Diversity in Action, the SmartSociety perspective
-
Upload
smart-society-project -
Category
Science
-
view
48 -
download
2
Transcript of Diversity in Action, the SmartSociety perspective
MichaelRovatsosTheUniversityofEdinburgh
UniversityofLeeds,9thJune2016
DiversityinAcConTheSmartSocietyperspecCve
SmartSociety
� HybridandDiversity-AwareCollecCveAdapCveSystems:“whenpeoplemeetmachinestobuildasmartersociety”
� 4-year€6.8MFETIntegratedProject,co-ordinatedbyUniversityofTrento
� BringstogetherAI,computerscience,humanfactors,privacy,ethics
ProjectContext� Hybridity
� VisionofpeopleandmachinescollaboraCngandcomplemenCngeachothertotacklehardsocietalproblems
� Diversity� DiversepopulaConsofinteracCnghumansandmachineswithdifferentknowledge,skills,objecCves,andexpectaCons
� Collec1ves� HowdowecomposeindividualinteracConstoobtaincollecCveacConandgloballycoherentsocialcomputaCons?
� Adapta1on� HowcanweunderstandandsupportcollecCveadaptaConinacomplexsocio-technicalsystems?
Ask&Share� Anappfordoingthingstogether� withoutknowingthewhat/who/how
� Combines� human-drivencrowdsourcing� automatedacCvityrecommendaCon
� Currentscenariotourism,butaimforgeneralmodelofcollecCveproblemsolving
Where’sthenovelty?� ConvenConalappsrelyon
� StaCccontentsuppliedapriori� Search/filtering/recommendaCon� “Browse&pick”feel
� LimitaCons� Limitedscalability&flexibility� Nolearningfromhumaninput� Noexplicituser/placormgoals� NoincenCvisaCon� Notransparency
What’sinthebox(es)Programmingmodel
Accountability&transparency
IncenCves Privacy-awarenessAcCvity&contextrecogniCon
SocialOrchestra1on
The“agentprotocol”view
task execution
allocationtask
compositiontask
feedback
discoverypeer
po
FAILED AGREE(t)
ADVERTISE(C)
REQUEST(D)
REJECT(t)
FEEDBACK(t, F )
NO SOLUTION
SOLUTIONS(T )
AGREE(t)
START(t)
UPDATE(t)
TaskcomposiCon� CombinatorialproblemofallocaCnggroupsofuserstosharedtasks,wheretaskrequestscomefromusers
� HardconstraintsrestrictthegroupingsandtaskproperCesthatcanberealisedinprinciple
� So>constraintsdeterminewhichcoaliConstructuresandtaskfeaturesarepreferredbysystemand/orusers
� Examples:� Ridesharingwheredriverssharetheircarswithotherpassengersforaspecificjourney(ourvanillaexample)
� MeeCngscheduling,ciCzensciencetasks,workforceshigallocaCon,clinicalworkflowmanagement,etcetc
DiversityintaskcomposiCon� IntradiConalmechanismdesign,globalallocaConsarecomputedgivenindividualpreferencesandglobalcriteria� E.g.socialwelfaremaximisaCon,ParetoopCmality,strategy-proofness,etc.
� MechanismsareproposedthatprovablysaCsfytheseproperCes,soluConcanthereforebeimposedonusers
� Diversityimpliesthatuserscannotreporttheirpreferences� Systemnevercapturesallrelevantdecisionvariables� SoluConscannotbecomputed/consideredexhausCvely� UClityofsoluConscannotbedeterminedbyusersapriori
Keyproblems:1.Howtocompute“opCmal”setsofsoluCons2.Howtoinfluenceusers’choices3.Howtolearnusers’preferences
FromtaskallocaContotaskrecommendaCon
UserandsystemuClity
User's utility function depends on user’s requirements and preferences
Global (system) utility function depends on social welfare and maximising task completion
MIPfirst
MIPothers
MIP*
Learning
Objective
Objective
Objective
Constraints
Constraints
Constraints
Hard feasibility constraints
MIP*
MIPfirst
CompuCngallocaCons
We want to modify users' utility artificially so that their choices lead to a feasible global solution • Explicit Approaches:
• intervention • (possible) future reward
• Implicit Approaches: • discounts • taxation
Influencingusers
MIPfirst
MIPothers
MIP*
MIPfirst
Objective
Constraints
Noiseless and Constant Noise Models
Logit Model (also goes into objective function)
Sponsored Solution
TaxaCon
Learningusers’preferences� DevelopedamodelofcoarsepreferencesincombinaConwithusertypingtoallowforfastpreferenceelicitaCon
� ShouldallowustoopCmisequerieswhenrecommendingsoluConsusingexisingac1velearningmethods� PosequeriesthatalsomaximisetheexpectedvalueofinformaCon,andincludethisintheopCmisaConprocedure
� Challenge:notsuitableforcollecCves,onlyforsingleusers� BasicallyopCmisaConconstraintswouldhavetoincludefullBaysianupdateofprobabilitydistribuConsonpreferences
Conclusions� CreaCnglarge-scalecollecCveintelligencethatuCliseshumanandmachinestrengthsrequiresaddressingdiversityatdifferentlevels
� Harnessingthepowerofdiversityinvolves� RelaxingassumpConsthatotherwiseallowustoprovidehardguarantees
� Tacklingnewalgorithmicchallenges(andrepresentaConalonesIhavenotmenConed
Conclusions� Inreturn,wegetsystemsthatbelermodelabroaderrangeofhumandecisionmaking� IncreaseslikelihoodthatpeoplewillengagewithfuturecollecCveintelligencesystems
� IhaveomiledmanyotheraspectsofSmartSocietywherediversityalsocreatesnewchallenges� AcCvityrecogniCon,transparency,privacy,etc.
Promise
Surveillance
Manipulation
Collective intelligence
Personalisation
Humans as cheap labour
Peril
Man-machine collaboration
Promisesandperils