KDD2018滴滴派单算法论文。
Wepresentanovelorderdispatchalgorithminlarge-scaleon-demandride-hailingplatforms.Whiletraditionalorderdispatchapproachesusuallyfocusonimmediatecustomersatisfaction,theproposedalgorithmisdesignedtoprovideamoreefficientwaytooptimizeresourceutilizationanduserexperienceinaglobalandmorefarsightedview.Inparticular,wemodelorderdispatchasalarge-scalesequentialdecision-makingproblem,wherethedecisionofassigninganordertoadriverisdeterminedbyacentralizedalgo-rithminacoordinatedway.Theproblemissolvedinalearningandplanningmanner:1)basedonhistoricaldata,wefirstsummarizedemandandsupplypatternsintoaspatiotemporalquantization,eachofwhichindicatestheexpectedvalueofadriverbeinginaparticularstate;2)aplanningstepisconductedinreal-time,whereeachdriver-order-pairisvaluedinconsiderationofbothimmedi-aterewardsandfuturegains,andthendispatchissolvedusingacombinatorialoptimizingalgorithm.ThroughextensiveofflineexperimentsandonlineABtests,theproposedapproachdeliversremarkableimprovementontheplatform’sefficiencyandhasbeensuccessfullydeployedintheproductionsystemofDidiChuxing.
1