinterpolation:Inputmodelshouldallowsmoothinterpolationbetweenstochasticoptimismandworst-casepessimism
• Performanceinterpolation:Algorithmshouldhaveperformanceratiothatinterpolatessmoothlybetweenthebetterguaranteeforstochasticinputsandtheworseguaranteefortheworst-case
Restofthetalk[JointwithGuyBlelloch,Kedar Dhamdhere &Suporn Pongnumkul,Summer2004]
• ListUpdateProblem:Competitive&AverageCaseAnalysis
• AHybridOnlineModel• SetupandResultsfromapreliminaryexperiment• Conjecture
ListUpdateProblemSelf-organizingsequentialsearch
• Unsortedlist
• Receivedasequenceofrequests
• Costofaccessingtheith elementofthelistisi.Afteraccess,canmoveitanywhereaheadinthelistforfree
• Cantransposeanypairofadjacentelementsatunitcost
y w z x v uL:
y w z x v uL:
ListUpdateExample
Action Cost Total Cost
y w z x v uL:
ListUpdateExample
Action Cost Total CostAccess x
y w z x v uL:
ListUpdateExample
Action Cost Total CostAccess x 4 4
y w z x v uL:
ListUpdateExample
Action Cost Total CostAccess x 4 4Access v
y w z x v uL:
ListUpdateExample
Action Cost Total CostAccess x 4 4Access v 5 9
y w z x v uL:
ListUpdateExample
Action Cost Total CostAccess x 4 4Access v 5 9
Move v forward to before w
y v w z x uL:
ListUpdateExample
Action Cost Total CostAccess x 4 4Access v 5 9
Move v forward to before w 0 9
y v w z x uL:
ListUpdateExample
Action Cost Total CostAccess x 4 4Access v 5 9
Move v forward to before w 0 9Access y
y v w z x uL:
ListUpdateExample
Action Cost Total CostAccess x 4 4Access v 5 9
Move v forward to before w 0 9Access y 1 10
y v w z x uL:
ListUpdateExample
Action Cost Total CostAccess x 4 4Access v 5 9
Move v forward to before w 0 9Access y 1 10Access v
y v w z x uL:
ListUpdateExample
Action Cost Total CostAccess x 4 4Access v 5 9
Move v forward to before w 0 9Access y 1 10Access v 2 12
y v z w x uL:
ListUpdateExample
Action Cost Total CostAccess x 4 4Access v 5 9
Move v forward to before w 0 9Access y 1 10Access v 2 12
Transpose w and z
y v w z x uL:
ListUpdateExample
Action Cost Total CostAccess x 4 4Access v 5 9
Move v forward to before w 0 9Access y 1 10Access v 2 12
Transpose w and z 1 13
AverageCaseAnalysis
• Assumeeachrequestcomesfromafixedprobabilitydistribution,independentofpreviousrequests.Supposetheith itemhasprobabilitypi.Designalgorithmstominimizetheexpectedcost.
• Optimalstrategyistokeepthelistsortedinnon-increasingorderofpi.
STAT=StaticList
• Listissortedinnon-increasingorderoftheprobabilities
• Nevermovesanything
• Goodforwhenwehaveagoodestimateoftheprobabilitydistribution.
FC:FrequencyCount
• Ifprobabilitydistributionisunknown,estimateitusingfrequencycounts
• Keeplistsortedaccordingtocounts
CompetitiveAnalysis
• Definition: Ananalysisinwhichtheperformanceofanonlinealgorithmiscomparedtothebestthatcouldhavebeenachievedifalltheinputshadbeenknowninadvance.
CompetitiveRatio
A:Our online algorithm
CA(s)
OPT:Optimal Offline
algorithm
COPT(s)
A is c-competitive if $ aCA(s) ≤ c COPT(s) + a
for all request sequences s
Move-to-Front(MTF)
[Sleator,Tarjan,CACM1985]Whenanelementisaccessed,moveittothefrontofthelist.
• Theorem:MTFhascompetitiveratio2againstoptimalofflinealgorithm.
PerformanceComparison
• E(FC)/E(STAT) =1[Rivest CACM1976]• E(MTF)/E(STAT)=𝜋/2 =1.58…[Chung,Hajela,SeymourSTOC85]
TStimestampalgorithm[AlbersSODA95,Albers&MitzenmacherAlgorithmica 1998]Ourworkismotivatedbythegoaltopresentauniversalalgorithmthatachievesagoodcompetitiveratiobutalsoperformsespeciallywellwhenrequestsaregeneratedbydistributions.(“Haveitall”)TS:Inserttherequesteditem,sayx,infrontofthefirstiteminthelistthathasbeenrequestedatmostoncesincethelastrequesttox.Ifxhasnotbeenrequestedsofar,leavethepositionofxunchanged.Theorem TSis2-competitive
E(TS)/E(STAT)=1.5whileforsomedistribution,E(MTF)/E(STAT)>1.57
NewHybridInterpolatingModel• Assumeafixedprobabilitydistribution
• Foreachrequest,withprobability,letadversarychangetherequest.
• ó AverageCaseAnalysis
• ó CompetitiveAnalysis
• ó Knownprobabilitydistributionwithuncertainty.
),...,,( 21 npppp =!
e0=e1=e10 << e e
Desiderata:InterpolatingAlgorithmforHybridModel
• Takesasinputestimatesofp ande
• Matchesbestaveragecaseperformancewhene islow,andmatchesbestcompetitiveratiowhene ishigh,andinterpolatesinbetween.
CandidateAlgorithm:Move-From-Back-Epsilon• Listinitiallysortedinnon-increasingorderofprobabilities.
• Whenanelementx isaccessed,promoteitpastothersthathaveprobabilityuptopx +e.
y w z x v uL:
MFBEExample(e =0.2)
Request Cost Total Cost
s: v y z yProb py = 0.5 pw = 0.2 pz = 0.15 px = 0.1 pv = 0.03 pu = 0.02
y w z x v uL:
MFBEExample(e =0.2)
Request Cost Total Cost
s: x v y vProb py = 0.5 pw = 0.2 pz = 0.15 px = 0.1 pv = 0.03 pu = 0.02
y w z x v uL:
MFBEExample(e =0.2)
Request Cost Total CostAccess x (and move to before w)
s: x v y vProb py = 0.5 pw = 0.2 pz = 0.15 px = 0.1 pv = 0.03 pu = 0.02
y x w z v uL:
MFBEExample(e =0.2)
Request Cost Total CostAccess x (and move to before w) 4 4
s: x v y vProb py = 0.5 px = 0.1 pw = 0.2 pz = 0.15 pv = 0.03 pu = 0.02
y x w z v uL:
MFBEExample(e =0.2)
Request Cost Total CostAccess x (and move to before w) 4 4
s: x v y vProb py = 0.5 px = 0.1 pw = 0.2 pz = 0.15 pv = 0.03 pu = 0.02
y x w z v uL:
MFBEExample(e =0.2)
Request Cost Total CostAccess x (and move to before w) 4 4Access v (and move to before x)
s: x v y vProb py = 0.5 px = 0.1 pw = 0.2 pz = 0.15 pv = 0.03 pu = 0.02
y v x w z uL:
MFBEExample(e =0.2)
Request Cost Total CostAccess x (and move to before w) 4 4Access v (and move to before x) 5 9
s: x v y vProb py = 0.5 pv = 0.03 px = 0.1 pw = 0.2 pz = 0.15 pu = 0.02
y v x w z uL:
MFBEExample(e =0.2)
Request Cost Total CostAccess x (and move to before w) 4 4Access v (and move to before x) 5 9
s: x v y vProb py = 0.5 pv = 0.03 px = 0.1 pw = 0.2 pz = 0.15 pu = 0.02
y v x w z uL:
MFBEExample(e =0.2)
Request Cost Total CostAccess x (and move to before w) 4 4Access v (and move to before x) 5 9
Access y
s: x v y vProb py = 0.5 pv = 0.03 px = 0.1 pw = 0.2 pz = 0.15 pu = 0.02
y v x w z uL:
MFBEExample(e =0.2)
Request Cost Total CostAccess x (and move to before w) 4 4Access v (and move to before x) 5 9
Access y 1 10
s: x v y vProb py = 0.5 pv = 0.03 px = 0.1 pw = 0.2 pz = 0.15 pu = 0.02
y v x w z uL:
MFBEExample(e =0.2)
Request Cost Total CostAccess x (and move to before w) 4 4Access v (and move to before x) 5 9
Access y 1 10
s: x v y vProb py = 0.5 pv = 0.03 px = 0.1 pw = 0.2 pz = 0.15 pu = 0.02
y v x w z uL:
MFBEExample(e =0.2)
Request Cost Total CostAccess x (and move to before w) 4 4Access v (and move to before x) 5 9
Access y 1 10Access v
s: x v y vProb py = 0.5 pv = 0.03 px = 0.1 pw = 0.2 pz = 0.15 pu = 0.02
y v x w z uL:
MFBEExample(e =0.2)
Request Cost Total CostAccess x (and move to before w) 4 4Access v (and move to before x) 5 9
Access y 1 10Access v 2 12
s: x v y vProb py = 0.5 pv = 0.03 px = 0.1 pw = 0.2 pz = 0.15 pu = 0.02
DifficultiesinProvingPropertiesofMFBE• MustcomparewithOPTratherthanSTAT• OPTcanbecomputedbyDynamicprogramming
• Trivialway=O((n!)2m)• Improvement=O((2n)(n!)m)
[Reingold,Westbrook,1996]
• HardtoderiveusefulpropertiesofOPT
Experiments(Pongnumkul ALADDINREU2004)• Motivation:Toseethebehaviorofalgorithmsinourhybridmodel.
• Measurement:Wemeasuretheperformanceofanonlinealgorithmbytheaveragecompetitiveratio.
OurExperiment
• Variablesinourexperiment• TypeofListUpdateAlgorithm(MTF,STAT,MFBE)
• TypeofProbabilityDistribution• TypeofAdversary
• Epsilon:e
OurExperiment
• Wegeneratearequestsequenceoflength100,withachosenprobabilitydistribution
• Then,withprobabilitye changetherequestsequence adverserially
OurExperiment• Recordthecostincurredbytheonlinealgorithm=CostA(s)
• UseDynamicProgrammingtofindoptimumcostofthatrequestsequence=CostOPT(s).
• CompetitiveRatio=CostA(s)/CostOPT(s)
• Repeatthis100timestofindtheaveragecompetitiveratio.
Distribution
• GeometricDistribution:• P[i]/ 1/2i
• UniformDistribution:• P[i]=1/n,n=lengthofthelist
• ZipfianDistribution(Zipf(2)):• P[i]/ 1/i2
CruelAdversary
• Thisisanadaptiveadversary
• Looksatthecurrentlistandrequestthelastiteminthelist.
y w z x v uL:
CruelAdversary,GeometricDistribution,n=6
Average Competitive Ratio
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Epsilon
Aver
age
Com
petit
ive
Rat
io
STAT
MFBE
MTF
CruelAdversary,GeometricDistribution,n=6(zoomedin)
Average Competitive Ratio
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Epsilon
Aver
age
Com
petit
ive
Rat
io
STAT
MFBE
MTF
OtherAdversaries
• Geometricadversarychooseselementsrandomly,accordingtothegeometricdistributiononthereversedSTATorder
• Uniformadversaryrequestselementsfromthelistuniformlyatrandom
• ObliviousAdversarydoesn’tlookatthecurrentlist
UniformAdversary,Zipfian2Distribution,n=6
Average Competitive Ratio
1
1.05
1.1
1.15
1.2
1.25
1.3
1.35
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Epsilon
Aver
age
Com
petit
ive
Rat
io
STAT
MFBE
MTF
ReversedGeometricAdversary,GeometricDistribution,n=6
Average Competitive Ratio
1
1.1
1.2
1.3
1.4
1.5
1.6
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Epsilon
Aver
age
Com
petit
ive
Rat
io
STAT
MFBE
MTF
Observations
• Theperformanceofanyalgorithminthishybridmodeldependsheavilyonthetypeofadversary.
• MFBEseemsbetterthantheworseofSTATandMTF.
Conjecture
(Nottheworst)TheaveragecompetitiveratioofMFBEisdominatedbythemaximumoftheaveragecompetitiveratiosofSTATandMTF.
• (Bestofall)Whatwewantedbutprobablynottrue:Avg competitiveratioofMFBEisatmosttheminimumoftheaveragecompetitiveratiosofSTATandMTF.
InterpolateModelANDPerformance:Buildhybridinputmodelsinterpolatingstochasticandadversarialinputstoderivenewalgorithmsthatdeterioratesmoothlyinperformance• Prove/disprove“MFBEisnottheworst”• Findasettingwherethehybridmodelinterpolatesstochasticandworst-caseinputs,andthealgorithminterpolatestheperformanceofaveragecaseandworstcasealgorithms
• Aretheremoreeffectiveapproachestotrade-offstochasticoptimismandworst-casepessimism?
Summary