cities and fleets - wordpress.com · 2018. 6. 14. · the current data science toolbox: •...
TRANSCRIPT
CitiesandfleetsDamonWischik
UNIVERSITY OF CAMBRIDGE
Dept.ofComputerScienceandTechnology
DRIVER CITY§ canchoosewhichroutetotake§ wantsthefastesttraveltime
§ providesinfrastructure§ (hopestomaximizetotalutility)
Theclassicmodelofrouting...[Wardropequilibrium]
...canleadtoperverseoutcomes,e.g.thecitybuildsanewroad,andeveryone’straveltimegoesup[Braess’sparadox]Thisisthe“priceofanarchy”.
USER RIDESHAREFLEET CITY§ canchoosemodeoftransport
§ wantslowestprice
§ cansetorigin-basedsurgemultipliers
§ needstorebalancethefleet§ wantstomaximizeprofit
§ setspublictransitfares§ (hopestomaximizetotalutility)
Inaworldwithmoredecisionmakers...
...thereismoreanarchyhencemorewaysforthingstogowrong,&alsonewoptionsforjoined-upcitymanagement.
Whatcouldgetbetter?Afleetoperator(ifit’sadominantplayer)willinternalizethecostofanarchy,soitwillavoidBraess’sparadox.
A B
C D
cityreducesA→Brailfare
£3
£2
£6
£2
Whatcouldgowrong?§ Therecanbeperverseoutcomes,inthespiritofBraess’sparadox
A B
C D
totalrailpassengers=20totalrideshareprofit=£38
£6
£2
£6
£2
£0.1
£1
£0.1
10pax
10pax
12pax
12pax
totalrailpassengers=10totalrideshareprofit=£21.93
Relatedwork:IntheInternet,weimplementedmultipathloadbalancing.Withtheright‘price’signals,thenetworkachievescompleteresourcepooling.
Threeflowssharefourresources,asthoughthenetworkweremadeupofasingleresource
resourcepoolingofroads
Therearenewoptionsforjoined-upcitymanagement
Can’talltrafficproblemsbesolvedwiththerightcongestionpricing?
Singapore’sElectronicRoadPricing
§ Thecitysimulatesavirtualroadnetworkwhosecapacityis95%ofwhat’sreallythere,andmeasurescongestion
§ Fleetoperatorsagreetosetroutesandpricesaccordingtovirtualcongestion[orairquality,…]
§ Thecitysendsreal-timevirtualcongestionsignals,andthefleetssendenoughdatathatthecitycanverifycompliance
§ Thestreetsarekeptfree-flowing
§ Inreturn,thefleetsarepermittedspecialaccesstorestrictedzones
Therearenewoptionsforjoined-upcitymanagement
What’sthewayforward?• Apps+algorithmsmovefasterthancities• FleetoperatorshavecrackedUX
(thusthey’vegotgreatdatasets+levers)• Eachcityhasitsown
issues,datasets,andcontrollevers
• Weshouldn’tjustbedevisingmodels• Weshoulddeviseadashboard
–adataplayground–foroperatorstoeasilyexploreperformance/policies/systemdesigns
• Betterthancongestioncharging,citiesandfleetscansolveproblemstogether
Whatdoyoudoinadataplayground?
WhatarethemostcommonstrategiesI’veused(asastatistician/modeller/programmer)andwhattoolswouldhavemademyjobeasier?
explore“userstories”simulatenewscenarios
compareandoptimizepolicies
Whatdoyoudoinadataplayground?
§ Thedataplaygroundshowsgranulardata(reconstructed,ifneedbe,usingmachinelearning)sothattheoperatorcanseeandmeasureeverything
byIanLewis,directorofsmartCambridge
§ Seereal-timedatafeedsandinferences§ Workwithallsortsofdatasets,sinceeachcityisdifferent
§ Intuitivelyexplorerichdata,especiallyabout“userstories”Example:howdotouriststravel?
§ Intuitivelyexplorerichdata,especiallyabout“userstories”Example:howdotouriststravel?
tripstothenightsafari
§ Intuitivelyexplorerichdata,especiallyabout“userstories”Example:howdotouriststravel?
safari-goers
§ Intuitivelyexplorerichdata,especiallyabout“userstories”Example:howdotouriststravel?
safari-goers
§ Intuitivelyexplorerichdata,especiallyabout“userstories”Example:howdotouriststravel?
tripsbysafari-goers
§ Intuitivelyexplorerichdata,especiallyabout“userstories”Example:howdotouriststravel?
tripsbysafari-goerssafari-goertruefalse
§ andsupportthisexplorationwithmachinelearning
§ i.e.thedataplaygroundprovidesaUXforinteractingwithhierarchicaldataExcelandTableauareorientedaroundtabulardataGoogleAnalyticsetc.areorientedaroundhierarchicaldatasets,buttheiranalysesarehardcoded
users
trips
waypoints
§ “Deconstruct”thesimulatorandembeditinthedataplaygroundbytreatingitascomposableoperationsondata
§ “Deconstruct”thesimulatorandembeditinthedataplayground:interactvisuallywiththesimulator’sinputandoutput
§ Short-circuitthe“data→model→simulation”pipeline:useresampleduserstoriesetc.
Thecurrentdatasciencetoolbox:
• emphasisonuserstories• ML-powereddataclustering+highlighting• composableempiricalsimulation
SQLplyr
pandasdata.table
ggplotd3
interactivevis
• everyqueryhasanaturalvisualization• interactwithviz⟺modifyquery
ThenextExcel:
spark
mlepdesim