reason1377.com

MoralpreferencesFrancescaRossiIBMT.
J.
WatsonResearchcenterfrossi@it.
ibm.
com1MotivationandIntroductionHowdohumansormachinesmakeadecisionWheneverwemakeadecision,weconsiderourpreferencesoverthepossi-bleoptions.
Also,inasocialcontext,collectivedecisionsaremadebyaggregatingthepreferencesoftheindividuals.
AIsystemsthatsupportindividualandcollectivedecisionmak-inghavebeenstudiedforalongtime,andseveralpreferencemodellingandreasoningframeworkshavebeendenedandexploitedinordertoproviderationalitytothedecisionpro-cessanditsresult.
However,littleefforthasbeendevotedtounderstandwhetherthisdecisionprocess,oritsresult,isethicalormoral.
Rationalitydoesnotimplymorality.
HowcanweembedmoralityintoadecisionprocessAndhowdoweensurethatthedecisionwemake,asanindividualoracollectivityofin-dividuals,aremoralInotherwords,howdowepassfromtheindividuals'personalpreferencestomoralbehaviouranddecisionmakingWhenwepassfromhumanstoAIsystems,thetaskofmodellingandembeddingmoralityandethicalprinciplesisevenmorevagueandelusive.
Aretheexistingethicaltheo-riesapplicablealsotoAIsystemsOnonehand,thingsseemeasiersincewecannarrowthescopeofanAIsystem,sothatthecontextualinformationcanhelpusindenethecorrectmoralvaluesitshouldworkaccordingto.
However,itisnotclearwhatmoralvaluesweshouldembedinthesystem,norhowtoembedthem.
Shouldwecodetheminasetofrules,orshouldweletthesystemlearnthevaluesbyobservingushumansPreferencesandethicaltheoriesarenotthatdifferentinonerespect:theybothdeneprioritiesoveractions.
So,canweuseexistingpreferenceformalismstoalsomodelethicalthe-oriesWediscusshowtoexploitandadaptcurrentpreferenceformalismsinordertomodelmoralityandethicstheories,aswellasthedynamicintegrationofmoralcodeintopersonalpreferences.
Wealsodiscusstheuseofmeta-preferences,sincemoralityseemstoneedawaytojudgepreferencesac-cordingtotheirmoralitylevel.
Itisimperativethatwebuildintelligentsystemswhichbe-havemorally.
Toworkandlivewithus,weneedtotrustsuchsystems,andthisrequiresthatweare"reasonably"surethatitbehavesmorally,accordingtovaluesthatarealignedtotheOnleavefromUniversityofPadova,Italyhumanones.
Otherwise,wewouldnotletarobottakecareofourelderlypeopleorourkids,noracartodriveforus,norwewouldlistentoadecisionsupportsysteminanyhealth-carescenario.
Ofcoursetheword"reasonable"makessensewhentheapplicationdomaindoesnotincludecriticalsitua-tions(likesuggestingafriendonasocialmediaoramovieinanonlinesellingsystem).
ButwhentheAIsystemishelping(orreplacing)humansincriticaldomainssuchashealthcare,thenweneedtohaveaguaranteethatnothingmorallywrongwillbedone.
Inthisextendedabstractweintroducesomeissuesinem-beddingmoralityintointelligentsystems.
Afewresearchquestionsaredened,withnoanswertothem,withthehopethatthediscussionraisedbythequestionswillshedsomelightontothepossibleanswers.
2PreferencemodellingandreasoningPreferenceshavebeenstudiedforalongtimeinAI,bothintheareaofknowledgerepresentationandinmulti-agentsys-tems.
Severalframeworkshavebeendenedtomodeldif-ferentkindsofpreferences,suchasqualitative(asin,e.
g.
,"Ipreferbluetored")andquantitativeones(asin,e.
g.
,"Igive5starstoBreakfastatTiffany'sand2starstoTerminator").
Ingeneralpreferencesaredeninganorderingoverasetofoptions.
Thisordercanbetotalandstrict,butinpracticeitmayhavealotoftiesandincomparability.
Whenthesetofoptionsisverylarge,andeachoptionisdenedbyasetoffeatures(suchasacar,whichcanbede-nedbyitmodel,itscolour,itsengine,etc.
),preferencescanbeexpressedoversinglefeaturesofsmallsetsofthem,ratherthanentireoptions(asin,e.
g.
,"IfIbuyaconvertible,Ipreferittoberedratherthanwhite").
Thisallowsforafasterandeasierpreferencespecicationphase,aswellasformoreef-cientpreferenceelicitation.
Severalwayshavebeendenedtopassfromsuchcompactwaystomodelpreferencesoverfeaturestothepreferenceorderingovertheoptions.
How-ever,itispossibletoreasonaboutsuchpreferenceswithoutgeneratingtheexponentiallylargeorderingovertheoptions,whichmakespreferencesreasoningtractableinsomecases.
Examplesofframeworktodothisareconstraints[Rossietal.
,2006],softconstraints[Mesegueretal.
,2005]andCP-nets[Boutilieretal.
,2004].
Onceanindividual'spreferencesoverthepossibleoptionsarespecied,weneedtobeabletondthemostpreferredoption,orthenextbestoption,ortocomparetwooptionsthatmaybepresentedtous.
Severalalgorithmstoperformssuchtaskshavebeendened[Brafmanetal.
,2010;Boutilieretal.
,2004].
Whenindividuals,orAIsystems,arepartofasocialen-vironmentandneedtomakecollectivedecisions,individ-ual'spreferencesareaggregated(forexampleviasomevot-ingrule)andanoptionischosenforthewholegroup.
Manyvotingruleshavebeendenedandstudied,aswellastheirproperties[Arrowetal.
,2002].
Issuessuchasmanipu-lation,control,bribery,aswellaspropertiessuchasfair-nessandunanimityhavelongbeinginvestigated,inordertodenedecisionsupportsystemsthatbehaveasdesired[Airiauetal.
,2011;Fargieretal.
,2012;Conitzeretal.
,2011;XiaandConitzer,2010;Langetal.
,2007;Pinietal.
,2011;Pozzaetal.
,2011;Gonzalesetal.
,2008;Maranetal.
,2013;PurringtonandDurfee,2007;LangandXia,2009].
3FrompreferencestomoralityTotrustanAIsystem,likeacompanionrobotoraself-drivingcar,weneedtobereasonablysurethatitbehavesmorally,ac-cordingtovaluesthatarealignedtothehumanones.
Other-wise,wewouldnotletarobottakecareofourelderlypeopleorourkids,noracartodriveforus,norwewouldlistentoadecisionsupportsysteminanyhealthcarescenario.
SoitisimperativethatweunderstandhowtoprovideAIsys-temswithmorality[MusschengaandvanHarskamp,2013;WallachandAllen,2009;Greeneetal.
,2016].
Moralityandethicalbehaviourarebasedonprioritisingac-tionsonthebasisofwhatismorallyrightorwrong.
Manyethicaltheorieshavebeendenedandstudiedinthepsychol-ogyliterature.
Theyincludethefollowingones:Consequentialism:Actionconsequencesareevaluatedinternsofascaleofgoodandbad,andanagentshouldchoosetheactionthatminimisethebadandmaximisesthegood.
VirtueEthics:Anagentshouldchooseactionsthatsat-isfysomepre-denedsetofvirtuesDeontologism:Actionsarepredenedasgoodorbad,andanagentshouldchoosethebestaction,nomattertheconsequences.
Nomatterwhichethicaltheoryonedecidestouse,theno-tionofrightandwrongofcoursedependsonthecontextinwhichhumans(ormachines)function,soformallyanethicaltheorycanbedenedasafunctionfromacontexttoapar-tialorderingoveractions.
Indeed,usuallywehaveapartialorderoveractions,sincesomeactionscouldbeincompara-bletoothers.
Asonemaynoticebylookingattheprevioussectiononpreferences,thisisnotthatdifferentfromwhatpreferencesdene:apartialorderoverpossibleoptions(ofactions,ordecisionsingeneral).
Soitmakessensetoinvesti-gatethepossibleuseofpreferenceframeworksinmodellingandembeddingmoralityintoAIsystems.
Researchquestion1:Areexistingpreferencemodellingandreasoningframeworksreadytobeusedalsotomodelandreasonwithethicalprinciplesandmoralcode,orweneedtoadaptthemorinventnewonesIfwehadthe"moral"partialorderandthe"preference"partialorderforeachindividual,onecouldtrytomergetheminsomeway,toobtaina"moralpreferenceordering".
Forex-ample,twoCP-netsmodellingthemoralandthepreferenceorderingscouldbesyntacticallyorsemanticallymergedviaoperatorsthatcouldgiveprioritytothemoralCP-netandletthepreferenceonedictatethebehaviouronlywhenitisnotinconictwiththemoralone.
Thetechnicaldetailshavenotbeenspelledoutyet,butonecouldimagineseveralreason-ablewaysofdoingthis.
Researchquestion2:Givenamoralandanethicalorder-ingoveractions,howtocombinethemGivensuchorder-ingsintheformsofCP-netsorsoftconstraints,orothercom-pactformalismstomodelpreferences,howtocombinethemWhatpropertiesshouldwedesireabouttheircombinationHowever,knowingthepreferencesofanindividualisal-readyadifculttask.
Elicitationandlearningframeworkhavebeeproposedinordertodothatinawaythatismostfaith-fultothe"real"preferencesoftheindividual.
Knowingthemoralorderingofanindividualisevenmoredifcult.
Andthisisevenmoresowhenweareinasocialcontext,sincethismaymakeindividualschangetheirmoralattitudesovertimebecauseofsocialinteraction.
TheexistingapproachestodeneethicalprinciplesinAIsystemsrangefromtryingtocodeethicalprinciplesintheformofrules,tolettingthesystem"learn"suchprinciplesfroma(possiblysupervised)observationofthebehaviourofhumansinsimilarsettings.
SomeAIsystemstrytolistthesetofrulestouseinself-drivingcarstosolveethicaldilemmaslikethetrolleyprob-lem.
However,suchapproachesareusuallynotgeneral,sinceitisunfeasibletoforeseeallpossiblesituationsinaverywidescenario.
Ontheotherhand,otherapproachesuse,forexam-ple,inversereinforcementlearning[NgandRussell,2000]totrytolearnmoralityfromhumanbehaviour.
Ipersonallyfeelthatthebestresultscouldbeobtainedbycombiningthesetwoapproaches,althoughitisnotclearyethowtodoitbest.
Researchquestion3:Howtocombinebottom-uplearningapproacheswithtop-downrule-basedapproachesindeningethicalprinciplesforAIsystemsResearchquestion4:Recently,themostsuccessfulAIsystemsarebasedonstatisticalmachinelearningapproachesthat,bytheirnature,donotprovideanaturalwaytoexplainorjustifytheirdecisions(orsuggestions),northeyassureopti-mality.
Ifweemploythisapproachalsoforembeddingmoral-ityintoamachine,howarewegoingtoprovethatnothingmorallywrongwillhappen4Moralitybymeta-preferencesAsmentionedabove,inasocialcontext,individualprefer-encesaretransformedlittlebylittlebyincorporatingreason-ableelementsfromthesocietalinteractionwithothermem-bersofthegroup.
Thisisoftencalled"reconciliation"ofin-dividualpreferenceswithsocialreason,andtakesplaceinthecontextofcollectivechoice.
Tobeabletodescribethedynamicmovingfromonepreferenceorderingoverthenextone(intime),andtomakesurethatthelaterpreferenceor-deringsareindeedbetterintermsofmorality,oneneedstohaveawaytojudgepreferencesaccordingtosomenotionofgoodandbad(inanyoftheabovementionedethicaltheo-ries).
Indeed,Sen[Sen,1974]claimsthatmoralityrequiresjudgementamongpreferences.
Toaccountforthis,hein-troducedthenotionofmetaranking(thatis,preferencesoverpreferences)whichenablestoformaliseindividualpreferencemodications.
Amoralcodecouldthenbedenedasrankingofpreferencerankings.
Thatis,themoralcodeisdenedbyastructurethat,byemployingnotionssuchasdistance,isabletorankpreferencesaccordingtotheirmoralitylevel.
Thedistanceintrinsicinthemoralcodecanthenbeusefulinmeasuringthedeviationofanysocialorindividualactionfromthemoralcodeitself.
Researchquestion5:Givenamoralcode,inasocialchoicecontext,whereindividualssubmittheirpreferenceor-deringandtheresultisacollectivepreferenceordering,howtomeasurethedeviationofthecollectiveorderingfromamoralcodeAndhowtomeasurethedeviationofindivid-ualsfromacollectivemoralcodeIfanindividualmodiesitspreferenceorderingfromamorallylowtoamorallyhigherordering,weshouldwanttousecollectivedecisionmakingsysteminwhichsuchamoveleadstocollectiveactionsofhighermorality.
Thatis,someformofmonotonicityshouldbedesired.
Researchquestion6:Whichpropertiesshouldbedesiredinamoralpreferenceaggregationenvironment5MoralityinnarrowAIsystemsIn[Greene,2014]itisshownthathumanmoraljudgmentdoesn'tcomefromadedicatedmoralsystem,butitisrathertheproductoftheinteractionofmanygeneral-purposebrainnetworks,eachworkingandbeingusefulinnarrowcontexts.
Soitseemsthathumansneedageneralpurposebraininordertobemoral.
IsittruealsoforAIsystemsResearchquestion7:CannarrowAIsystemsbemoralIfhumansbringalloftheirgeneralintelligencetobearwhenmakingmoraldecisions,evenfairlysimpleones,doesthatthatmeanthatwehavetosolveArticialGeneralIntelligenceinordertoproducesomethinguseful6ConclusionsIntelligentsystemsaregoingtobemoreandmorepervasiveinoureverydaylives.
Tonamejustafewapplications,theywilltakecareofelderlypeopleandkids,theywilldriveforus,andtheywillsuggestdoctorshowtocureadisease.
How-ever,wecannotletthemdoallthisveryusefulandbenecialtasksifwedon'ttrustthem.
Tobuildtrust,weneedtobesurethattheyactinamorallyacceptableway.
Soitisimpor-tanttounderstandhowtoembedmoralvaluesintointelligentmachines.
Existingpreferencemodellingandreasoningframeworkcanbeastartingpoint,sincetheydeneprioritiesoverac-tions,justlikeanethicaltheorydoes.
However,manymoreissuesareinvolvedwhenwemixpreferences(thatareatthecoreofdecisionmaking)andmorality,bothattheindividuallevelandinasocialcontext.
Wehavelistedsomeofthesequestions,hopingthatthisshortpapercangeneratesomean-swers.
References[Airiauetal.
,2011]S.
Airiau,U.
Endriss,U.
Grandi,D.
Porello,andJ.
Uckelman.
Aggregatingdependencygraphsintovotingagendasinmulti-issueelections.
InPro-ceedingsofIJCAI2011,pages18–23,2011.
[Arrowetal.
,2002]K.
J.
Arrow,A.
K.
Sen,andK.
Suzu-mura.
HandbookofSocialChoiceandWelfare.
North-Holland,2002.
[Boutilieretal.
,2004]C.
Boutilier,R.
I.
Brafman,C.
Domshlak,H.
H.
Hoos,andD.
Poole.
CP-nets:Atoolforrepresentingandreasoningwithconditionalceterisparibuspreferencestatements.
JAIR,21:135–191,2004.
[Brafmanetal.
,2010]R.
I.
Brafman,F.
Rossi,D.
Salvagnin,K.
B.
Venable,andT.
Walsh.
Findingthenextsolutioninconstraint-andpreference-basedknowledgerepresen-tationformalisms.
InProceedingsofKR2010,2010.
[Conitzeretal.
,2011]V.
Conitzer,J.
Lang,andL.
Xia.
Hy-percubewisepreferenceaggregationinmulti-issuedo-mains.
InProceedingsofIJCAI2011,pages158–163,2011.
[Fargieretal.
,2012]H.
Fargier,J.
Lang,J.
Mengin,andN.
Schmidt.
Issue-by-issuevoting:anexperimentaleval-uation.
InProceedingsofMPREF2012,2012.
[Gonzalesetal.
,2008]C.
Gonzales,P.
Perny,andS.
Queiroz.
Preferenceaggregationwithgraphicalutilitymodels.
InProceedingsofAAAI2008,pages1037–1042,2008.
[Greeneetal.
,2016]JoshuaGreene,FrancescaRossi,JohnTasioulas,KristenBrentVenable,andBrianWilliams.
Embeddingethicalprinciplesincollectivedecisionsup-portsystems.
InProceedingsAAAI2016.
AAAIPress,2016.
[Greene,2014]JoshuaGreene.
Thecognitiveneuroscienceofmoraljudgmentanddecisionmaking.
InTheCognitiveNeurosciencesV(ed.
M.
S.
Cazzaniga).
MITPress,2014.
[LangandXia,2009]J.
LangandL.
Xia.
Sequentialcompo-sitionofvotingrulesinmulti-issuedomains.
Mathemati-calsocialsciences,57:304–324,2009.
[Langetal.
,2007]J.
Lang,M.
S.
Pini,F.
Rossi,K.
B.
Ven-able,andT.
Walsh.
Winnerdeterminationinsequentialmajorityvoting.
InProceedingsofIJCAI2007,pages1372–1377,2007.
[Maranetal.
,2013]A.
Maran,N.
Maudet,M.
S.
Pini,F.
Rossi,andK.
B.
Venable.
Aframeworkforaggregat-inginuencedCP-netsanditsresistancetobribery.
InProceedingsofAAAI2013,2013.
[Mesegueretal.
,2005]P.
Meseguer,F.
Rossi,andT.
Schiex.
Softconstraints.
InP.
VanBeekF.
RossiandT.
Walsh,editors,HandbookofConstraintProgramming.
Elsevier,2005.
[MusschengaandvanHarskamp,2013]BertMusschengaandAnton(eds.
)vanHarskamp.
WhatMakesUsMoralOnthecapacitiesandconditionsforbeingmoral.
Springer,2013.
[NgandRussell,2000]AndrewY.
NgandStuartRussell.
Algorithmsforinversereinforcementlearning.
InPro-ceedingsoftheSeventeenthInternationalConferenceonMachineLearning.
MorganKaufmann,2000.
[Pinietal.
,2011]M.
S.
Pini,F.
Rossi,K.
B.
Venable,andT.
Walsh.
Incompletenessandincomparabilityinprefer-enceaggregation:Complexityresults.
Artif.
Intell.
,175(7-8):1272–1289,2011.
[Pozzaetal.
,2011]G.
DallaPozza,M.
S.
Pini,F.
Rossi,andK.
B.
Venable.
Multi-agentsoftconstraintaggregationviasequentialvoting.
InProceedingsofIJCAI2011,pages172–177,2011.
[PurringtonandDurfee,2007]K.
PurringtonandE.
H.
Dur-fee.
Makingsocialchoicesfromindividuals'CP-nets.
InProceedingsofAAMAS2007,pages1122–1124,2007.
[Rossietal.
,2006]F.
Rossi,P.
VanBeek,andT.
Walsh.
HandbookofConstraintProgramming.
Elsevier,2006.
[Sen,1974]AmartyaSen.
Choice,orderingandmorality.
InPracticalReason,KrnerS.
(ed).
Oxford,1974.
[WallachandAllen,2009]WendellWallachandColinAllen.
MoralMachines.
Oxford,2009.
[XiaandConitzer,2010]L.
XiaandV.
Conitzer.
Strategy-proofvotingrulesovermulti-issuedomainswithrestrictedpreferences.
InProceedingsofWINE2010,pages402–414,2010.