交通灯控制系统外文翻译.docx

交通灯控制系统外文翻译.docx

《交通灯控制系统外文翻译.docx》由会员分享，可在线阅读，更多相关《交通灯控制系统外文翻译.docx（10页珍藏版）》请在冰豆网上搜索。

交通灯控制系统外文翻译

本科生毕业设计（论文）

外文文献翻译

毕业设计题目：

交通灯智能控制系统

学院：

信息科学与工程学院

专业班级：

测控技术与仪器0703班

学生姓名：

王欣

指导教师：

桑海峰

2011年3月19日

外文原文

IntelligentTrafficLightControl

MarcoWiering,JellevanVeenen,JillesVreeken,andArneKoopmanIntelligentSystemsGroup

InstituteofInformationandComputingSciencesUtrechtUniversity

Padualaan14,3508TBUtrecht,TheNetherlands

email:

marco@cs.uu.nl

July9,2004

Abstract

Vehiculartravelisincreasingthroughouttheworld,particularlyinlargeurbanareas.Thereforetheneedarisesforsimulatingandoptimizingtrafficcontrolalgorithmstobetteraccommodatethisincreasingdemand.Inthispaperwestudythesimulationandoptimizationoftrafficlightcontrollersinacityandpresentanadaptiveoptimizationalgorithmbasedonreinforcementlearning.Wehaveimplementedatrafficlightsimulator,GreenLightDistrict,thatallowsustoexperimentwithdifferentinfrastructuresandtocomparedifferenttrafficlightcontrollers.Experimentalresultsindicatethatouradaptivetrafficlightcontrollersoutperformotherfixedcontrollersonallstudiedinfrastructures.

Keywords:

IntelligentTrafficLightControl,ReinforcementLearning,Multi-AgentSystems（MAS）,SmartInfrastructures,TransportationResearch

1Introduction

Transportationresearchhasthegoaltooptimizetransportationflowofpeopleandgoods.Asthenumberofroadusersconstantlyincreases,andresourcesprovidedbycurrentinfrastructuresarelimited,intelligentcontroloftrafficwillbecomeaveryimportantissueinthefuture.However,somelimitationstotheusageofintelligenttrafficcontrolexist.Avoidingtrafficjamsforexampleisthoughttobebeneficialtobothenvironmentandeconomy,butimprovedtraffic-flowmayalsoleadtoanincreaseindemand[Levinson,2003].

Thereareseveralmodelsfortrafficsimulation.Inourresearchwefocusonmicroscopicmodelsthatmodelthebehaviorofindividualvehicles,andtherebycansimulatedynamicsofgroupsofvehicles.Researchhasshownthatsuchmodelsyieldrealisticbehavior[NagelandSchreckenberg,1992,WahleandSchreckenberg,2001].

Carsinurbantrafficcanexperiencelongtraveltimesduetoinefficienttrafficlightcontrol.Optimalcontroloftrafficlightsusingsophisticatedsensorsandintelligentoptimizationalgorithmsmightthereforebeverybeneficial.Optimizationoftrafficlightswitchingincreasesroadcapacityandtrafficflow,andcanpreventtrafficcongestions.Trafficlightcontrolisacomplexoptimizationproblemandseveralintelligentalgorithms,suchasfuzzylogic,evolutionaryalgorithms,andreinforcementlearning（RL）havealreadybeenusedinattemptstosolveit.Inthispaperwedescribeamodel-based,multi-agentreinforcementlearningalgorithmforcontrollingtrafficlights.

Inourapproach,reinforcementlearning[SuttonandBarto,1998,Kaelblingetal.,1996]withroad-user-basedvaluefunctions[Wiering,2000]isusedtodetermineoptimaldecisionsforeachtrafficlight.Thedecisionisbasedonacumulativevoteofallroadusersstandingforatrafficjunction,whereeachcarvotesusingitsestimatedadvantage（orgain）ofsettingitslighttogreen.Thegain-valueisthedifferencebetweenthetotaltimeitexpectstowaitduringtherestofitstripifthelightforwhichitiscurrentlystandingisred,andifitisgreen.Thewaitingtimeuntilcarsarriveattheirdestinationisestimatedbymonitoringcarsflowingthroughtheinfrastructureandusingreinforcementlearning（RL）algorithms.

Wecomparetheperformanceofourmodel-basedRLmethodtothatofothercontrollersusingtheGreenLightDistrictsimulator（GLD）.GLDisatrafficsimulatorthatallowsustodesignarbitraryinfrastructuresandtrafficpatterns,monitortrafficflowstatisticssuchasaveragewaitingtimes,andtestdifferenttrafficlightcontrollers.Theexperimentalresultsshowthatincrowdedtraffic,theRLcontrollersoutperformallothertestednon-adaptivecontrollers.Wealsotesttheuseofthelearnedaveragewaitingtimesforchoosingroutesofcarsthroughthecity（co-learning）,andshowthatbyusingco-learningroaduserscanavoidbottlenecks.

Thispaperisorganizedasfollows.Section2describeshowtrafficcanbemodelled,predicted,andcontrolled.Insection3reinforcementlearningisexplainedandsomeofitsapplicationsareshown.Section4surveysseveralpreviousapproachestotrafficlightcontrol,andintroducesournewalgorithm.Section5describesthesimulatorweusedforourexperiments,andinsection6ourexperimentsandtheirresultsaregiven.Weconcludeinsection7.

2ModellingandControllingTraffic

Inthissection,wefocusontheuseofinformationtechnologyintransportation.Alotofgroundcanbegainedinthisarea,andIntelligentTransportationSystems（ITS）gainedinterestofseveralgovernmentsandcommercialcompanies[Ten-TexpertgrouponITS,2002,WhitePaper,2001,EPA98,1998].

ITSresearchincludesin-carsafetysystems,simulatingeffectsofinfrastructuralchanges,routeplanning,optimizationoftransport,andsmartinfrastructures.Itsmaingoalsare:

improvingsafety,minimizingtraveltime,andincreasingthecapacityofinfrastructures.Suchimprovementsarebeneficialtohealth,economy,andtheenvironment,andthisshowsintheallocatedbudgetforITS.

Inthispaperwearemainlyinterestedintheoptimizationoftrafficflow,thuseffectivelyminimizingaveragetraveling（orwaiting）timesforcars.Acommontoolforanalyzingtrafficisthetrafficsimulator.Inthissectionwewillfirstdescribetwotechniquescommonlyusedtomodeltraffic.Wewillthendescribehowmodelscanbeusedtoobtainreal-timetrafficinformationorpredicttrafficconditions.Afterwardswedescribehowinformationcanbecommunicatedasameansofcontrollingtraffic,andwhattheeffectofthiscommunicationontrafficconditionswillbe.Finally,wedescriberesearchinwhichallcarsarecontrolledusingcomputers.

2.1ModellingTraffic.

Trafficdynamicsbareresemblancewith,forexample,thedynamicsoffluidsandthoseofsandinapipe.Differentapproachestomodellingtrafficflowcanbeusedtoexplainphenomenaspecifictotraffic,likethespontaneousformationoftrafficjams.Therearetwocommonapproachesformodellingtraffic;macroscopicandmicroscopicmodels.

2.1.1Macroscopicmodels.

Macroscopictrafficmodelsarebasedongas-kineticmodelsanduseequationsrelatingtrafficdensitytovelocity[LighthillandWhitham,1955,Helbingetal.,2002].Theseequationscanbeextendedwithtermsforbuild-upandrelaxationofpressuretoaccountforphenomenalikestop-and-gotrafficandspontaneouscongestions[Helbingetal.,2002,JinandZhang,2003,BrouckeandVaraiya,1996].Althoughmacroscopicmodelscanbetunedtosimulatecertaindriverbehaviors,theydonotofferadirect,flexible,wayofmodellingandoptimizingthem,makingthemlesssuitedforourresearch.

2.1.2Microscopicmodels.

Incontrasttomacroscopicmodels,microscopictrafficmodelsofferawayofsimulatingvariousdriverbehaviors.Amicroscopicmodelconsistsofaninfrastructurethatisoccupiedbyasetofvehicles.Eachvehicleinteractswithitsenvironmentaccordingtoitsownrules.Dependingontheserules,differentkindsofbehavioremergewhengroupsofvehiclesinteract.

CellularAutomata.Onespecificwayofdesigningandsimulating（simple）drivingrulesofcarsonaninfrastructure,isbyusingcellularautomata（CA）.CAusediscretepartiallyconnectedcellsthatcanbeinaspecificstate.Forexample,aroad-cellcancontainacarorisempty.Localtransitionrulesdeterminethedynamicsofthesystemandevensimplerulescanleadtochaoticdynamics.NagelandSchreckenberg（1992）describeaCAmodelfortrafficsimulation.Ateachdiscretetime-step,vehiclesincreasetheirspeedbyacertainamountuntiltheyreachtheirmaximumvelocity.Incaseofaslowermovingvehicleahead,thespeedwillbedecreasedtoavoidcollision.Somerandomnessisintroducedbyaddingforeachvehicleasmallchanceofslowingdown.ExperimentsshowedrealisticbehaviorofthisCAmodelonasingleroadwithemergingbehaviorsliketheformationofstart-stopwaveswhentrafficdensityincreases.

CognitiveMulti-AgentSystems.AmoreadvancedapproachtotrafficsimulationandoptimizationistheCognitiveMulti-AgentSystemapproach（CMAS）,inwhichagentsinteractandcommunicatewitheachotherandtheinfrastructure.Acognitiveagentisanentitythatautonomouslytriestoreachsomegoalstateusingminimaleffort.Itreceivesinformationfromtheenvironmentusingitssensors,believescertainthingsaboutitsenvironment,andusesthesebeliefsandinputstoselectanaction.Becauseeachagentisasingleentity,itcanoptimize（e.g.,byusinglearningcapabilities）itswayofselectingactions.Furthermore,usingheterogeneousmulti-agentsystems,differentagentscanhavedifferentsensors,goals,behaviors,andlearningcapabilities,thusallowingustoexperimentwithaverywiderangeof（microscopic）trafficmodels.

Dia（2002）usedaCMASbasedonastudyofrealdriverstomodelthedrivers’responsetotravelinformation.Inasurveytakenatacongestedcorridor,factorsinfluencingthechoiceofrouteanddeparturetimewerestudied.Theresultswereusedtomodeladriverpopulation,wheredriversrespondtopresentedtravelinformationdifferently.Usingthispopulation,theeffectofdifferentinformationsystemsontheareawherethesurveywastakencouldbesimulated.Theresearchseemspromising,thoughnoresultswerepresented.

Atrafficpredictionmodelthathasbeenappliedtoareal-lifesituation,isdescribedin[WahleandSchreckenberg,2001].Themodelisam