ebook img

ALICE: Analysis & Learning Iterative Consecutive Executions PDF

265 Pages·2016·6.31 MB·English
by  
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview ALICE: Analysis & Learning Iterative Consecutive Executions

ALICE: Analysis & Learning Iterative Consecutive Executions Helga Ingimundardóttir Faculty of Industrial Eng., Mechanical Eng. and Computer Science University of Iceland 2016 Dissertationforthedegreeofdoctorofphilosophy ALICE: Analysis & Learning Iterative Consecutive Executions HelgaIngimundardóttir SchoolofEngineeringandNaturalSciences FacultyofIndustrialEng.,MechanicalEng.andComputerScience Reykjavík,June2016 AdissertationpresentedtotheUniversityofIcelandSchoolofEngineeringandNaturalSciences incandidacyforthedegreeofdoctorofphilosophy. Doctoralcommittee Prof.TómasPhilipRúnarsson FacultyofEngineering,UniversityofIceland Prof.GunnarStefánsson FacultyofPhysicalSciences,UniversityofIceland Prof.MichèleSebag ResearchdirectorCNRS,LRI,UniversitéParis-Sud Opponents Prof.EdmundKieranBurke ScienceandEngineering,QueenMaryUniversityofLondon Prof.KateSmith-Miles SchoolofMathematicalSciences,MonashUniversity ALICE:Analysis&LearningIterativeConsecutiveExecutions ©2016HelgaIngimundardóttir PrintedinIcelandbySvansprent ISBN978-9935-9307-9-8 Alltfyrirmóðurmína iv Eitherthewellwasverydeep,orshefellveryslowly,forshehad plentyoftimeasshewentdowntolookaboutherandtowonder whatwasgoingtohappennext. Narrator Abstract Overtheyearstherehavebeenmanyapproachestocreatedispatchingrulesforscheduling. Re- centpasteffortshavefocusedondirectsearchmethods(e.g. geneticprogramming)ortraining ondata(e.g.supervisedlearning).Thedissertationwillexaminethelatterandgiveaframework calledAnalysis&LearningIterativeConsecutiveExecutions(ALICE)onhowtodoiteffectively. Definingtrainingdataasfφ(x(k));y(k)gK 2 Dthedissertationwillshow:i)samplesφ(x) i i k=1 i D shouldrepresenttheinduceddatadistribution . Thisdonebyupdatingthelearnedmodelin anactiveimitationlearningfashion;ii)y islabelledusinganexpertpolicyviaasolver;iii)data i needstobebalanced, asthesetisunbalancedw.r.t. thedispatchingstepk, andiv)toimprove uponlocalisedstepwisefeaturesφ,it’spossibletoincorporate(K(cid:0)k)roll-outswherethelearned modelcanbeconstruedasadeterministicpilotheuristic. Whenqueryinganexpertpolicy,thereisanabundanceofvaluableinformationthatcanbe utilisedforlearningnewmodels.Forinstance,it’spossibletoseekoutwhentheschedulingpro- cessismostsusceptibletofailure. Furthermore,generallystepwiseoptimality(orclassification accuracy)impliesgoodendperformance,hereminimisingthefinalmakespan. However,asthe impactofsuboptimalmovesisnotfullyunderstood,thenthemeasureneedstobeadjustedfor itsintendedtrajectory. Usingtheseguidelines,itbecomeseasiertocreatecustomdispatchingrulesforone’sparticular application. For this several different distributions of job-shop will be considered. Moreover, themachinelearningapproachisbasedonpreferencelearning,i.e.,whichpost-decisionstateis preferabletoanother. However,thatcouldeasilybesubstitutedforotherlearningmethodsor appliedtoothershop-constraintsorfamilyofschedulingproblemsthatarebasedoniteratively applyingdispatchingrules. v Niður,niður,niður!Ætlaðiþettaaldreiaðtakaenda?Hvaðskyldi éghafahrapaðmargakílómetra? Lísa Ágrip Tilerumargaraðferðirviðaðbúatilákvarðanareglurfyriráætlanagerð. Undanfariðhefuráher- slan í fræðunum verið á beina leit (t.d. gentíska bestun) eða gagnaþjálfun, en ein aðferð við þaðsíðarnefndaerstýrðurlærdómur. Íritgerðinniverðursúaðferðskoðuðnánarogsettfram líkan kallað Lærdómur ítrekunarreiknirita og samtakagreining algríma (LÍSA) um hvernig megi framkvæmaþessagreininguáskilvirkanmáta. Látumþjálfunargögninverafφ(x(k));y(k)gK 2 Dogritgerðinmunsýna: i)úrtökφ(x) i i k=1 i D þurfaaðveraísamræmiviðgagnadreifinguna semverðurunninúrhenni. Þettaergertmeð þvíaðuppfæralærðalíkaniðmeðvirkunámsferlibyggðuáeftirlíkingum;ii)y ermerktmeðþví i aðnotaendurgjöfsérfræðings(gertmeðbestun); iii)gögninþurfaaðveraíjafnvægi, þarsem gagnasettiðeríójafnvægimeðtillititilskrefsk;einnigiv)tilaðbetrumbætalýsinguánúverandi stöðuφ,erhægtaðnotaútspilunfyrirnæstu(K(cid:0)k)skref,þaðeraðendalokumákvarðanaferilsins. Þámátúlkalærðalíkaniðsemfyrirframákveðnaútspilunarreglu. Þegar sérfræðingur er spurður, verður til mikið af gagnlegum upplýsingum sem hægt er að nýtatilaðlæranýlíkön. Tilaðmyndaerhægtaðkomastaðþvíhvenæríákvarðanaferlinuer líklegastaðmistökeigisérstað.Yfirleittgefaháarlíkuráþvíaðbestaákvörðunsétekin(eðaþjál- funarnákvæmni)tilkynnagóðalokaframmistöðu,þ.e.íþessusamhengiaðlágmarkaheildartíma fyriralltákvarðanaferlið. Þarsemafleiðingarrangraákvarðanaeruekkialltafþekktar,þáerbetra aðuppfæramatiðmeðtillititilákvarðanatökunnarsjálfrar. Með þessari greiningu er einfaldara að búa til sérhæfðar ákvarðanareglur fyrir hverja nýja notkun. Í ritgerðinni verða skoðaðar nokkrar mismunandi tegundir af verkniðurröðun á vélar. Þaraðaukiverðurvélnámiðbyggtáákjósanlegribestun,þarsemgerðurergreinarmunuráþví hvaðastöðurerubetrikosturenaðrar. Ákjósanlegribestunværiþóhægtaðskiptaútfyriraðrar námsaðferðir,hægtværiaðbætaviðfleiriskorðumáverkefniðeðabeitasömunámsaðferðáaðra tegundafverkefnumafsvipuðumtoga. vi IthinkIshouldunderstandthatbetter,ifIhaditwrittendown:but Ican’tquitefollowitasyousayit. Alice Contents Listingoffigures xi Listingoftables xiii Listingofalgorithms xv Listingofpublications xvi Nomenclature xviii I Prologue 1 1 Introduction 3 1.1 Rice’sframeworkforalgorithmselection . . . . . . . . . . . . . . . . . . . . 4 1.2 Previouswork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 1.4 Supplementarymaterial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.5 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2 Job-shopSchedulingProblem 13 2.1 Mathematicalformulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.2 Constructionheuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.3 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.4 Singleprioritybaseddispatchingrules . . . . . . . . . . . . . . . . . . . . . . 21 2.5 Featuresforjob-shop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.6 Compositedispatchingrules . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.7 Rice’sframeworkforjob-shop . . . . . . . . . . . . . . . . . . . . . . . . . . 27 3 Problemgenerators 29 3.1 Job-shop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 vii 3.2 Flow-shop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 3.3 Benchmarkproblemsuite . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 4 Problemdifficulty 35 4.1 Distributiondifficulty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 4.2 Definingeasyversushardschedules . . . . . . . . . . . . . . . . . . . . . . . 39 4.3 Consistencyofprobleminstances . . . . . . . . . . . . . . . . . . . . . . . . 42 4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 5 EvolutionarySearch 47 5.1 Experimentalsetting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 5.2 Performancemeasures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 5.3 Experimentalstudy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 5.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 6 GeneratingTrainingData 57 6.1 Job-shoptreerepresentation . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 6.2 Labellingschedulesw.r.t.optimaldecisions . . . . . . . . . . . . . . . . . . . 59 6.3 Computationalgrowth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 6.4 Trajectorysamplingstrategies . . . . . . . . . . . . . . . . . . . . . . . . . . 60 7 AnalysingSolutions 65 7.1 Makingoptimaldecisions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 7.2 Makingsuboptimaldecisions . . . . . . . . . . . . . . . . . . . . . . . . . . 66 7.3 Optimalityofextremalfeatures . . . . . . . . . . . . . . . . . . . . . . . . . 69 7.4 Simpleblendeddispatchingrule . . . . . . . . . . . . . . . . . . . . . . . . . 77 7.5 Featureevolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 7.6 Emergenceofproblemdifficulty . . . . . . . . . . . . . . . . . . . . . . . . . 79 7.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 8 PreferenceLearning 87 8.1 Ordinalregressionforjob-shop . . . . . . . . . . . . . . . . . . . . . . . . . 87 8.2 Selectingpreferencepairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 8.3 Scalabilityofdispatchingrules . . . . . . . . . . . . . . . . . . . . . . . . . . 90 8.4 Rankingstrategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 8.5 Trajectorystrategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 8.6 Stepwisesamplingbias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 8.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 viii

Description:
I can't quite follow it as you say it. Alice. Contents. Listing of figures xi .. 2.3 Gantt chart of a partial JSP schedule 8.4 Main statistics for deviation from optimality, ρ, based on various stepwise sam- 11.2 Main statistics for {φi}24 .. “The only real mistake is the one from which we le
See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.