concretely

SCI论⽂写作常⽤句⼦、句式、词汇（总结）100

pproach,…

rk特点，特征

ideaunderlyingourmethodistotrainthemodel’sinitialparameterssuchthat…

spectto

ous声名狼藉的，众所周知的

enttheresultinTable1.

ppliedtometa-learningforRL

ilthealgorithminAlogrithm3.

ymaylackcomputationaleffiency

entnetworkscansupportmeta-learninginafullysupervidcontext.

-levelalgorithminthedeepRLspace

atureofthislineofwork

ality合理性

gence收敛性

getraveltime（withrespectto关于）

zingtrafficflowandminimizingtheaveragewaitingtimearethegoalsofintelligenttrafficcontrol.

erimentresultsshowaninspiringimprovement

uatetheeffectivenessandefficiencyofourMetaLight…

sbadontheintuitiveprincipleof…

20.,andrewardisameasureoftransportationefficiency.

portantly,

-badandactor-criticbadRLmethods

onameoftheexperiment实验的备忘录

ricaltransformationlikeflippingandrotation.对称变换，对折和翻转

own减⼩scaleup增⼤

orthwerefertotheversionofthealgorithmweputforwardasPPO+Demonstrations(PPO+D).

at,…Inaddition,…Meanwhile,…

cVolume:交通量

theyreachreasonableperformance.

sthevaluesoftheunenactionstoreasonablevalues,

31.…hasattractedincreasinginterestsrecently.

lly,thecommunicationprotocolbetweenagentsismanuallyspecifiedandnotaltered

duringtraining.通常地，…

ation，n.复现（实验）；(绘画等的)复制；拷贝

ms，同义词

,inordertoefficientlytakeadvantageofthecontextualinformation,wedesigna…

llelandinadvance.

rly,Zhangetal.[23]leveragebi-directionalLSTMtocapturethecontextualinformation.

ction,weformalizetheproblemofTTE,andthenintroducethefeaturesweudinthepropodframework.

edgraph有向图

podmethodtakesinsightsfromthegmentbadmethods,whichisabletoefficientlyreducethe

podmethodtakesinsightsfromthegmentbadmethods,whichisabletoefficientlyreducethe

respontimeofarequest.

’sfuturetrafficconditionisstronglycorrelatedwiththehistoricaltrafficconditionsofthelinkitlfaswellas

itsneighborlinks.

tethetrafficconditionsobrvedongraph…

nstratetheefficiencyofourmethod,weconductextensiveexperiments

with…,wedevelopanencoder-decoderframeworkfordynamicgraphs.

假设，认定，认为

tely,wepropoatwo-timescaledeeptemporalpointprocessmodelthatcapturestheinterleaveddynamicsof

theobrvedprocess.=Specifically

ctedgraph⽆向图，directedgraph有向图。

r,therecentoverwhelmingsuccessofconvolutionalneuralnetworks(CNNs)…

end(为此),wepropoanewneuralnetworkmoduledubbed…

edtoexistingmodulesoperatinginextrinsic…

and-designedfeaturesonpointcloudshavelongbeenpropodingraphicsandvision,

uctivelearning:unlabelleddataisthetestingdata（转导推理）

ivelearning:unlabelleddataisnotthetestingdata（归纳式学习）

thatreasoningproblemsfordynamicgraphscanbeforextrapolationorinterpolation

ervedprocessnamely：dynamicsofthenetwork(realizedastopologicalevolution)

csonthenetwork(realizedasactivitiesbetweennodes).

nglatentreprentationsofnodesingraphsisanimportantandubiquitoustaskwithwidespreadapplications

suchaslinkprediction,nodeclassification,andgraphvisualization.

way,wecantakeinsightsfromthegment-badapproaches,

59.…showsthatitisapracticalandrobustsolutionforlarge-scalereal-worldrvices.

rtempiricalresultsandanalysisinthissubction.

ively,thepressureofaninterctioncouldbeenas

62.,whichisdesignedtofullyexploitthejointrelationsofspatialandtemporalinformation.

paperweexploreasimpleneuralmodel,calledCommNet,thatus…

ofoneLSTMperpersondoesnotcapturehuman-humaninteractions.

t,whichusgraphattentionalnetworkstofacilitatecommunication.

neousandHeterogeneousScenariosinDifferentCities.

estofourknowledge,wearethefirsttougraphattentionalnetworkinthettingoftrafficsignalcontrol

paper,wepropotouthegraphattentionalnetwork[26]tolearnthedynamicsoftheinfluences,

69.…weextendtheprevioussingle-headattentionintheneuralnetworktomulti-headattentionasmuchrecentworkdid

[25,26].

teriordataliketheroadandweatherconditionmighthelptoboostmodelperformance.

icideaistolearnalow-dimensionalvectorforeachnode…

eusuallyreprentedasaquenceofgraphsnapshotsfromdifferenttimesteps(Leskovecetal.,2007).

riinabroadspectrumofproblemsrangingfrombiologyandparticlephysicstosocialnetworksand

recommendationsystems.

esstate-of-the-artperformanceonveraltransductiveandinductivepredictiontasksfordynamicgraphs.

etheplethoraofdifferentmodelsfordeeplearningongraphs,fewapproacheshavebeenpropodthusfarfor

dealingwithgraphsthatprentsomesortofdynamicnature(ngfeaturesorconnectivityovertime).

76.,thispaperunprecedentedlytreattheproblemoftrafficsimulationasalearningproblem,

(point)istoaccuratelyimitatetheobrvedvehiclebehaviorsand

comethedeficiencyin…,anaturalconsiderationistolearn…fromreal-worldobrvations,insteadof

exclusivelyrelyingonunrealisticphysicalmodels.

paper,wefollowthegenerativeadversarialimitationlearning(GAIL)[8]thatincorporates…,andextendittothe

multi-agentcontextinthetrafficsimulationproblem.

urework,wewillworkonimprovingthemodel’sgeneralizationabilityindifferenttrafficenvironment.

eriorresultsofBCismainlyduetoitsfailureinmodelinginteractionsbetweenvehicles,whichismoreobvious

inmulti-interctioncas.

paper,wepropoanovelGCNmodel,whichwetermasShortestPathGraphAttentionNetwork(SPAGAN).

83.…allowsforamoreinformativeandintactexplorationofthegraphstructureand…

ralltrainingworkflowofSPAGAN.

suallydividedintotwocategories,alapproacheslearnthenodeembedding

lapproachesdefineconvolutions

directlyonthegraph,like.

,weexplicitlymodel(显⽰建模)thetemporalcorrelationsofinteractionsbyadoptinganextraLSTM.

87.(GCN)andgatedgraphconvolutionneuralnetwork(GGNN)[18]havedemonstratedground-breakingperformanceon

manytaskslike…

eparallelprocessing，MPP.⼤规模并⾏处理

ultssuggestthatTCNsconvincinglyoutperformbalinerecurrentarchitectures…

etranslation,audiosynthesis,languagemodeling

uldcauthedegradationofpredictiveperformanceofthemodel.

rthnotingthat3DGATcanbeemployedtootherspatialtemporalapplicationsaswell.

93.3DGATilaboratelydesignedtofullyexploitthejointrelationsofspatialandtemporalinformation.

areConSTGATwithveralstrongbalinemethodsusingreal-worlddatats.

themcomparable,weuthesamettingof…

96.…suchascomputervision,naturallanguageprocessing,recommendersystems,andotherdomainsbyemployingthe

graphstructures.

ilthe3D-attentionmechanism…,whichisanextensionofgraphattentionnetwork[15],andisillustratedin

Figure3.

herimprovetheperformance,wecombine…and…

yzetheeffectofthecontextualinformationofaroute.

oanovelspatial-temporalgraphattentionnetwork,whichdealswiththegeographicinformationand

temporalinformationofthetrafficconditionssimultaneously.