﻿
 UFDC Home myUFDC Home  |   Help
<%BANNER%>

# Combinational Approaches to Solve Scheduling Problems

## Material Information

Title: Combinational Approaches to Solve Scheduling Problems
Physical Description: 1 online resource (158 p.)
Language: english
Creator: Nemani, Ashish
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2009

## Subjects

Subjects / Keywords: column, combinatorial, curfew, decomposition, intermodal, load, location, negative, neighborhood, optimization, railroad, scheduling, vlsn
Industrial and Systems Engineering -- Dissertations, Academic -- UF
Genre: Industrial and Systems Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

## Notes

Abstract: In this dissertation, we discuss classical scheduling problems especially in transportation industries. All of these problems come from real life applications and are very critical from the academic as well as the financial point of view. Being NP-hard problems there doesn t exist any well-defined algorithm which can solve them efficiently with good running time, and thus rules of thumbs are still being followed in practice, with very myopic use of optimization procedures. These decisions are usually worth billions of dollars per year and even a slight improvement will have a significant economic impact. Through efforts described in this document, we try to develop some holistic approaches along with heuristics, to get efficient and effective results for these decision problems. This proposed work has the potential of implementation in commercial grade software. We first suggest some hybrid approaches to solve the intermodal load planning problem, which generate very effective solutions within minutes. In our second problem, subset-disjoint minimum cost cycle problem, we suggest several exact and heuristic approaches to find the minimum cost cycles which contains at most one node from any subset. These problems occur very often as a subproblem of other other combinatorial problems. In the third problem, location routing problem, we suggest column generation algorithm and show its effectiveness by doing experiments with benchmark problems. In our last problem, which is very critical in all railway industries, we propose several models based on based on mixed integer programming, heuristics, and other hybrid approaches. These algorithms show a significant improvement over the current practices.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Ashish Nemani.
Thesis: Thesis (Ph.D.)--University of Florida, 2009.

## Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2009
System ID: UFE0041090:00001

## Material Information

Title: Combinational Approaches to Solve Scheduling Problems
Physical Description: 1 online resource (158 p.)
Language: english
Creator: Nemani, Ashish
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2009

## Subjects

Subjects / Keywords: column, combinatorial, curfew, decomposition, intermodal, load, location, negative, neighborhood, optimization, railroad, scheduling, vlsn
Industrial and Systems Engineering -- Dissertations, Academic -- UF
Genre: Industrial and Systems Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

## Notes

Abstract: In this dissertation, we discuss classical scheduling problems especially in transportation industries. All of these problems come from real life applications and are very critical from the academic as well as the financial point of view. Being NP-hard problems there doesn t exist any well-defined algorithm which can solve them efficiently with good running time, and thus rules of thumbs are still being followed in practice, with very myopic use of optimization procedures. These decisions are usually worth billions of dollars per year and even a slight improvement will have a significant economic impact. Through efforts described in this document, we try to develop some holistic approaches along with heuristics, to get efficient and effective results for these decision problems. This proposed work has the potential of implementation in commercial grade software. We first suggest some hybrid approaches to solve the intermodal load planning problem, which generate very effective solutions within minutes. In our second problem, subset-disjoint minimum cost cycle problem, we suggest several exact and heuristic approaches to find the minimum cost cycles which contains at most one node from any subset. These problems occur very often as a subproblem of other other combinatorial problems. In the third problem, location routing problem, we suggest column generation algorithm and show its effectiveness by doing experiments with benchmark problems. In our last problem, which is very critical in all railway industries, we propose several models based on based on mixed integer programming, heuristics, and other hybrid approaches. These algorithms show a significant improvement over the current practices.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Ashish Nemani.
Thesis: Thesis (Ph.D.)--University of Florida, 2009.

## Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2009
System ID: UFE0041090:00001

Full Text

PAGE 2

2

PAGE 3

3

PAGE 4

PAGE 5

page ACKNOWLEDGMENTS .................................. 4 LISTOFTABLES ...................................... 8 LISTOFFIGURES ..................................... 9 ABSTRACT ......................................... 10 CHAPTER 1INTRODUCTION ................................... 12 2LOADPLANNINGPROBLEMATANINTERMODALRAILROADTERMINAL 18 2.1Introduction ................................... 18 2.2DenitionsandBackground .......................... 21 2.3ProblemDescription .............................. 24 2.3.1BasicOperations ............................ 25 2.3.2GeneralRulesofLoadingUnitsonRailcars ............. 26 2.3.3HitchUtilization ............................. 27 2.3.4AerodynamicEfciency ........................ 27 2.4ModelFormulation ............................... 28 2.5Heuristics:VeryLargeScaleNeighborhoodSearchAlgorithms ...... 33 2.5.1ConstructionHeuristics ........................ 35 2.5.2TheNeighborhoodSearchStructure ................. 35 2.5.3ImprovementGraph .......................... 36 2.5.4CyclicExchanges ............................ 37 2.5.5PathExchanges ............................ 38 2.5.6IdentifyingPathExchanges ...................... 38 2.6VeryLargeScaleNeighborhoodWithTabuSearch ............. 39 2.7HybridApproach ................................ 41 2.8ComputationalResults ............................. 41 2.8.1ComparisonofExactApproachandHeuristics ........... 42 2.8.2ResultsatDifferentStagesofHybridApproach ........... 44 2.9FlexibilityoftheModel ............................. 45 2.9.1LocalCarryCost ............................ 45 2.9.2HandlingTimeataTerminal ..................... 46 2.9.3DailyorWeeklyLoadPlan ...................... 46 2.10Conclusions ................................... 47 3SUBSETDISJOINTMINIMUMCOSTCYCLEDETECTION ........... 48 3.1Introduction ................................... 48 3.2NetworkReduction ............................... 52 3.3ExactAlgorithmsforSubsetDisjointMinimumCostCycles ........ 53 5

PAGE 6

............................ 54 3.3.2DynamicLabelingAlgorithm(DLA)forSDMCCProblems ..... 56 3.3.3AlgorithmAccelerationforDLA .................... 59 3.3.4DLAwithAccelerationforSDNCCProblems ............. 59 3.3.5FixedLabelingAlgorithm(FLA)forSDMCCProblems ....... 60 3.3.6All-PairslikePullAlgorithms(APPull) ................. 62 3.3.7All-PairslikePushAlgorithms(APPush) ............... 68 3.3.8StrategytoAvoidInefcientLabelExtensions ............ 68 3.4HeuristicAlgorithmsforSubsetDisjointMinimumCostCycles ....... 70 3.4.1LimitedUnprocessedLabelsHeuristics ............... 70 3.4.2LimitedCycleLengthHeuristics .................... 71 3.4.3LimitedCycleNumberHeuristics ................... 71 3.5ComputationalAnalysis ............................ 72 3.6Conclusions .................................. 85 4COLUMNGENERATIONAPPROACHFORLOCATIONROUTINGPROBLEMS 89 4.1Introduction ................................... 89 4.2ProblemDescription .............................. 90 4.3MathematicalModel .............................. 91 4.4ColumnGenerationAlgorithm ......................... 95 4.4.1TheMasterProblem .......................... 95 4.4.2ThePricingProblem .......................... 99 4.4.3TheAlgorithm .............................. 104 4.5ImplementationDetails ............................. 109 4.5.1TestProblems .............................. 109 4.5.2ComputationalPlatform ........................ 109 4.5.3ComputationalResults ......................... 110 4.6SummaryandConclusions .......................... 112 5SOLVINGTHECURFEWPLANNINGPROBLEM ................ 115 5.1Introduction ................................... 115 5.2ProblemDescription .............................. 118 5.3LiteratureReview ................................ 120 5.4Time-SpaceNetworkFormulation(TSNF) .................. 122 5.4.1MathematicalModel .......................... 124 5.4.2ComputationalAnalysis ........................ 127 5.5Duty-GenerationModel(DGM) ........................ 129 5.5.1Duty-GenerationPhase1:VariableReduction ............ 131 5.5.2Duty-GenerationPhase2:ProjectScheduling ............ 131 5.5.3Duty-GenerationPhase3:Crew-Scheduling ............. 133 5.5.4ComputationalAnalysis ........................ 134 5.6Column-GenerationModel(CGM) ...................... 135 5.6.1TheMasterProblem(MP) ....................... 136 5.6.2ThePricingProblem .......................... 138 6

PAGE 7

........................ 139 5.7Decomposition-BasedDutyGenerationModel ............... 141 5.8ComputationalResults ............................. 143 5.9Conclusions ................................... 147 6GENERALCONCLUSIONANDFUTURERESEARCH ............. 149 REFERENCES ....................................... 152 BIOGRAPHICALSKETCH ................................ 158 7

PAGE 8

Table page 2-1Examplesofrailcarsandcandidatepatterns .................... 24 2-2GapsatdifferenttimeintervalsbyIPAM. ...................... 43 2-3Comparativeanalysisofheuristics. ......................... 43 2-4Developmentofhybridapproach(timeinseconds). ................ 44 3-1ComparativeanalysisofexactSDMCCalgorithms:smallinstances. ...... 77 3-2ComparativeanalysisofexactSDMCCalgorithms:mediuminstances ..... 78 3-3ComparativeanalysisofexactSDMCCalgorithms:largeinstances ....... 79 3-4ComparativeanalysisofexactSDNCCalgorithms:smallinstances ....... 81 3-5ComparativeanalysisofexactSDNCCalgorithms:mediuminstances ..... 82 3-6ComparativeanalysisofexactSDNCCalgorithms:largeinstances ....... 83 3-7ComparativeanalysisofheuristicSDMCCalgorithms .............. 86 3-8ComparativeanalysisofheuristicSDNCCalgorithms .............. 87 4-1ConstraintsandassociateddualvaluesfortheMasterProblem ......... 100 4-2ResultsontheLRPbenchmarkinstances ..................... 110 4-3ComparisononPerl'sbenchmarkproblems .................... 111 4-4ComparisononbenchmarkproblemsthatarecompiledbyBarreto[ 11 ] .... 112 5-1Detailsofreal-lifeinstances ............................. 143 5-2Improvementinnumberofviolationsfor2007 ................... 145 5-3Improvementinnumberofviolationsfor2007 ................... 145 5-4Comparisonofalgorithms:optimalitygap(smallinstances) ........... 147 8

PAGE 9

Figure page 2-1Simpliedmodelofarailroadterminal. ....................... 22 2-2Patternsforarailcar. ................................. 25 2-3Flowofunitsatarailroadterminal. ......................... 26 2-450%hitchutilization. ................................. 27 2-5100%hitchutilization. ................................ 27 2-6NetworkforLPP. ................................... 29 2-7Nodescreationintheimprovementgraph ..................... 37 3-1Illustratingsubset-disjointminimumcostcyle. ................... 49 3-2Illustratingsubset-disjointnegativecostcyle. ................... 50 4-1ExampledatafortheLRP .............................. 91 4-2DynamicprogrammingalgorithmforESPPRC ................... 102 4-3Columngenerationalgorithm ............................ 106 5-1Space-Timenetworkforthecurfewplanningproblem. .............. 123 5-2VariablematrixD[p]forprojectp 129 5-3DynamicprogrammingalgorithmforESPPRC. .................. 140 5-4Algorithmframeworkofthecolumn-generationprocedure. ............ 141 9

PAGE 10

10

PAGE 11

11

PAGE 12

12

PAGE 13

13

PAGE 14

14

PAGE 15

15

PAGE 16

16

PAGE 17

2 dealswiththeintermodalloadschedulingproblem.Chapter 4 discussesthelocationroutingproblem.Chapter 3 dealswiththesubsetdisjointminimumornegativecycleproblems.Chapter 5 dealswiththecurfewplanningalgorithm.Finally,wegiveageneralconclusionandfutureworkinthe 17

PAGE 18

PAGE 19

PAGE 20

PAGE 21

1. Wegiveadetaileddescriptionoftheintermodalrailwayterminalandtheentities(Section 2.2 ). 2. Weexplaintherestrictionsonassigningcontainersortrailersonrailcars,theguidingcriteriafortheloadingprocess,aswellasothercriticalfactors(Section 2.3 ). 3. WeformulatetheloadplanningproblemasanMIPonanappropriatelydenednetworkwithgeneralizedowanddescribetheformulation(Section 2.4 ). 4. Weproposeaheuristic,single-unitmulti-exchangeneighborhood(SUMEN)search,whichgeneratesalocallyoptimalsolutionwithintwoseconds(Section 2.5 ). 5. Wedescribethesecondheuristic,multi-unitmulti-exchangeneighborhood(MUMEN)search,whichgeneratesnear-optimalsolutionsinaveryshorttime.Wethenimproveitbycombiningwiththetabusearch(MUMENT).WethensuggestauniquehybridprocedurethatproducessolutionsbetterthanthosegeneratedbyCPLEXinmorethanonehour(Section 2.7 ). 6. Weperformextensivecomputationalinvestigationsoftheexact,heuristics,andhybridapproachesandreporttheseresults(Section 2.8 ).Wealsodiscusstheapplicabilityofeachapproachinreal-lifeapplications. 7. Wedescribetheexibilityofourapproachesforsolvingdifferentvariationsoftheproblem(Section 2.9 ).Thisincludesdiscussingthedynamicversionoftheproblemandsuggestinghowitcanbesolvedbydirectlyextendingourapproach. 8. Finally,wepresentourconclusionsandsuggestdirectionsforfutureresearch(Section 2.10 ). 21

PAGE 22

PAGE 23

PAGE 24

PAGE 25

PAGE 26

PAGE 27

PAGE 28

PAGE 29

2-6 .Itcontainsthreesetsofnodes,describedasi,ii,andiii.(i)Therstsetcorrespondstotheunitsgroupedintofamilies.Alltheunitsofafamilyareofthesameshape,weight,anddimensions,buttheymaydifferinrevenueearnedorpenaltyincurredifdelayed.(ii)Thesecondsetbelongstoallthefeasiblesetsofpatterns,and(iii)thethirdsetbelongstoalltherailcars.Thearcbetweenaunit-nodeandapattern-nodeshowsthattheunitcanbeassignedinthatpatternatacertaincost.Anarcexistsbetweenapattern-nodeandarailcar-nodeiftheunitscanbeassignedontherailcarinthatpattern.Duetothepresenceofthesetwosetsofarcs,theowthroughthepattern-nodesisconservedandhence,thenetworkisageneralizedownetwork[ 5 ]. Figure2-6. NetworkforLPP. Thefollowingnotationisusedtodescribethemixedintegerprogrammingformulationoftheloadplanningproblem(LPP).Theproposedmodelwillbedenotedastheintegerprogrammingassignmentmodel(IPAM).Parameters:B:setofalltheunits. 29

PAGE 31

58 ].Constraint 2 isthemassbalancefortheinowandoutowofcontainersandtrailersatthepatternnodes.Constraints 2 and 2 enforcetheconditionthatarailcarcanbeassignedinonlyonepatternandaunitcanbeloadedinasinglepatternandthustoasinglerailcar,respectively.Constraints 2 and 2 ensurethatthe 31

PAGE 32

PAGE 33

PAGE 34

3 ]foranoverviewofVLSNsearchalgorithms.Thesealgorithmsaredesignedtosolvethepartitionproblemsasshownintheliterature.Ourproblemdoesn'tfallexactlyintothiscategoryduetotheaerodynamicfactorthatisbasedonthegapbetweenadjacentrailcars.Wehandlethisissuewhilendingthenegativecyclesbydynamicallyupdatingthearccosts.Thismodicationpreventsthebasicstructureofthealgorithmfromchanging,andletsusconsidertheLPPasaspecialcaseofthepartitionproblem,withthesetofrailcarsandunitspartitionedintomanysubsetsbasedonmanyrulesdiscussedlaterinthissection.Apartitionproblemistopartitionaset,sayS,intodifferentsubsets,sayS1,S2,S3,...,Sk,suchthatthecostofpartitionisminimum,wherethecostofpartitionisthesumofthecostofeachpart.ThepartitionproblemwassolvedusingVLSNbyThompsonandOrlin[ 71 ]andThompsonandPsaraftis[ 72 ].Ahujaetal.[ 6 ],[ 7 ]appliedthisapproachtothecapacitatedminimumspanningtreeproblemandtotheweapontargetassignmentproblem[ 4 ].WewillpresentabriefoverviewofthisapproachwhenitisappliedtotheLPP.Aneighborhoodsearchalgorithmstartswithafeasiblesolutionoftheoptimizationproblemandsuccessivelyimprovesitbyreplacingitwithanimprovedneighboruntilitreachesalocaloptimum.Thequalityofthelocallyoptimalsolutiondependsbothuponthequalityofthestartingfeasiblesolutionandthestructureoftheneighborhood; 34

PAGE 35

35

PAGE 36

PAGE 37

PAGE 38

38

PAGE 39

2.5.1 ,whosecomplexityisO(pnm*|I|).TheimprovementgraphG1(A)containsO(n+|B|)nodesandO((n+|B|)2)arcs,andthecostofallarcscanbecomputedinO((n+|B|)2)time.Weuseadynamicprogramming-basedalgorithmbyAhujaetal.[ 4 ]toobtainsubset-disjointcycles.Thisalgorithmrstlooksforprotabletwo-exchangesinvolvingtwonodesonly;ifnoprotabletwo-exchangeisfound,itlooksforprotablethree-exchangesinvolvingthreenodes,andsoon.Thealgorithmeitherndsaprotablemulti-exchangeorterminateswhenitisunabletondamulti-exchangeinvolvingtnodes(wesett=5).Intheformercase,weimprovethecurrentsolution,andinthelattercasewedeclarethecurrentsolutiontobelocallyoptimalandstop.TherunningtimeofthedynamicprogrammingalgorithmisO((n+|B|)2*2t)periteration,anditistypicallymuchfastersincemostcyclicexchangesfoundbythealgorithmaresimpleswapping.TheVLSNsearchheuristics,particularlyMUMEN,generatedgood-qualitysolutionswithin30secondsformostoftheinstances(seeSection 2.8 ).Forsomeinstances,thegapsarehighindicatingthattheVLSNsearchgetstrappedinthelocaloptima.Toovercomethis,weperturbthesolutionandallowsomeunprotableexchanges.Weusethetabusearchconceptsforthesemodicationsasdescribedintheupcomingsection. 2.8 ).Theoptimalitygapsaregoodandcanbeimprovedfurtherifmoretimeisavailable.Thenatureoftheconstraintsrendersthefeasibleregionhighlyscattered,andthuseventheverylarge-scaleneighborhoodsearchcanconvergetoa 39

PAGE 40

PAGE 41

PAGE 42

2.5 ,weperformedsomeexperimentsontherelaxedloadplanningproblemobtainedbyignoringtheaerodynamicfactor.Weobservedthatthisrestrictionhasanegligibleeffectonthequalityofthesolution.Outofthe15instancesinvestigated,onlyonegeneratedaslightlybetterresult.Thisexperimentsuggeststhattherestrictiondoesn'treducetheneighborhoodsizesignicantlyenoughtoworsenthesolution. 2-2 .Weobservedthatthesolutionqualitycurveissharpintherstveminutes,butitsslopekeepsdecreasingastimepasses,thereforejustifyingourterminationcriterionofthreeminutesinthehybridapproach.Next,wecomparedthesolutionqualityofdifferentheuristics:SUMEN,MUMEN,andMUMENTinTable 2-3 .WeobservedthatalthoughtheSUMENsearch 42

PAGE 43

Table2-2. GapsatdifferenttimeintervalsbyIPAM. 803200.00%0.00%0.00%0.00%0.00%1.6mins3100.00%0.00%0.00%0.00%0.00%2.6mins3000.00%0.00%0.00%0.00%0.00%5.3mins2900.03%0.00%0.00%0.00%0.00%4.4mins2800.00%0.00%0.00%0.00%0.00%3.4mins2701.29%0.00%0.00%0.00%0.00%5.9mins2600.21%0.00%0.00%0.00%0.00%4.6mins250NoSol.9.40%9.40%9.40%0.00%31.4mins240NoSol.NoSol.29.22%29.22%0.00%23.4mins230NoSol.24.98%24.98%24.98%0.00%47.7mins220NoSol.NoSol.NoSol.5.27%0.00%27.0mins210NoSol.NoSol.NoSol.3.28%1.18%1,172.7mins200NoSol.NoSol.NoSol.NoSol.3.84%2,291.2mins195NoSol.NoSol.91.28%91.28%9.15%1,237.7mins190NoSol.NoSol.93.57%4.93%2.36%1,892.0mins18524.36%24.04%24.04%15.73%10.57%1,264.0mins18059.24%59.24%59.24%59.24%38.90%1,546.7mins Comparativeanalysisofheuristics. GapTime(Secs)GapTime(Secs)GapTime(Secs) 8032010.77%24.72%134.72%1831010.49%13.82%123.82%1530012.36%14.74%154.74%2129012.12%12.38%152.38%1928013.66%16.00%94.06%1627014.70%14.90%104.90%1426018.38%27.91%104.77%2125017.17%13.08%93.08%1424021.51%26.13%113.10%1923023.84%12.72%102.72%1722027.61%14.81%81.32%2121027.02%12.67%72.67%1420017.96%13.08%53.08%919524.61%13.31%43.31%819026.14%15.51%54.48%818510.77%26.49%45.09%1218010.49%23.65%43.65%7

PAGE 44

Developmentofhybridapproach(timeinseconds). GapTimeGapTimeGapTime 803200.00%970.00%510.00%1013100.00%1580.00%790.00%793000.00%1800.00%1920.00%1922900.03%1800.00%880.00%882800.00%1800.00%1110.00%1112701.29%1800.00%760.00%762600.21%1800.00%980.00%98250NoSol.1801.51%2130.08%231240NoSol.1800.00%2260.00%226230NoSol.1801.91%2120.29%222220NoSol.1801.01%2010.61%209210NoSol.1801.70%1970.50%215200NoSol.1801.88%1950.71%208195NoSol.1802.20%1980.96%206190NoSol.1804.97%1931.18%19918524.36%1804.81%2071.32%21318059.24%1801.89%1990.07%205 2-4 andderivedtheobservationsbelowfromit.(i)TheIPAMtakesmuchlesstimetondanoptimalsolutionwhenthenumberofunitsishigherthanthetotalcapacityofthetrain.Thisisbecausewhentherearelargenumbersofunitsofdifferenttypes,eachrailcarcanbeassignedinthemostoptimalway,sotherewillbefewerbranchesinthebranch-and-boundtree.However,asthenumberofunitsnearsthetotalcapacityofthetrain,thenumberofbrancheswillincreaseinordertondthebesttoftheunitstotherailcars.ThequalityofthesolutiongeneratedbyIPAMinthetimetakenbyheuristicswasvariable,andinmostcasesitwasnotbetterthanheuristics.Itimprovedwiththetimeasmorebrancheswereexamined.(ii)TheMUMENTsearchheuristicgavereasonablygoodsolutionsforallinstances,andwhenwecompareditwiththesolutionofIPAMgeneratedwithinthesametime-interval,thedifferencebecameprettysignicant.(iii)Intuitively,agoodstartingpointshouldimprovethesolutionqualityaswellastherun-timeforabranch-and-boundalgorithm,butthiswasnottrueforall 44

PAGE 45

PAGE 46

PAGE 47

47

PAGE 48

3-1 .ThegraphinFigure 3-1 hasnode1insubset1,node2insubset2,node3insubset3,andnodes4and5insubset4.TheSDMCC,whichistheoptimalsolutionofSDMCC 48

PAGE 49

Figure3-1. Illustratingsubset-disjointminimumcostcyle. WeillustratethesubsetdisjointnegativecostcycleinFigure 3-2 .WekeepthenodepartitionssameasinFigure 3-1 ,andchangeonlythecostofarcstosimplifytheillustration.Inthisnetwork,wehavetwofeasiblenegativecycles:(i)(1-2-3-1)ofcost-4,and(ii)(1-2-5-3-1)ofcost-10.Therefore,theoptimalcycleofSDNCCproblemis(1-2-3-5-1).Thereexistsonemorenegativecycle(1-2-4-5-3-1)ofcost-12,butweeliminateitfromconsiderationasitnotsubset-disjoint.Itshouldbenotedthatinthisexampletheminimumissameasthenegativecycle(1-2-3-5-1).Fromthegivendenitionsandillustrations,wecaneasilyconcludefollowingtworelationsbetweenSDMCCproblemandSDNCCproblem: 1. ThefeasiblesetofnegativecycleproblemwillbeasubsetofthefeasiblesetofminimumcycleprobleminthesamegraphG. 2. Ifthereexistsafeasiblesolutionforthenegativecycleproblem,theoptimalsolutionofminimumcycleandnegativecyclewillbesame. 49

PAGE 50

Illustratingsubset-disjointnegativecostcyle. Thesubset-disjointminimumandnegativecostcycleproblemsareusuallythecorestructureofmostoftheroutingbasedoptimizationproblems,andmultiple-exchangebasedneighborhoodsearchalgorithms.IthasbeenprovedNP-completein[ 33 ].Wenextdiscusssomeofitsdirectandindirectapplications. 1. Prizecollectingtravellingsalesmanproblem[ 8 ]Asalesmantravelsbetweenpairsofcitiesatacostdependingonlyonthepair,andgetsaprizeineverycitythathevisits.However,heneedstopayapositivepenaltyforeverycitythathefailstovisit.Theobjectiveoftheproblemistomaximizehisearning(prizecollected-travelcosts-penalties).Thereisnoconstraintontheminimumreward.WecanreducethisproblemtotheSDMCCproblembyconsideringeachcityasnode.Wealsoassumethateachnodehasitsownsubset,i.e.eachsubsetcontainsonlyonenode.Arcs(i,j)2Aareconstructedbetweennodepair(i,j),ifthesalesmancantraveltocityjaftercityi.Thecostofeacharc(i,j)2A,isassignedasthesumofthreecostcomponents:(i)travelcostfromcityitocityj,(ii)negativeofthepenaltyofnotvisitingcityi,and(iii)prizecollectedforvisitingcityj.Aftertheconstructionofthenetwork,weapplytheSDMCCalgorithmstondtheminimumcyclewhichhasonetoonecorrespondencetotheoptimalsolutionoftheprizecollectingtravellingsalesmanproblem. 2. Verylargescaleneighborhoodexchangesearch[ 6 ]Theverylargescaleneighborhoodsearchisgenerallyusedinthecontextofsolvingpartitioningproblems,whicharehighlycombinatorialandhugesizeproblems.GivenasetSofnelements,theobjectivetondapartitionT=fT1,T2,...,TkgofS,whereTissetofsubsetsofS.Eachpairofsubsetsmustbemutuallyexclusive,andtheunion 50

PAGE 51

2 .AlthoughSDMCCorSDNCCproblemsareverycommonlyuseddirectlyorindirectlyinseveralalgorithms,wedon'thaveanyliteraturededicatedtothem.Generally,somedynamicprogrammingisusedwhichiscustomizedtoparticularproblemsandmaynotbeapplicabletootherclassofproblems.Inthischapter,wehavedevelopedseveralgenericalgorithmswhichcanbeappliedinallproblemswhereminimumcycleistobefoundrepetitively.Thecontributionsofthispaperinclude: 1. WesuggestfourexactalgorithmsfortheSDMCCproblemsandSDNCCproblems.Thesealgorithmscanbeseenasthegeneralizedversionofthedynamicprogrammingalgorithmsgenerallyusedfortheconstrainedshortestpathproblems. 2. Networkreduction:Thecomplexityofexactalgorithmsdependsonthesizeofnetwork(nodesandarcs).Weintroduceseveralstrategiestoreducethesizeofthenetworkandthereby,todecreasethecomplexityofthealgorithm. 51

PAGE 52

Algorithmacceleration:Wepresentsomenewandsomeexistingstrategiesusedinotherapplications,toimprovethepracticalrunningtimeofthealgorithms.Thesestrategiesdon'timprovetheworstcasecomplexitiesofalgorithms;however,theyprovetobeveryeffectiveinpracticalapplications. 4. Theexactalgorithmsmayperformveryefcientlyevenonlargeinstances,thesolutiontimemustbeevenlowerthanfractionofsecondsinsomealgorithmssuchasVLSNsearch.Weproposeseveralheuristicsbasedontheexactalgorithmsbycontrollingdifferentparametersoftheexactalgorithms. 5. Wenallygiveanextensivecomputationalanalysisbyrunningallalgorithmsondifferentclassesofproblems. 1. Subsetofnodei=Subsetofnodej. 2. Nodeihasnoincomingarc. 3. Nodejhasnooutgoingarc. 4. Forallk2Vsuchthat(j,k)2A^(i=k_s(k)6=s(i))^s(k)6=s(j)thereexistsaq2Ss(j),q6=jsuchthat(q,k)2A,(i,q)2Aandciq+cqk
PAGE 53

53

PAGE 54

54

PAGE 55

55

PAGE 56

3.3.1 ,andremovesitfromthelist.not_dominated(l0,L(l0)):Itreturnstrueifthelabell0isnotdominatedbyanyofthelabelsintheset(L(l0)).remove_dominated(l0,L(l0),U):Ifthefunctionl0isnotdominatedbyanylabelpresentinthesetoflabelswiththesameendnode,itispossible 56

PAGE 57

57

PAGE 58

PAGE 59

3.3.4DLAwithAccelerationforSDNCCProblemsThedynamiclabelingalgorithmfortheSDNCCproblemisalmostsameasthatfortheSDMCCproblems.However,thealgorithmcanbeacceleratedatseverallevelsbyusingthepropertyspeciconlytonegativecycles.Property1:Foranynegativecostcycleq=(i1,i2,...,il,i1)oflengthlinG,thereexistsj21,2,...,lsuchthatc(pk)<08k=1,2,l,wherepk=(ij,ij+1,...,ij+k),andallsubscriptsarithmeticisperformedmodulol.Usingtheaboveproperty,wecanconcludethatforanyoptimalsolutionofSDNCCproblem,theremustexistanodeisuchthateverypathinthesubgraphofGobtained 59

PAGE 60

3.3.3 ,and(ii)describedusingproperty1.Unfortunately,wecannotusebothtechniquesatthesametime.Inrstpropertyweassumethatallcycleswithlowernumberednodesmusthavebeenfoundwhenlowernumberednodeswasconsideredassource.While,ifweusethesecondtechnique,somecyclemaystartfromthelargernodeandpassthroughasmallernode,whichdoesn'thaveanyoutgoingarcwithnegativecost,tocompleteanegativecostcycle.Generally,anetworkcontainsverysmallpercentageofnegativecostarcs,soifweusethesecondaccelerationtechnique,wewilleliminatemanysourcenodesandreducethenumberoflabelscreatedwhilelabelextensionprocess.However,ifthepercentageofnegativearcsisveryhigh,weshouldusetherstaccelerationtechniqueasnegativecostcriteriawouldbesatisedbyalmostallnodes.Theworst-casecomplexityofDLAforSDMCCandSDNCCproblemsremainsame.Wenextgivethedetailsofthealgorithmforcompleteness. 3.3.2 fortheDLA,however,usesdifferentstrategytostorethelabels.Inthexedlabelalgorithm,likeDLA,wemaintainasetofunprocessedlabelsU,butunlikeDLA,wemaintainthesetLiofnon-dominated,andprocessedlabelsateverynode 60

PAGE 61

61

PAGE 62

3.3.3 isvalidforFLA,too.Weincludetheminitsalgorithmicframework.ThecorrespondingalgorithmforSDNCCproblemscanbeobtainedbymakingsamechangesaswedidwhileconvertingtheDLAofSDMCCproblemtoSDNCCproblem.Moreexplicity,wereplacetheline3byforall2Nwithci<0;i2Ndo;andline10byforall(i,j)2+((l))ANDc(l)+cij<0do.ThecomplexityofDLAandFLAforbothversionsoftheproblemremainssame.However,therearetwomaindifferencesbetweenthesetwoalgorithms:(i)numberoflabelsstoredinDLAapproachismuchhigherthanthatofFLA,and(ii)thenumberofchecksperformedforestablishingthedominanceoflabelsiscomparativelyhigherinFLA.WediscusstheeffectofthesedifferencesinthecomputationalanalysisinSection 3.5 62

PAGE 63

5 ].ForeachsourcenodeiinN,werstndallpathsoflength1,thenextendthemtopathoflength2,andsoontillwereachthemaximumpathlengthn.Itisnotatruegeneralizationasinouralgorithms,wegenerateallpathsoflengthkonlyfromthecurrentsourcei2Nandmovetondingpathsoflengthk+1.However,inactual"all-pairshortestpathdynamicalgorithms"all 63

PAGE 64

64

PAGE 65

3.3.3 .TheAPPullalgorithmscanbeeasilymodiedtosolvetheSDNCCversionoftheproblem.Todifferentiatethesealgorithmswecallthem,APPull-MandAPPull-Nalgorithms,respectively.TheAPPull-Nalgorithmsusetheaccelerationstrategyspecicfornegativecyclesassumingthatthepercentageofnegativecostarcsinthenetworkisverylow.Theextend-possiblefunctioninline13,examinesiftheextendedlabel'scostremainsnegative,besidescheckingtheresourcevector.Wepresentthedetailsofthealgorithmforcompletenessofthediscussion.TheAPPullalgorithmscanbeconvertedtoapuregeneralizedversionofallpairshortestpathalgorithms,byexchangingthe"forloop"ofline3andline14.However,pureversionstoresverylargenumberoflabels,andslowsdownthecompletealgorithm.TheAPPullalgorithmreducesthememoryrequirementbystoringfewerlabels,andgeneratessolutionsinmuchlowerrunningtime.WeshowthedetailedcomputationalresultsinSection 3.5 65

PAGE 68

PAGE 70

PAGE 71

PAGE 72

3.5 31 ].ThesebenchmarkproblemsweregeneratedbyconsideringdifferentinstancesofthecapacitatedminimumspanningtreeproblemsinpaperbyAhujaetal.[ 6 ].Thepapersolvesthecapacitatedminimumspanningtreeproblemsusingvery-largescaleneighborhoodsearch,inwhichthesubset-disjointnegativecycleproblemissolvedasthesubproblem.Wetakeeachinstanceofthespanningtreeproblemfrom[ 6 ],andgroupallitssubproblems.Thesegroupsofsubproblemsserveastheinputforcomparingouralgorithms.Theinstancesforcapacitatedminimumspanningtreeproblemsweregeneratedintwoways: 72

PAGE 73

PAGE 74

3-1 ,forthemediumsizeproblemsinTable 3-2 ,andforthelargesizeproblemsinTable 3-3 .Inallthetables"P"denotesproblem-groups,"Sets"denotesthenumberofsubsetsinthegraph,"Nodes"denotesthenumberofnodesinthegraph,"Arcs"denotesthearcdensityinpercentage,"Neg.Arcs"denotesthedensityofnegativearcsinthegraph,"AML"denotestheaveragemaximumnumberoflabelsstoredbyalgorithms,and"time"denotesthetimeconsumedinseconds.InTable 3-1 ,wecompareallalgorithmsonthesmallsizeinstances.Theproblemshavebeensortedindecreasingorderofnumberofsubsets,andforthesamenumberofsubsets,intheincreasingorderofnumberofnodes.Inalltheseinstances,fromtimeaswellasmemoryperspective,thedynamiclabelingalgorithmperformedmuchbetterthanthexedlabelingalgorithm.For17outof20instances,thetimerequiredbyDLAwasalmosthalfofthatrequiredbyFLA.TheaveragemaximumnumberoflabelsstoredbyFLAwas1.5-2timesofaveragemaintainedbyDLA.ItisnotveryintuitiveintherstglanceasintheFLA,westorelabelsforeachnode,onlyifitiscertainthatthelabelwouldneverberemovedthroughdominancetest.Inotherwords,ateachnode,wemaintainthelistofprocessedlabelswhicharexed.Duringthelabelextensionstep,weinsertallnewlycreatedlabels,onlyinthesetofunprocessedlabels.Wetakeoutlabelsonebyonefromthislistandcheckifitisdominated.Itshouldbenotedthat,whenthelabelisinsertedintotheunprocessedlabelset,itisnotdominatedanyofthelabelxedatthecorrespondingnode.However,bythetimethatlabelisremovedfromtheunprocessedlabelsetforprocessing,itmaybedominatedbythenewlabelscreatedandxedatthatnode.However,intheDLA,thelabelsstoredinsetoflabelsmaintainedateachnodemaygetdominatedbysomefuturelabels,andhence,mayberemoved.Intheremovaloperation,alllabelsfromtheunprocessedlabel-setaswellasfromthenode-specic-label-set,whicharedominatedbythenewlycreatedlabelareremoved.So,althoughwemayhavetoperformlabelremovalandlabelentrystepsmultipletime, 74

PAGE 75

3-2 comparesthesamesetofalgorithmsonproblemclassesofmediumsize.Weconsiderthesamesetofnodes,andincreasethenumberofsubsets(14-17)inwhichtheseareclassied.Thecomparativebehaviorofallfourexactalgorithmsisalmostsameaswehavediscussedforproblemgroupsofsmallsize.ThedifferenceinDLAandFLAwithregardtothenumberoflabelsstoredaswellassolutiontime,becomesmoreprominentastheabsolutedifferenceisveryhigh(althoughdifferenceinratioissame).Allbutfourproblemsweresolvedwithinfractionofseconds.Thesefourgroupswerelargest(199)withrespecttothenumberofnodes.ThecomparativeperformanceofAPPullalgorithmwasmuchbetterinthisproblemclassthanthesmallerclass.APPullalgorithmperformedbetterthanAPPushinveoutof20problem-groups, 75

PAGE 76

3-3 displaystheresultsforthelargeproblemgroups.Thenumberofnodesvaryfrom80to199,andnumberofsubsetsfrom17to21.Thearcdensityisalsohigherfromlasttwoproblemclasses.comparesthesamesetofalgorithmsonproblemclassesofmediumsize.Inthiscasetoo,weobservethesamepatternofperformanceamongfourexactalgorithms.However,thedifferenceinsolutionbetween"DLA-FLA",and"APPull-APPush"algorithmswasverysignicant.Forexample,oneoftheproblem-grouptook137.36secondsforDLA,255.88secondsforFLA,20.97secondsforAPPull,and32.77secondsforAPPush.ItclearlyshowstheefciencyofAPPullandAPPushalgorithms.AmongAPPullandAPPushalgorithms,theAPPullalgorithmsperformedbetterthanAPPushinallproblem-groups.ThisbehaviorisquiteinterestingasforsmallinstancesAPPushalgorithmshasbetterrunningtime.So,wecanconcludethatAPPullalgorithmsarebestfortheseclassesofproblems.However,eventhemostefcientalgorithmtookmorethan20secondsfortwoproblem-groups,andmorethan10secondsforonegroup.FormostofthoseproblemswhichsolveSDMCCrepetitivelyassub-problems,thisrunningtimeisunacceptable.So,wedevelopedsomeheuristics,andwediscusstheseresultsaftertheanalysisofSDNCCalgorithms.WediscusstheresultsofSDNCCproblemsforproblem-groupsofsmallsizeinTable 3-4 .Allalgorithmsperformedveryefcientlyinalltheseinstancesgeneratingoptimumresultsinlessthan0.07seconds.Thememoryrequirementwasalsoverylowstoringatmost1300labels.ThepatternbetweenDLAandFLAissameasitwasintheSDMCCproblemasshowninTable 3-1 .AmongAPPullandAPPushalgorithms,APPushperformedbetterinallinstancesfromtimeperspective.However,thepatternbetweenDLAandAPPullalgorithmwasoppositeofSDMCCresults.TheDLAperformedconsistentlybetterthanAPPullalgorithmsfromtimeperspective,andthenumberoflabelsstoredwereinthesamerange(slightlymoreinDLA).Therefore, 76

PAGE 77

ComparativeanalysisofexactSDMCCalgorithms:smallinstances. AMLTimeAMLTimeAMLTimeAMLTime 33-12139969.812.644,0260.257,0580.553,8200.183,8200.1635-812,139967.822.833,2740.205,0560.443,2170.163,2170.1330-15129975.013.683,6500.246,3490.553,5070.173,5070.1527-9129969.282.792,8500.184,4750.402,7450.142,7450.1239-161119962.281.415,2060.748,8081.704,8800.734,8800.5942-71119968.251.587,5801.1913,5262.807,2671.037,2670.8445-111119975.251.247,3241.0313,4872.577,0780.897,0780.7448-141119976.251.105,7910.9510,3282.305,5820.845,5820.682-1110,114085.959.597310.029870.036440.026440.0011-810,114085.969.705260.027980.025000.015000.015-6104085.719.996880.038100.026600.026600.008-11104085.7111.478140.031,0010.027430.017430.0118-14108083.725.821,2720.041,7970.091,1300.041,1300.0324-14108083.728.698640.021,3850.067810.027810.0215-12108081.925.907450.021,1960.046770.026770.0221-11108080.305.515280.018020.034830.024830.013-584081.409.892830.013520.022540.002540.016-984081.408.703750.014490.013080.003080.009-118,94081.7811.874650.015990.014090.014090.0112-77,84078.3410.091880.012440.011630.001630.00

PAGE 78

ComparativeanalysisofexactSDMCCalgorithms:mediuminstances AMLTimeAMLTimeAMLTimeAMLTime 14-1916,178090.387.049,1450.4718,9891.149,0460.189,0460.2023-1616,178090.397.7816,7110.5422,3161.0515,9920.2015,9920.2320-16168090.367.3912,5570.4423,4430.9212,4920.1712,4920.1832-2315,169971.113.7910,0310.4516,5331.029,5720.259,5720.2526-7159970.484.978,3700.6113,4791.238,3230.308,3230.3038-1714,1519964.842.0826,1523.8252,8658.8925,1672.0225,1672.0329-14149965.954.579,0910.5814,7891.238,9060.308,9060.2941-71419974.091.9912,3562.4425,1725.8912,1161.6412,1161.5044-101419972.791.5322,7155.5340,04011.8522,0992.7022,0992.8047-61419976.491.5915,3543.0231,8407.8015,0861.8315,0861.721-5144090.7111.902,6840.074,3470.112,6220.032,6220.034-4144090.4812.552,1240.053,1740.092,0710.032,0710.037-10144090.4817.251,9850.053,1390.091,9610.021,9610.0210-6144090.4814.771,6340.033,1620.071,5950.041,5950.02

PAGE 79

ComparativeanalysisofexactSDMCCalgorithms:largeinstances AMLTimeAMLTimeAMLTimeAMLTime 13-11218092.879.2415,1721.0635,7392.4815,0940.3415,0940.4019-27218092.929.1937,3512.2161,0554.1435,1260.6535,1260.8316-920,218092.849.6121,3501.5738,1343.1220,7720.4620,7720.5625-5319,219977.446.02137,25672.18219,269117.29136,5486.15136,54812.1722-11208092.6810.4650,6923.2081,2805.6750,2250.8050,2251.0331-17199969.534.4923,5352.1941,2014.4223,2170.7623,2170.9134-1818,199971.814.1318,2651.5233,4203.2617,6050.5717,6050.6528-2518,199966.755.8251,3365.8879,14010.6051,0421.4951,0421.9346-211819973.512.65140,916137.36233,972255.88139,41620.97139,41632.7743-221819976.582.1492,28147.42131,57987.2991,71110.4591,71114.6637-171819973.532.6537,5236.9087,11617.9736,4753.0036,4753.2740-5417,1819971.533.47171,963132.81283,052264.11170,94120.95170,94134.25

PAGE 80

3-5 ,weshowtheoutputofallalgorithmsonproblems-groupsofintermediatesize.Allthealgorithmswereveryeffectiveintheseclassesofproblemsaswellsolving12outof14groupswithin0.015secondsandtwogroupswithin0.15seconds.ThecomparativebehaviorofDLAandFLAremainssameasforallothergroups.TheAPPushalgorithmsfoundoptimalcyclesmuchquickerthan(almost1.5times)APPullalgorithmsforallinstances.However,amongDLAandAPPushalgorithms,therewasnocleardifferenceinperformance.Ineightoutof14problem-groupsDLAwasquickerthanAPPushalgorithms,whileinvegroupsAPPushalgorithmsoutperformedDLA.Therewastieinonegroup,whichshouldbecountedtowardsAPPushalgorithmsifmemoryrequirementisourslightestconcern.ThenumberoflabelsstoredwasconsistentlylowerinAPPushalgorithms.InTable 3-6 ,weanalysetheperformanceofeachexactalgorithmforSDNCCproblemsonlargesizeinstances.TherewasnochangeinthepatternofDLAandFLAfromourearlieranalysis.TheAPPushalgorithmsagainoutperformedAPPullalgorithmsinallproblemgroups.ThemostinterestingobservationwastheshiftinthebehaviorofDLAandAPPushalgorithms.Intheseproblemgroups,DLAwascompletelyovershadowedbyAPPushalgorithmsandperformedbetterinonlytwooutof14instances.Inthosetwoinstancestoo,thetimedifferencewasnotverysignicant,withAPPushalgorithmstoringmuchlessnumberoflabels.Inthemostdifcultinstanceofthisclass,DLAalgorithmtook24.5secondstondtheoptimalresults,however,theAPPushalgorithmgenerateditin5.4seconds.Itshowsthatiftheproblemsizeisnotsmalltomediumrange,DLAisabetterapproachbutifthesizeisverylarge,APPushalgorithmsperformthebest.WenowtesttheheuristicsapproachessuggestedinSection 3.4 ,whichcontrolthreeparametersoftheexactalgorithms:(i)Numberofunprocessedlables,(ii) 80

PAGE 81

ComparativeanalysisofexactSDNCCalgorithms:smallinstances AMLTimeAMLTimeAMLTimeAMLTime 33-12139969.812.641,0210.0171,1620.0309430.0349430.02235-812,139967.822.835480.0126320.0155260.0255260.01830-15129975.013.685460.0156160.0225120.0255120.01927-9129969.282.791860.0122040.0191650.0261650.01739-161119962.281.414770.0625440.0594070.1444070.10742-71119968.251.581,2330.0831,5390.0871,1510.1721,1510.12745-11111991991.241,3690.0771,7740.0901,2640.1511,2640.11248-141119976.251.11,0590.0741,3020.0799850.1419850.1112-1110,114085.959.592470.0002730.0042260.0012260.00111-810,114085.969.71610.0021590.0061490.0041490.0025-6104085.719.991320.0031190.0051150.0081150.0038-11104085.7111.472620.0012660.0092390.0092390.00318-14108083.725.824460.0063990.0093700.0083700.00724-14108083.728.693990.0035040.0083290.0103290.00815-12108081.925.92300.0062500.0071940.0091940.00721-11108080.35.513410.0043310.0043100.0103100.0073-584081.49.89920.003980.000840.006840.0006-984081.48.7960.000920.005770.003770.0029-118,94081.7811.872340.0042510.0071900.0061900.00112-77,84078.3410.09890.002850.000780.002780.000

PAGE 82

ComparativeanalysisofexactSDNCCalgorithms:mediuminstances AMLTimeAMLTimeAMLTimeAMLTime 14-1916,178090.387.042,1710.0203,3810.0392,0210.0322,0210.01923-1616,178090.397.782,1630.0222,2560.0262,1200.0262,1200.02120-16168090.367.391,5200.0181,8340.0261,4780.0331,4780.01732-2315,169971.113.791,3080.0181,4660.0271,2360.0371,2360.02326-7159970.484.975250.0185070.0224970.0404970.02238-1714,1519964.842.083,4850.1286,4210.1953,2860.2733,2860.16729-14149965.954.572,7540.0263,1620.0452,6630.0472,6630.03041-71419974.091.991,0070.0921,1500.0949150.2359150.14744-101419972.791.533,4800.1594,3000.2223,3470.2803,3470.18647-61419976.491.591,9940.0942,5580.1041,8720.2061,8720.1431-5144090.7111.96050.0036030.0095750.0095750.0064-4144090.4812.555240.0045130.0125070.0195070.0047-10144090.4817.251,0980.0131,1570.0281,0660.0091,0660.00610-6144090.4814.777030.0138590.0266840.0056840.000

PAGE 83

ComparativeanalysisofexactSDNCCalgorithms:largeinstances AMLTimeAMLTimeAMLTimeAMLTime 13-11218092.879.242,2990.0262,4320.0402,2640.0362,2640.02319-27218092.929.1911,1710.17711,3520.25210,7480.09410,7480.08216-920,218092.849.611,9560.0232,3750.0311,8610.0261,8610.01925-5319,219977.446.0240,0115.79159,7599.67739,6430.89339,6431.21822-11208092.6810.464,7970.0576,9830.0994,5290.0484,5290.04031-17199969.534.492,9210.0463,3180.0652,8530.0622,8530.03834-1818,199971.814.131,5950.0281,8990.0541,5330.0511,5330.03528-2518,199966.755.8221,1740.63224,7461.01720,9200.21420,9200.22046-211819973.512.6525,2901.72836,0192.89324,9290.86424,9290.78043-221819976.582.1416,2050.78618,7781.01616,0430.50916,0430.42937-171819973.532.652,4300.1343,1660.1522,3200.3202,3200.19340-5417,1819971.533.47133,62124.586185,56337.862132,9093.064132,9095.485

PAGE 84

3-7 ,and 3-8 ,"P"denotestheproblem-groups,"Opt."denotestheoptimalsolution,"Dev."denotesthedeviationofheuristicsolutionfromtheoptimal,and"Time"denotestherunningtimeinseconds.FortheLULheuristics,wexedthelimitofunprocessedlabelsto10,50,and1000forAPPullandAPPushalgorithms.Thesolutionqualityandrunningtime,asexpectedimprovedwithincreaseinthenumberoflabelsallowed.Asweincreasedthelabellimitfrom10to1000,thegapbetweentheheuristicsolutionandoptimalsolutionimprovedfrom30%toalmost1%whiletimeincreasedfromonesecondtothreesecond.Inthemostdifcultinstance,thetimetakentogenerateoptimalsolution,forAPPullandAPPushalgorithms,was20secondsand34secondswhichimprovedto3.74secondsand3.60secondsrespectively.Theoptimalitygapwasaround2%.Bothgeneratedoptimalsolutionforoneproblem-group.AmongAPPullandAPPushalgorithmswithLUL,APPullgeneratedsolutionswithbettergapwhileAPPushwasfasterspeciallywithlowernumberoflabels.ForLCL,wevariedthelimitonthemaximumcyclelengthto2,4,and8.BothAPPullandAPPushversionofalgorithmsgeneratedpoorsolutionswithcyclelengthtwo,whichdenotesasimpleswapping.Itisduetothefactthataverageoptimalcyclelengthsinallproblemsgroupvaryfromeightto10.Withthecyclelengthlimit8,therunningtimewasveryhighconsideringthatitisaheuristic.TheLULoutperformedLCLheuristicsinsolutionqualityaswellasrunningtime.Finally,we 84

PAGE 85

85

PAGE 86

ComparativeanalysisofheuristicSDMCCalgorithms APPullAPPushAPPullAPPushAPPullAPPush POpt.LimitDev.TimeDev.TimeLengthDev.TimeDev.TimeLimitDev.TimeDev.Time 46-21-12.90103.900.914.000.5027.760.117.760.1137.861.346.811.60501.101.001.100.5842.810.612.810.5755.522.164.522.7610000.143.670.143.6980.339.920.3313.74101.626.851.3311.5243-22-7.23102.230.942.270.5124.730.114.730.1132.411.902.502.21500.591.020.640.6142.090.622.090.5851.143.491.234.2710000.093.570.093.5680.456.370.458.24100.148.140.1410.6237-17-5.29101.060.870.820.4821.820.111.820.1131.762.001.762.17500.180.880.240.5240.240.560.240.5050.882.250.882.6010000.002.200.002.0780.002.760.003.06100.062.950.063.1540-54-18.22105.480.935.740.5028.330.118.330.11311.612.1711.873.26502.331.002.650.5945.060.595.060.5559.742.4310.203.7610000.393.740.413.6082.028.922.0212.54105.374.395.936.48

PAGE 87

ComparativeanalysisofheuristicSDNCCalgorithms APPullAPPushAPPullAPPushAPPullAPPush POpt.LimitDev.TimeDev.TimeLengthDev.TimeDev.TimeLimitDev.TimeDev.Time 46-21-12.90102.710.3633.100.20227.760.1117.760.11335.290.1805.290.138500.380.4440.290.23342.810.2732.810.16853.240.2713.240.20810000.000.6510.000.46380.330.5580.330.423100.710.5930.900.43343-22-7.23103.050.2912.820.18324.950.1144.950.12232.860.2022.950.152500.180.3250.320.20442.140.2462.140.17051.680.2561.680.17710000.050.4290.050.29880.450.3790.450.284100.270.3590.360.32937-17-5.29100.650.3080.710.19022.240.1112.240.11330.120.3110.120.192500.000.3250.060.19840.240.2750.240.16750.000.3770.000.21110000.000.3340.000.19380.000.3360.000.195100.000.3680.000.20640-54-18.22105.980.3655.980.21028.430.1188.430.11738.330.2088.370.144501.540.4251.440.22745.060.3005.060.18056.150.2326.090.15910000.000.6940.000.49582.020.8672.020.893102.440.4212.500.385

PAGE 88

88

PAGE 89

62 ],newspaperdistribution[ 41 ],postboxlocation[ 46 ],wastecollection[ 44 ],goodsdistribution[ 65 ]andparceldelivery[ 74 ]arejustsomeoftheexamples.Theproblemthatintegratesthefacilitylocationandvehicleroutingproblemstogetheristhecombinedlocation-routingproblem(LRP)[ 64 75 ].TheLRPconsistsofselectingasubsetoflocationsamonggivencandidatefacilitylocationsanddeterminingthevehicleroutestovisitallthecustomersfromtheselectedfacilitylocations.TheobjectiveinLRPistominimizethesumofthexedcosts,variablecostsanddeliverycostssuchthatcapacitiesofselectedfacilitylocationsandvehiclesarenotexceededwhilethedemandofeachcustomerissatised.Thexedcostsincludethecostsofestablishingdepotonacandidatelocationandthevariablecostsdependontheperunitthroughputtoacustomerfromtheselectedfacility.TheLRPreectstheinterdependencebetweenlocationcostsandroutingcosts[ 69 75 ].Hence,itisacombinationoftwoalreadydifcultproblems:FacilityLocationProblem(FLP)andVehicleRoutingProblem(VRP).BothFLPandVRPhavebeenshowntobeNP-hard[ 23 42 52 ],thustheLRPisNP-hardaswell. 89

PAGE 90

4.2 ,wegivetheproblemdescriptionandinsection 4.3 ,weintroducethemixedintegerlinearprogrammingformulationfortheLRP.Proposedcolumn-generationalgorithmispresentedindetailinsection 4.4 .Wediscussthemasterproblemandthepricingsubproblem.Wepresentthedynamicprogrammingalgorithmdevelopedforthepricingsubproblem.Attheendofthissection,wegivetheoverallalgorithmandexplainthestepsofthealgorithm.Insection 4.5 ,theimplementationdetailsandresultsaregiven. 90

PAGE 91

4-1 Xcoord. Ycoord. Capacity(T) FixedCost(FC) VariableCost(VC) 1 24 24 500 50 0.02 2 26 21 500 50 0.02 3 33 21 500 50 0.03 4 22 24 500 50 0.03 5 32 22 500 50 0.04 VehicleNumber(K) 25 VehicleCapacity(Q) 300 CostPerMile(CM) 0.75 CustNo. Xcoord. Ycoord. Demand 1 23 35 125 2 45 43 84 3 12 45 60 ... ... ... ... ... ... ... ... ... ... ... ... 34 30 16 160 35 10 40 40 36 34 60 80 ExampledatafortheLRP 91

PAGE 93

Theobjectivefunctionistominimizethetotalcostwhichiscomposedofthesumofxedcostsofestablishingdepots,thesumofvariablecoststhatdependoneachdepot'sthroughputandthesumofdeliverycosts.Constraints( 4 )statethatthereis 93

PAGE 94

4 )statethatthereisexactlyonearccomingintoeachcustomer.Constraints( 4 )ensurethatthenumberofvehiclesleavingadeparturedepotequalstothenumberofvehiclesreturningtoitscorrespondingarrivaldepot.Constraint( 4 )enforcesthatthetotalnumberofvehiclesusedshouldnotexceedtheupperboundonthenumberofvehiclesanditshouldbegreaterthanKLBwhichiscalculatedasdPi2Cdi=Qe.Inconstraintset( 4 ),thevalidinequalitythatwasproposedbyKulkarniandBhave[ 45 ]isusedinordertoeliminatecyclesthatcanoccurbetweencustomers.Constraints( 4 )givethelowerandupperboundoftheUivariables.Theloadonvehicleatdeparturefromanodeicanbeatleastequaltothedemandofthatnodeanditcanatmostbeequaltothecapacityofthevehicle.Constraints( 4 )ensurethatnodemandisassignedtoalocationifthatlocationisnotselected(thatis,yj=0)andifacandidatelocationisselected(thatis,yj=1)thenthesumofthedemandsthatareassignedtotheselectedfacilityshouldnotexceedthecapacityofthefacilitythatwillbeconstructedonthatlocation.Constraints( 4 )specifythatcustomericanbeassignedtoalocationjonlyifthatlocationisselected.Constraints( 4 )requirethateachcustomerisassignedtoexactlyonedepot.Constraints( 4 )statethatifacustomerisnotassignedtoalocationthentherecannotbeanarcthatgoesfromthatcustomertothatlocation'sarrivalnode.Similarly,constraints( 4 )statethatifacustomerisnotassignedtoalocationthentherecannotbeanarcthatgoesfromthedeparturenodeofthatlocationtothiscustomer.Constraints( 4 )and( 4 )ensurethatifcustomerjfollowscustomeriandifcustomeriisassignedtocandidatelocationkthencustomerjmustalsobeallocatedonthatsamecandidatelocation.Theremainingconstraints( 4 ),( 4 )and( 4 )specifythattheroutingvariables(xij),theallocationvariables(zik)andthelocationvariables(yj)cantakevalues0or1.ThegivenexactmodelisobservedtosolvesmallsizeLRPs(upto15customersand4depots)optimallyinthecomputationalexperimentsthatweconducted.Hence,we 94

PAGE 95

PAGE 96

4.4.2 .Themathematicalformulationofthemasterproblemisgivenbelow:Sets:R:setoffeasibleroutesJ=f1,2,..,Mg:setofcandidatelocationsC=f1,2,..,Ng:setofcustomersModelParameters:K:Numberofavailablevehiclesir:1ifthecustomeriisassignedtorouter,0otherwise;iC,rRjr:1iftherouteroriginatesfromfacilityj,0otherwise;jJ,rRPDr:Demandofrouter;rRCr:Costofrouter;rRTj:Capacityofcandidatefacilityj;jJFCj:Fixedcostofcandidatefacilityj;jJDecisionVariables:xr:1iftherouterisselected,0otherwise;rRyj:1ifthecandidatefacilityjisselected,0otherwise;jJMinimize 96

PAGE 97

TheobjectivefunctionoftheMPistominimizethecostoftheselectedroutesamongthesetoffeasibleroutesRandthexedcostofselectedfacilitylocations.ObjectivefunctionvaluesoftheMPcorrespondtotheexactLRPmodelpresentedpreviously.Notethatthevariablecostsforeachselectedlocationisincorporatedtothecostofeachroutefound.Hence,thecostofanewrouter=fi1,i2,..,itgcanbefoundbytheequation( 4 ): 4 )aredenedforeachcustomerandformthesetpartitioningconstraints.TheseconstraintsstatethateachcustomeriisservedbyexactlyonerouteamongthesetoffeasibleroutesR.Constraintset( 4 )restrictsthenumberofpathsoriginatingfromacandidatefacilitylocation.Ifacandidatelocationjisselectedthenitrequiresthatatleastonerouteoriginatesfromthatlocationandthemaximumnumberofroutesthatcanoriginatefromthelocationissettotheavailablenumberofvehicles.Ifthecandidatelocationjisnotselectedthenitrequiresthatnorouteswilloriginatefromthislocation.Constraintset( 4 )isrstgivenin[ 14 ].OneclassicalconstraintthatcanbewritteninsteadofthisconstraintcouldbexrPjJjryj,8rR.Theclassical 97

PAGE 98

4 )istighterwhencomparedtotheclassicalone.Therefore,relaxationvaluesofthemodelwithconstraintset( 4 )willbegreaterthanorequaltorelaxationvalueswiththeclassicalmodel.Anotheradvantageisthereducednumberofconstraintssince( 4 )isdenedforeachcustomerandcandidatelocation.However,theclassicalconstraintisdenedforeachfeasiblerouteR.Notethatbyusingtheirandjrparameters,( 4 )canberephrasedastheconstraint( 4 ),thatis,theyareequivalent.Constraint( 4 )setsthelowerboundonthenumberoffacilitiesthatcanbeselectedamongthecandidateset.LBwiscalculatedasdPi2Cdi=TmaxewhereTmaxisthemaximumcandidatefacilitycapacityamongall.Constraint( 4 )statesthatthetotalnumberofroutesselected(orvehiclesused)shouldnotexceedthethenumberofavailablevehiclesanditshouldbegreaterthanLBvwhichiscalculatedasdPi2Cdi=Qe.Constraintset( 4 )requiresthatsumofthedemandsofthecustomersassignedtoafacilityshouldnotexceedthefacility'scapacityifthefacilityisselected.Theconstraints( 4 )and( 4 )specifythattherouteselectionvariables(xr)andthelocationvariables(yj)cantakevalues0or1. 98

PAGE 99

4-1 showstheconstraintnumberandassociateddualvaluesthatareobtainedasthemasterproblemissolved.Thus,themodiedarccost^cijintheconstructednetwork 99

PAGE 100

ConstraintsandassociateddualvaluesfortheMasterProblem MPconstraintnumberDualvalue 4 0q[i]ifiCandjC[N+1cij 0 29 ].Anintegermodelwhosesolutionidentiesanelementaryshortestpathonthedenedsubnetworkforalocationjwouldgiveonlythecolumnwiththemostnegativereducedcost.However,thegoalistodeterminewhetheranegativereducedcostcolumnexistsornot.Soobtainingasetofgoodroutesforeachlocationjismoreimportantthanndingthebestrouteforafastevolvingalgorithm.Thatiswhy,forthepricingsubproblemweusedynamicprogrammingtondasetofroutesthathavenegativereducedcosts.ThestructureofthealgorithmweusedforsolvingtheESPPRCispresentedinFigure 4-2 .Apathiscallednon-elementaryifanodeisvisitedmorethanoncehenceifthepathcontainsanyloop;otherwiseitiscalledanelementarypath.The 100

PAGE 101

PAGE 102

DynamicprogrammingalgorithmforESPPRC 102

PAGE 103

PAGE 104

PAGE 105

4.4.1 .Columngenerationtechniqueisparticularlyappliedforthesolutionoftherelaxedmasterproblem(RMP)whichhappenstohaveanexponentialnumberoffeasiblecolumns.TheRMPistheLPrelaxationoftheMP.Asinanycolumngenerationtechnique,westartwithasubsetoffeasiblecolumnsandwetrytondnewpromisingcolumnsinthesubproblemwhichisalsocalledthepricingproblem.ForthepricingproblemwesolveelementaryshortestpathproblemwithresourceconstraintsbymodifyingthealgorithmofLarsen[ 51 ].TheobjectivefunctionvalueoftheRMPformsalowerboundtotheoriginalproblemwhennomorefeasiblerouteswithnegativereducedcostcanbefoundbythesubproblem.TheoutlineofourcolumngenerationalgorithmisgiveninFigure 4-3 .Intheinitializationstep,werstimportthedataforthegivenproblem.Afterdeterminingthenumberofcustomers(N)andthenumberofcandidatelocations(M),thesetofdeparturelocations(D)andthesetofarrivallocations(A)areestablished(Step1.2).ThearctravelcostsarecalculatedbymultiplyingtheEuclideandistancesbetweenpairsofcoordinateswiththecostpermileparameter(CM).Thedistancesarecalculatedwithvedecimalpointandtruncationwiththeformulagiveninequation( 4 ). 105

PAGE 106

Columngenerationalgorithm usedcouldnotbeincorporatedinthemasterproblemmodelinthiscase.Constraints( 4 )and( 4 )oftheMPmodelmustberemovedtohaveafeasiblesolutionandtherelaxationgapincreasesiftheyarenotutilized;therefore,theintegerMPissolvedinalongercomputingtime.Second,thenatureoftheLRPrequirestollthevehiclecapacityasmuchaspossibletoforcesmallernumberofdepotsandvehicles.Thatiswhy,multi-noderoutesthatutilizethevehiclecapacitybetterarelikelytobepresentinoptimalsolutions.Asexpected,usingsingle-noderoutesonlygivesworseinitialfeasiblesolutionsinthesensethattheyarefarfromoptimal.Hence,thenumberofpricing 106

PAGE 107

PAGE 108

108

PAGE 109

4.5.1TestProblemsWetrytosolveageneralLRPwherewehavecapacitatedmultiplefacilitiesandmultiplevehicles.TherearenocommonlyacceptedandwidelystudiedbenchmarkproblemsforthistypeofLRP.The3benchmarkproblemsofPerl[ 64 ]aretheonlyproblemsetsthatmatchwiththegeneralLRPwestudy.LRPbenchmarksofPerl[ 64 ]arestudiedin[ 65 ],[ 37 ]and[ 76 ].Theauthorsofthesepaperscomparetheircomputationalresultswitheachother.AnotherLRPdatasetwhichisclosetothegeneralLRPwestudyisprovidedbyOr[ 61 ]andisstudiedin[ 62 ].However,facilitiesareuncapacitatedintheLRPbenchmarkofOr[ 61 ].Recently,Barreto[ 11 ]triedtoformastandardizeddatabaseforthegeneralLRPsbyrespectingtheformatproposedbyPerl[ 64 ].TheinstancesthatarecompiledbyBarreto[ 11 ]aremostlyfromtheliteratureofvehicleroutingproblems.TheseedpapersorthesesfortheseinstancesareduetoPerl[ 64 ],Daskin[ 26 ],Gaskell[ 34 ],Min[ 56 ],Or[ 61 ],Srivastava[ 70 ]andChristodes[ 21 ].InordertoformastandardizeddatabaseBarretotookthedepotvariablecostsaszeroandtakethecostpermileparameteras1inalltheinstances.However,whilemakingcomputationaltestswerespectedtheoriginaldataofPerl's3benchmarks.Therefore,wecouldmakecomparisonswiththesolutionsof[ 65 ],[ 37 ]and[ 76 ].Fortheotherinstances,wecouldcompareourresultsonlywith[ 12 ]. 40 ].MixedintegerandprimalsimplexoptimizersofCPLEXareusedtosolveMIPandLPmodelsrespectively.ThealgorithmisimplementedwithJavaAPIofILOGCPLEXConcertTechnology. 109

PAGE 110

PAGE 111

ComparisononPerl'sbenchmarkproblems Perl[ 64 ]{1}2355.58{2,10,12}105795.62{2,4,5}117789.96Hansenetal.[ 37 ]{1}2355.58{2,7,12,13}105617.67{2,4,6}117551.61Wuetal.[ 76 ]{1}2355.58{5,10,12}105532.28{2,4,6}127781.21ProposedAlgorithm{1}2355.58{2,8,12}105465.51{2,4,6}117496.62 64 ],Hansenetal.[ 37 ]andWuetal.[ 76 ].Allthe3methodsareheuristicmethodsasdiscussedintheliterature.TheresultsaresummarizedinTable 4-3 .ItisobservedthatthesolutionfoundforthePerl-12c-2disthesameastheheuristicsolutionsandthesolutionsfoundforPerl-55c-15dandPerl-85c-7dislowerthanallthepreviousresults.InTable 4-4 ,wecomparetheresultsreportedinBarretoetal.[ 12 ]withourresults.InthetableprovidedbyBarretoetal.[ 12 ],the(IP)columnisthebestknownsolutionobtainedusingtheclusteringheuristic.Thelowerbound(LB)columnwasobtainedbyBarreto[ 11 ]witharelaxed2-indexintegerlinearprogrammingformulation.ItwasnotpossibletogetLBfortheinstanceMin-134c-8dusingtherelaxed2-indexintegerlinearprogrammingformulation.Usingthecolumngenerationalgorithm,avalidLBcouldnotbefoundin5oftheinstancessincethealgorithmterminatedduetotimelimit,meaningthatthealgorithmwascontinuingtogeneratepromisingcolumnswhenthetimelimithasbeenreached.Theoptimalityisreachedin2ofthe14instances(Gaskell-29c-5dandMin-27c-5d)byusingtheclusteringheuristicmethod.In8ofthe14instancestheproposedintegersolutionsislowerthantheBarreto'ssolutions.Thesolutionsobtainedinbothofthemethodsarethesamein2ofthe14instances.Intheremaining4instancesBarretoobtainedbetterresultscomparedtooursolutions.Inallthese4instancesouralgorithmterminatedduetotimelimit.FortheinstanceChristodes-100c-10d,althoughtheproposedalgorithmterminatedduetotimelimit,thesolutionfoundisbetterthantheBarreto'ssolution.Barretoetal.[ 12 ]reportsthattheintegralitygap(Gap%)fallsbetweenaminimumof0%andamaximumof19.01%with 111

PAGE 112

4-4 ,theaverageandthemedianofgapofBarretoetal.[ 12 ]is3.90%and2.24%respectively.Fortheproposedalgorithm,theintegralitygap(Gap%)fallsbetweenaminimumof0%andamaximumof4.31%withanaverageof1.24%andamedianof0.59%. Table4-4. ComparisononbenchmarkproblemsthatarecompiledbyBarreto[ 11 ] 12 ]ProposedAlgorithm ProblemNameVcapIPLBGap(%)IPLBGap(%) 112

PAGE 113

113

PAGE 114

PAGE 115

PAGE 116

PAGE 117

5.2 117

PAGE 118

5.3 .Inthispaperweconsidertheglobalproblem,includingallreal-lifeattributes,anddiscusssomeholisticapproaches.Theseapproachesaretherstattemptsatcapturingandsolvingallissuesinonemodel,sothattheycanbeimplementedinreal-lifescenarios.Weproposethreeexactalgorithmswhichvaryintheircapabilitiesandeffectivenessforhandlingdifferentaspectsoftheproblem.Wealsosuggestadecomposition-basedheuristicwhichgeneratespracticalresultswithinareasonabletime-frame.Ourmodelsconsiderallmajorfactorsaffectingtheefciencyofthecurfewplanandareexibleenoughtobeextendedforcapturingnewrequirements.InSection 5.4 ,weformulatetheCPPasaMixedIntegerProgramming(MIP)Problemonanappropriatelydenedtime-spacenetwork.Thismodelisusefulforunderstandingtheintricaciesandcomplexitiesoftheproblem.Thecontributionsofthispaperinclude: 1. Wepresentaunique,duty-generationmodelfortheCPP.Itsframeworkcapturesallissuesdiscussedearlier,anditisveryexibleandcompact(Section 5.5 ). 2. Weproposeacolumn-generationapproachtoimprovethesolutionsobtainedfromtheduty-generationmodel.Althoughitispartiallysuccessfulinthecurrentscopeoftheproblem,itgeneratesgoodsolutionsinthescenariowhereprojectcrashingisnotallowed(Section 5.6 ). 3. Wesuggestadecomposition-basedheuristicthatgeneratesverypracticalandexcellentqualitysolutionswhichcanbeimplementeddirectlyintoreal-lifeapplications.Thecomputationalresultsonreal-worldinstancestakenfromalargerailroadshowasignicantimprovementinthenumberofdisruptionsorviolations(Section 5.7 ). 118

PAGE 119

1. 2. 3. 4. 5. 6. 119

PAGE 120

39 ].Itproposesamodelaimedatdeterminingthebestallocationofmaintenanceactivitiesandcrewsforminimizingthedisruptioninthescheduledtrainsandreducingthecompletiontime.Heconsidersasubsetofthepreviouslydescribedconstraintsandgivesanon-linearformulationoftheproblem.Thenon-linearityandlargesizeoftheproblemforcetheuseofheuristicsandhenceatabu-searchheuristichasbeenproposedthatgeneratestheneighborsbyswappingtheorderingofjobs,crews,orboth.Higginsetal.[ 39 ]usethismodelforthecomputationaltestingonan89-km.trackcorridorandimprovethemanuallycreatedschedulesbyalmost7%.AdynamicschedulegenerationtechniquefortherollinghorizonisdiscussedbyBruceetal.[ 18 ].Theirstudyisbasedonreal-timedatafromtheHongKongsubwayrailwaysystem.Lakeetal.[ 47 ]havedonestudiesonAustralianfreightoperations.Theysaythatrailwaysoperateundertheconictingobjectiveofminimizingtheinfrastructurecostswhilemaintainingorimprovingtheservicequalities.Theseconictshaveintensiedduetoindustrialgrowth,hightrafcvolume,andotherrelatedfactors.Theydevelopthemodelfortheshort-termmaintenanceschedulingoftrack-maintenanceactivitiesafterthetrainschedulesandmaintenanceactivitieshavebeenplanned.Lakeetal.[ 47 ]solveitusingatwo-phaseheuristicstechnique,withtherststagegeneratingafeasiblesolution,andthesecondstageimplementingsimulatedannealing.Theyapplythesealgorithmsontwenty-tracksegmentsandonaseven-dayschedulingperiod.Foroccasionallyusedtracks,e.g.inAustraliaandsomeEuropeancountries,Budaietal.[ 19 ]showthatthetrack-possessionismodeledin-betweenoperations.Theyintroduceaslightlydifferentversionoftheproblemwherethemainpurposeiscompletingtheprojectwithinthetrack'sfreetime.Itcreatesadynamicschedulefor 120

PAGE 121

PAGE 122

PAGE 123

Space-Timenetworkforthecurfewplanningproblem. theregetrelocatedtoeitherthestartingnodeofaprojectordirectlytothesinknode.Wecreateseveralrelocationpoolsineachweekbasedonthegeographicalstructureofthenetworkandthejob'svolume.Allstartingnodesandnishingnodesareconnectedthrougharcs,calledrelocationarcs,totherelocationpoolsofthesameweekwhicharewithinapredenedfactorofmaximumdistancelimitbasedonpreviousexperiments.Althoughtherelocationpoolsreducetheexactnessoftheproblem,theymaketheproblemtractable.Inourformulation,acrewmayrelocatefromprojectptoproject 123

PAGE 124

5.2 canberelaxed,theproblemcanbeexpressedasalinearprogram.Thesolutionofthislinearformulationwouldalwaysresultinanintegeroptimalsolutionduetothetotalunimodularityoftheconstraintmatrix.Unfortunately,thesideconstraintsmakethisproblemverydifcult.However,wecanreducethenumberofintegervariablesbydeningeverythinglinearexceptforthevariablescorrespondingtotheowontheprojectarcs.Theowconservationconstraintsmixedwiththeobjectivefunctionforceallothervariablestobeintegers.Toincreasethefeasibilityoftheproblemandndtheinitialsolutioneasily,werelaxthecontinuityrestrictionsandintroducesojournarcsbetweencontiguousrelocation-poolnodes(seeFigure 5-1 ).Thisallowsthecrewstoremaininthepoolandbeassignedlaterintheprojectspan.Butwediscouragetheseowsbyassigningveryhighcostsonthesearcs.Tofurtherimprovethefeasibility,wecanallowcrewstotravelfromarelocationpooltoanotheratveryhighcosts,sinceitwillintroducemanyinfeasibletravelsduringweekends.Pleasenotethatrelocation-poolnodesforeachprojectineachweekappeartwiceinthenetwork(Figure 5-1 )tomakethediagramcleaner,thoughtheyarethesameintheoriginalnetwork.Nextweshowthemathematicalmodelforthetime-spacenetwork. 124

PAGE 125

125

PAGE 126

126

PAGE 127

5 maintainthebalancedowofeachcrew-typethroughoutthenetwork.Constraints 5 keepacheckonthenumberofcrewsemployedwhileconstraints 5 statethateachprojectmustbecompletedbyexactlyonecrew.Constraints 5 controlthevalueofsubdivisionvariables.Constraints 5 enforcethemaximumabsolutecurfewsallowedperweek,andConstraints 5 tracktheprecedencerelations.Constraints 5 maintaintheservice-corridorabsolutecurfewrestrictions.Constraints 5 donotallowtwomutuallyexclusivesubdivisionstobecomeactivesimultaneously.Constraints 5 aretheintegralityconstraints,althoughmanyofthesecanberelaxedtolinearvariablesasdiscussedearlierinthesection. 127

PAGE 128

5.8 ).Weusethesamedatasetsfortestingeachformulation.TheorderofnumberofvariablesareO(|P|*W*|k|/2*r),whererisaveragenumberofprojectswhicharewithinthemaximumdistancelimitofeachother.TheorderofnumberofconstraintsareO(W*(1+|P|+k+2|S|+|U|)+|P|).Forthereal-lifeinstances,theaveragenumberofvariableswasaround600,000,andtheaveragenumberofconstraintswasaround65,000.WesolvethisformulationusingCPLEX11.2.Thebasicmodelwithoutusingtherelocationpooldidn'tgiveanysolutionwithin24hours.Wedidn'tallowtheprocesstorunformorethan24hours,becauseitlosesthepracticalityoftheapproach.Atimelimitofeighthoursisgoodfromthepracticalpointofview.Theintroductionofrelocationpoolssignicantlyimprovesthesolutiontimetotwohours,butthenumberoftravelsviolatingthemaximumdistancelimitincreasedtoanunacceptablelimit.Therailprojectsandtieprojectscanbedividedandsolvedindependentlyandnallycombinedtocreateaglobalsolution.Theprocessofcombiningneedssomere-optimization,assomeconstraintsgetviolatedwhentheindependentsolutionsaremerged.Wesolvethesedecomposedproblemsandareabletondthefeasiblesolutionsforthesameprobleminstancewithinveminutesforeachpart.Theprocessofcombiningtakesanother15minutes,butitresultsinseveralotherviolationswithpenalty.Thoughthismodelisnotverysuccessful,itisgoodforbetterunderstandingtheproblemandidentifyingthecriticalissues.Themaindrawbackofthemodelisitsincapabilityinhandlingprojectcrashingwithoutwhichbothreal-lifeinstancescauselargenumbersofviolations.Projectcrashinghelpsreducethenumberofabsolutecurfewsinaweek,andthusimprovesthefeasibility.Toincludeit,wewouldhavetointroduceanothersetofarcsforeachproject,andthatwouldmakethemodelevenmoreintractable.Theseresultsguideustowarddevelopingamorecompactformulationthatcanhandletheprojectcrashingandotherconstraintsmoreefciently.Wediscussthismodel,theduty-generationmodel,inthefollowingsection. 128

PAGE 129

CrewSet1CrewSet2CrewSet3CrewSet4 1 0 0 1 0 0 1 0 0 1 0 0 1 1 0 1 1 0 1 1 0 0 1 0 1 1 0 1 1 0 0 1 0 0 0 0 1 1 0 0 1 0 0 0 0 0 0 0 . . . . . 0 0 1 0 0 1 0 0 1 0 0 0 0 0 1 0 0 1 0 0 1 0 0 1 Figure5-2. VariablematrixD[p]forprojectp. WecreateamatrixD[p]foreachprojectpwhichisacollectionofallcandidatedutiesoftheprojectp.Thedutiesareidentiedbytheiruniqueid.Inthisformulation,wehavetheexibilityofcrashingtheprojectsbycombiningtwosmallcrewsortwolargecrews.Thisfeaturecannotbeimplementedintherstmodel,aseachnewcrew-setdoublesthesizeoftheproblem.Acrew-setisthecombinationofcrewsemployedtocompleteaproject.Wecreatefourcrew-setsforeachcrew-type(railandtie):(i)1-smallcrew,(ii)1-largecrew,(iii)2-smallcrew,and(iv)2-largecrew;wherethelasttwoare 129

PAGE 130

5-2 ,wheretheprojecttakesfour,three,two,oroneweektocompleteifdonebythefourcrew-setsdiscussedabove,respectively.Nextwepresentthemathematicalformulation:Index:d:ColumnofmatrixD.awd:ValueatwthrowanddthcolumnofmatrixD.c:Elementofacrew-setCS.Parameters:D:Setofdutiesforallprojects.D[p]:Setofdutiesfoeprojectp.D[p][c]:Setofdutiesfortheprojectpdonebycrew-setc.kd:Numberofcrew-typekindutyd.d:1ifdutydrequiresanabsolutecurfew,otherwise0.Ids:1ifdutydisinsubdivisions,otherwise0.du=1ifdutydisinservicecorridoru,otherwise0.fd:Startingweekofdutyd,i.e.theweekinwhichrst1comesincolumnd.ld:Finishingweekofdutyd,i.e.theweekinwhichlast1comesincolumnd.ct[d]:crewtypeinvolvedindutyd.Variables:xd=1ifdthcolumnofmatrixDistheduty-scheduleforitsproject;0otherwise.ysw:1ifsubdivisionsisunderabsolutecurfewinweekw;otherwise0.vcurfw:Violationinat-most-curfewconstraintsinweekw.vsubms1s2w:Violationinmutuallyexclusiveconstraintsinweekw.vservuw:Violationinservicecorridorconstraintsinweekwforcorridoru.vprecpq:Violationinprecedenceconstraintsforpairpq. 130

PAGE 132

5 ensurethecompletionofeachprojectbyexactlyonecrewset.Constraints 5 and 5 maintainthecontinuityrestrictioninallweeks,exceptforthelastoneaswerelaxtheconstraintinthelastweek.Aswediscussedintime-spacenetworkmodel, 132

PAGE 133

5 controlthesubdivisionvariablesassignment.Constraints 5 keepcheckonthenumberoftheabsolutecurfewsineachweek.Constraints 5 and 5 ,repsectively,donotallowmutuallyexclusiveprojectsandsubdivisionstobeactivesimultaneously.Constraints 5 areforprecedencerelationsamongtheprojects,andConstraints 5 maintaintheservicecorridorrestrictions.Wetrytopartiallyenforcethedistanceconstraintsthroughset 5 and 5 .TheDGMhasbeendevelopedatthecrew-typelevel,sothereisnodirectwaytoforcethemaximumdistancelimit.Thebasicconceptisthatforeachdutyendinginweekw,thereshouldbeatleastonedutycontainingatleastoneofthesamecrew-typeswithinthealloweddistancelimitstartinginweekw+1;andforeachpairofdutiesendinginweekw,thereshouldbeatleasttwodutiesstartinginweekw+1,andsoon.Settingalloftheseasconstraintswouldhavemadetheproblemintractable.Sowetryadifferentmethodtopartiallyforcetheseconstraints.InConstraints 5 ,westatethatforeachdutyendinginweekw,thereshouldbeatleastonedutycontainingatleastoneofthesamecrew-typeswithinthedistancelimitstartinginweekw+1.InConstraints 5 ,wesaythatforeachdutyendinginweekw,thereshouldbeatleastonedutycontainingatleastoneofthesamecrew-typesstartinginweekw1.Thesesetsofconstraintstogetherhandlethedistanceconstraintsveryeffectively,asshowninthecomputationalsection(Section 5.8 ).Thesetwosetsofconstraintsarereferredtoaspartialdistanceconstraintsthroughoutthepaper.Werelaxtheserestrictionsforlastveweeksfortworeasons:(i)toimprovethefeasibility,and(ii)becauseitmightbethelastprojectdonebythecrew.Set 5 aretheintegralityconstraints.Theoutputofthisphaseisthestartingweekandcrew-typeemployedforeachproject,andtheyareusedinthenextphasetogeneratethecrewschedules. 133

PAGE 134

1. 2. 3. 5 ].Asallsideconstraintsaresatisedintheprojectschedulingphase,optimizingtheowoveranetworkwillgeneratethebestcrew-schedules. 134

PAGE 135

PAGE 136

136

PAGE 137

137

PAGE 138

5 ensurethateachprojectiscoveredbyexactlyonerouteandConstraints 5 keeptrackofthenumberofcrewsemployed.Constraints 5 and 5 forcethenumberofabsolutecurfewswithinthegivenlimit.Constraints 5 and 5 ensurethemutuallyexclusivityconstraintsatthesubdivisionandprojectlevels,respectively.Constraints 5 keepacheckonthenumberofservicecorridorsunderabsolutecurfew,andConstraints 5 maintaintheprecedencerelations.Constraints 5 enforcethevariables'integrality.Thevaluesinparenthesesarethedualvaluesusedinthepricingproblems.WenowdiscussthePricingProblem,whichissolvedrepetitivelytoinsertthepotentialcandidateroutesintotheRMP. 5.6.1 ).WeneedtorevisethearccostsofthepricingproblemseachtimewesolvetheRMPandseeknewcandidatestoenter.Wealsoneedtodynamicallycalculatethearccosts,becausethestart-timesofunscheduledprojectsareunknown. 138

PAGE 139

PAGE 140

:Updatedarccosts,timewindowofeachnode,andprecedencematrix. Output :Asetofelementarypathsfromnode0tonodeN+1thathavenegativereducedcostsanddonotviolateanycapacityconstraints. Step1 :InitializationGettheupdatedarccostsassociatedwiththelastRMPsolved.CreateLabel({0,0}).Initializeasetofsetsthatstorestheelementsofeachdistinctpathfound.Initializethepredecessorset. Step2 :WHILEu6=N+1DO{u,tu}=BestLabel(unprocessedlabels)IFu
PAGE 141

:Distancebetweenprojects,numberofcrewsofeachtype,precedencematrix,andtimewindowsfortheprojects. Output :Scheduleforeachcrew. Step1 :Initialization.Gettheinputs.Createdeparturedepot(D)andarrivaldepot(A)sets.Initializeroutes:Constructmulti-noderoutesusingheuristics.PrepareinitialsetofroutesfortheMP. Step2 :ConstructtheMP. Step3 :SolvetheRMPwithcurrentsetoffeasibleroutes.Updatethelowerbound.GetthedualvaluesforeachconstraintoftheRMP.Modifythearccostsusingthedualvalues.FOReachcrewtypeDOSolvesubproblemswithmodiedcosts.IFReducedCostr<0foranynewpathr,gotoStep4.ELSEgotoStep5. Step4 :AddthenewpromisingcolumnstotheRMPandgotoStep3. Step5 :Optimalsolutionfortheoriginalrelaxedmasterproblemhasbeenfound.Figure5-4. Algorithmframeworkofthecolumn-generationprocedure. 141

PAGE 142

142

PAGE 143

PAGE 144

PAGE 145

Table5-2. Improvementinnumberofviolationsfor2007 15 ]TSNFDGMCGMDecomposedDGM Improvementinnumberofviolationsfor2007 15 ]TSNFDGMCGMDecomposedDGM 145

PAGE 146

15 ].Theirformulationcapturesallissuesoftheproblem,butitisintractableevenformodest-sizedinstances.WesolvedtheexactmodelusingCPLEX11.2ona2.5GHzPCwith2GBofmemory.Wesettheterminationcriteriabasedonamaximumavailabletimeofeighthours,evenforsmallerinstances.Itgeneratedoptimalsolutionsforsixoutoftenve-weekinstances,whiletherewereoptimalitygapsof3%to8%fortherest[ 15 ].TheresultsobtainedbyBogetal.[ 15 ],TSNF,DGM,CGM,andDGM*(explainednext)forve-weekinstancesareshowninTable 5-4 .TheTSNFgeneratesoptimalsolutionsforve-weekinstancesinaroundsixminutes,butprojectcrashingisnotincluded.ThequalityoftheTSNFisnotacceptableevenforthesesmallinstancesasthegapvariesfrom2%to15%,whichclearlydemonstratestheimportanceofprojectcrashing.WetestedtheDGMbyputtingallpartialdistanceconstraints(equivalenttothedecomposedmodelforsmallerinstances)intheseinstances.Itgeneratedoptimalsolutionsforeightoutofteninstances.Intheremainingtwoinstances,thenumberofviolationswasonlyonemorethantheoptimalsolutions.WeusedtheseschedulesastheinitialsolutionsoftheCGM,whichimprovedthesolutionofoneofthesetwo.Wealsotestedtheeffectofusingthepair-wiseduties'distance 146

PAGE 147

5.5 .TheresultsareshowninTable 5-4 incolumnDGM*.Itconvergedtotheoptimalsolutionfornineoutofteninstances.Theseresultsshowthatthepartial-distanceconstraintsareveryeffectiveinthesereal-lifeinstancesthathaveprojectsscatteredoveralargegeographicalregion.Wealsotestedouralgorithmsonten-weekinstances.Theexactalgorithmscouldnotgenerateevenonefeasiblesolutionforanyoftheseinstances,againemphasizingtheveryhighcomplexityoftheproblem.Theresultsofothermodelsshowasimilarbehavior,withtheDGMgeneratingthebestresultsconsistently. Table5-4. Comparisonofalgorithms:optimalitygap(smallinstances) 15 ]TSNFDGMCGMDGM* 147

PAGE 148

148

PAGE 149

149

PAGE 150

150

PAGE 151

151

PAGE 152

[1] RailroadFacts,2004.AssociationofAmericanRailroads. [2] TheimportanceofAdequateRailInvestment,2005.AssociationofAmericanRailroads. [3] Ahuja,R.K.,Ergun,O.,Orlin,J.B.,Punnen,A.P.,2002.Asurveyofverylargescaleneighborhoodsearchtechniques.DiscreteAppliedMathematics123,75. [4] Ahuja,R.K.,Kumar,A.,Jha,K.C.,2007.Exactandheuristicalgorithmsfortheweapon-targetassignmentproblem.OperationsResearch55,1136. [5] Ahuja,R.K.,Magnanti,T.,Orlin,J.B.,1993.NetworkFlows:Theory,AlgorithmsandApplications,Chapter15.PrenticeHall,EnglewoodCliffs,NJ,566. [6] Ahuja,R.K.,Orlin,J.B.,Sharma,D.,2001.Multi-exchangeneighborhoodsearchalgorithmsforthecapacitatedminimumspanningtreeproblem.MathematicalProgramming91,71. [7] Ahuja,R.K.,Orlin,J.B.,Sharma,D.,2003.Acompositeverylarge-scaleneighborhoodstructureforthecapacitatedminimumspanningtreeproblem.OperationsResearchLetters31,185. [8] Balas,E.,2006.Theprizecollectingtravellingsalesmanproblem.Networks19(6),621. [9] Bard,J.F.,Kontoravdis,J.F.,Yu.G.,2002.Abranch-and-cutprocedureforthevehicleroutingproblemwithtimewindows.TransportationScience36(2),250. [10] Barnhart,C.,Johnson,E.L.,Nemhauser,G.L.,Savelsbergh,M.W.P.,Vance,P.H.,1998.Branch-and-Price:Columngenerationforsolvinghugeintegerprograms.OperationsResearch46,315. [11] Barreto,S.S.,2004.Analysisandmodellingoflocation-routingproblems.PhDthesis,AveiroUniversity. [12] Barreto,S.S.,Ferreira,C.,Paixao,J.,Santos,B.S.,2007.Usingclusteringanalysisinacapacitatedlocation-routingproblem.EuropeanJournalofOperationsResearch179,968. [13] Beasley,J.E.http://www.ms.ic.ac.uk/info.html [14] Berger,R.T.,Coullard,C.R.,Daskin,M.S.,2007.Location-routingproblemswithdistanceconstraints.TransportationScience41(1),29. [15] Bog,S.,Nemani,A.K.,Ahuja,R.K.,2009.Iterativealgorithmsforthecurfewplanningproblem.SubmittedtotheJournaloftheOperationalResearchSociety. 152

PAGE 153

Bontekoning,Y.M.,Kreutzberger,E.,2001.New-generationterminals:Aperformanceevaluationstudy.DelftUniversityPress,Delft. [17] Bostel,N.,Dejax,P.,1998.Modelsandalgorithmsforcontainerallocationproblemsontrainsinarapidtransshipmentyard.TransportationScience32,370. [18] Bruce,S.N.,Chow,C.K.P.,Hui,L.C.K.,Yong,A.M.K.,1999.Railwaytrackpossessionassignmentusingconstraintsatisfaction.ArticialIntelligence12,599. [19] Budai,G.,Huisman,D.,Dekker,R.,2004.Schedulingpreventiverailwaymaintenanceactivities.IEEEInternationalConferenceonSystems,ManandCybernetics,4171. [20] Chabrier,A.,2005.Vehicleroutingproblemwithelementaryshortestpathbasedcolumngeneration.Computers&OperationsResearch33,2972. [21] Christodes,N.,Eilon,S.,1969.Analgorithmforthevehicle-dispatchingproblem.OperationalResearchQuarterly20(3),309. [22] Clarke,G.,Wright,J.W.,1964.Schedulingofvehiclesfromacentraldepottoanumberofdeliverypoints.OperationsResearch12,568. [23] Cornuejols,G.,Fisher,M.L.,Nemhauser,G.L.,1977.Locationofbankaccountstooptimizeoat:Ananalyticstudyofexactandapproximatealgorithms.ManagementScience23(8),789. [24] Corry,P.,Kozan,E.,2006.Anassignmentmodelfordynamicloadplanningofintermodaltrains.ComputersandOperationsResearch33,1. [25] Corry,P.,Kozan,E.,1987.Full-scalerailcartestingtodeterminetheeffectofposition-in-trainonaerodynamicresistance.ReportofAssociationofAmericanRailroads,Publication,Sd-705. [26] Daskin,M.S.,1995.Networkanddiscretelocation:models,algorithmsandapplications.JohnWiley&Sons,Inc.,NewYork. [27] Desrochers,M.,Soumis,F.,1988.Ageneralizedpermanentlabellingalgorithmfortheshortestpathproblemwithtimewindows.INFOR26(3),191. [28] DeWitt,W.,Clinger,J.,2000.Intermodalfreighttransportation:Transportationinthenewmillennium:stateoftheartandfuturedirectionsUperspectivefromtransportationresearchboardstandingcommittees.NationalResearchCouncil,WashingtonDC33. [29] Dror,M.,1994.Noteonthecomplexityoftheshortestpathmodelsforcolumngenerationinvrptw.OperationsResearch42,977. 153

PAGE 154

PAGE 155

Labbe,M.,Laporte,G.,1986.Maximizinguserconvenienceandportalserviceefciencyinpostboxlocation.BelgianJournalofOperationsResearch,StatisticsandComputerScience26,21. [47] Lake,M.,Ferreira,L.,Murray,M.,2000.Minimisingcostsinschedulingrailwaytrackmaintenance.InternationalConferenceonComputersinRailwaysVII,Bologne,Italy,895. [48] Lai,Y.C.,Barkan,C.P.L.,2005.Optionsforimprovingtheenergyefciencyofintermodalfreighttrains.TransportationResearchRecord1916:JournaloftheTransportationResearchBoard,47. [49] Lai,Y.C.,Barkan,C.P.L.,2006.Improvingenergyefciencyofintermodaltrainsusingmachinevisionandoperationsresearchanalysis.Proceedingsofthe7thWorldCongressonRailwayResearch,Montreal. [50] Lai,Y.C.,Barkan,C.P.L.,Onal,H.,2007.Optimizingtheaerodynamicefciencyofintermodalfreighttrains.TransportationResearchPartE:LogisticsandTransportationReview44,820. [51] Larsen,J.,1999.ParallelizationoftheVehicleRoutingProblemwithTimeWindows.PhDthesis,InstituteofMathematicalModeling,TechnicalUniversityofDenmark,Lyngby. [52] Lenstra,J.K.,RinnooyKan,A.H.G.,1981.Complexityofvehicleroutingandschedulingproblems.Networks11,221. [53] Lbbecke,M.E.,Desrosiers,J.,2005.Selectedtopicsincolumngeneration.OperationsResearch53,1007. [54] Macharis,C.,Bontekoning,Y.M.,2003.OpportunitiesforORinintermodalfreighttransportresearch:Areview.EuropeanJournalofOperationalResearch154,400. [55] McKenzie,D.R.,North,M.C.,Smith,D.S.,1989.IntermodalTransportationUThewholestory.Simmons-BoardmanBooks. [56] Min,H.,Current,J.,Schilling,D.,1992.Themultipledepotvehicleroutingproblemwithbackhauling.JournalofBusinessLogistics13(1),259. [57] Min,H.,Jayaraman,V.,Srivastava,R.,1998.Combinedlocation-routingproblems:asynthesisandfutureresearchdirections.EuropeanJournalofOperationalResearch108,1. [58] Nemhauser,G.L.,Wolsey,L.A.,1999.IntegerandCombinatorialOptimization.JohnWileyandSons,Inc. [59] Newman,A.L.,Yano,C.A.,2000.Schedulingdirectandindirecttrainsandcontainersinanintermodalsetting.TransportationScience34,256. 155

PAGE 156

Oke,S.,2004.Maintenancescheduling:description,status,andfuturedirections.SouthAfricanJournalofIndustrialEngineering15(1),101. [61] Or,I.,1976.Travelingsalesman-typecombinatorialproblemsandtheirrelationtothelogisticsofregionalbloodbanking.PhDthesis,NorthwesternUniversity. [62] Or,I.,Pierskalla,W.P.,1979.Atransportationlocation-allocationmodelforregionalbloodbanking.AIIETransactions11,86. [63] Patterson,Z.,Ewing,G.,Haider,M.,2006.TherstcarrierchoicesurveyfortheQuebeccity-Windsorcorridor:shipperpreferencessuggestuphillbattleforincreasingrailfreightmarketshare.ConferencepaperSTRC. [64] Perl,J.,1983.AUniedWarehouseLocation-RoutingAnalysis.PhDthesis,NorthwesternUniversity. [65] Perl,J.,Daskin,M.S.,1985.Awarehouselocation-routingproblem.TransportationResearchB19B(5),381. [66] Powell,W.B.,Carvalho,T.A.,1998.Real-timeoptimizationofcontainersandatcarsforintermodaloperations.TransportationScience32,110. [67] Powell,W.B.,Carvalho,T.A.,1998.Dynamiccontroloflogisticsqueueingnetworkforlarge-scaleeetmanagement.TransportationScience32,90. [68] Robl,E.H.,2002.UnderstandingIntermodal:Aportableprimerontodaysmultimodaltransportationequipmentandsystems.Ebook. [69] Salhi,S.,Rand,G.K.,1989.Theeffectofignoringrouteswhenlocatingdepots.EuropeanJournalofOperationalResearch38,150. [70] Srivastava,R.,1986.Algorithmsforsolvingthelocation-routingproblem.PhDthesis,TheOhioStateUniversity. [71] Thompson,P.M.,Orlin,J.B.,1989.Thetheoryofcyclictransfers.MITOperationsResearchCenterReport,200-289. [72] Thompson,P.M.,Psaraftis,H.N.,1993.Cyclictransferalgorithmsformulti-vehicleroutingandschedulingproblems.OperationsResearch41,935-946. [73] Thompson,P.M.,Psaraftis,H.N.,2005.FreightRailTransportation:Areviewofthe2004experience.CongressoftheUnitedStatesCongressionalBudgetOfce. [74] Wasner,M.,Zapfel,G.,2004.Anintegratedmulti-depothub-locationvehicleroutingmodelfornetworkplanningofparcelservice.InternationalJournalofProductionEconomics90,403. [75] Webb,M.H.J.,1968.Costfunctionsinthelocationofdepotformultipledeliveryjourneys.OperationalResearchQuarterly19,311. 156

PAGE 157

Wu,T-H.,Low,C.,Bai,J-W.,2002.Heuristicsolutionstomulti-depotlocation-routingproblems.Computers&OperationsResearch29,1393. 157

PAGE 158