Escolar Documentos
Profissional Documentos
Cultura Documentos
BayesianEpistemology
AlanHjekandStephanHartmann
1.Introduction
Bayesianismisourleadingtheoryofuncertainty.Epistemologyisdefinedasthe
Bayesianism, after all, studies the properties and dynamics of degrees of belief,
understoodtobeprobabilities.Traditionalepistemology,ontheotherhand,placesthe
singularlynonprobabilisticnotionofknowledgeatcentrestage,andtotheextentthat
ittraffics inbelief,thatnotiondoesnotcomeindegrees.Sohowcantherebea
Bayesianepistemology?
Accordingtooneview,therecannot:Bayesianismfailstodojusticetoessential
traditional epistemology: where the latter has been mired in endless debates over
illuminatevariouslongstandingproblemsofepistemology,whilenotaddressingall
ofthem;andwhileBayesianismopensupfascinatingnewareasofresearch,itbyno
meansclosesdownthestaplepreoccupationsoftraditionalepistemology.
Thecontrastbetweenthetwoepistemologiescanbetracedbacktothemid17th
justificationsforhisclaimstoknowledgeinthefaceofpowerfulskepticalarguments.
No more than four years after his death, Pascal and Fermat inaugurated the
11/23/08 2
probabilistic revolution, writ large in the PortRoyale Logic, in which the many
shadesofuncertaintyarerepresentedwithprobabilities,andrationaldecisionmaking
isamatterofmaximizingexpectedutilities(aswenowcallthem).Correspondingly,
theCartesianconcernforknowledgefadesintothebackground,andamorenuanced
representationofepistemicstateshasthelimelight.Theisticbeliefprovidesavivid
exampleofthecontrastingorientations.Descartessoughtcertaintyintheexistenceof
Godgroundedinapodeicticdemonstrations.Pascal,bycontrast,explicitlyshunned
suchallegedproofs,arguinginsteadthatoursituationwithrespecttoGodislikea
gamble,andthatbeliefinGodisthebestbetthusturningthequestionoftheistic
BayesianepistemologyowesitsnametotheReverendThomasBayes,who,acentury
epistemologywouldbefarmoreappropriate,aswewillsee.
plethoraofpositionsasiftheyformamonolithicwhole.Otherarticlesinthisvolume
distinguishcarefullyamongvariouspositionsthatourbroadbannerconflates.Forour
purposes,theystartoutregardingknowledgeandbeliefasthecentralconceptsof
epistemology,andthentogoontostudytheproperties,grounds,andlimitsofthese
binarynotions.WealsospeakofBayesianismasifitisaunifiedschoolofthought,
wheninfacttherearenumerousintramuraldisputes.I.J.Good(1971)calculatesthat
thereare(atleast)46,656waystobeaBayesian,whilewewillmostlypretendthat
thereisjustone.Byandlarge,thevariousdistinctionsamongBayesianswillnot
11/23/08 3
matter for our purposes. As a good (indeed, a Good) Bayesian might say, our
traditionalproblemscanbeframed,andprogresscanbemadeonthem,usingthe
tools of probability theory. But Bayesian epistemology does not merely recreate
traditionalepistemology;thankstoitsconsiderableexpressivepower,italsoopensup
newlinesofenquiry.
2.WhatisBayesianEpistemology?
Kolmogorovs(1933)axiomatization.Let beanonemptyset.Afield(algebra)on
realnumbersobeying:
1)P(a)0forallaF. (Nonnegativity)
CallPaprobabilityfunction,and( ,F,P)aprobabilityspace.
Onecouldinsteadattachprobabilitiestomembersofacollectionofsentencesofa
formallanguage,closedundertruthfunctionalcombinations;thisismorecommonin
Bayesian confirmation theory. A lively area of current debate concerns just how
finely grained such contents of probability attributions should be. For example,
variousproblemsofselflocationsuggestthatprobabilitiesshouldattachtocentred
11/23/08 4
propositions,e.g.<possibleworld,individual,time>triples.
unconditionalprobabilities:
P(a b)
P(a|b)= ,providedP(b)>0.
P(b)
P( b | a)P(a )
P(a|b) =
P( b)
P(b | a)P(a)
=
P(b | a)P(a) + P(b | a)P( a)
More generally, suppose there is a partition of hypotheses {h1, h2, ... , hn}, and
evidencee.Thenforeachi,
P(e | hi )P(hi )
P(hi|e)=
P(e | h )P(h ) .
j j
j=
Joyce(2008)foramoredetaileddiscussionofvarietiesandusesofBayestheorem.
pieceofevidenceeandahypothesish:
econfirmsh(relativetoP)iffP(h|e)>P(h).
various measures of evidential support (see Eells and Fitelson 2000 and Fitelson
1999).
11/23/08 5
Bayesianismcanbeunderstoodascombiningasynchronicthesisaboutthedegrees
ofbelieforcredencesofarationalagentatagiventime,andadiachronicthesisabout
howtheyevolveinresponsetoevidence.Synchronically,theagentscredencesare
function Pinitial, and that she becomes certain of e (where e is the strongest such
proposition).ThenhernewcredencefunctionPnewisrelatedtoPinitialasfollows:
(Conditionalization) Pnew(x)=Pinitial(x|e)(providedPinitial(e)>0).
Jeffreyconditionalization allowsforlessdecisivelearningexperiencesinwhich
herprobabilitiesacrossapartition{e1,e2,...}changeto{Pnew(e1),Pnew(e2),..., },
wherenoneofthesevaluesneedbe0or1:
Pnew(x)= iPinitial(x|ei)Pnew(ei)(providedPinitial(ei)>0).
(Jeffrey1983).Pnew(x)iscalledtheposteriorprobabilityfunction.
functionandrepeatedlyupdatesby(Jeffrey)conditionalizationasevidencecomesin.
epistemological odyssey with considerable rigidity about how the agent should
respondtoevidence.ButBayesianismisathemethatadmitsofmanyvariationssee
Good(1971)andHjekandHall(2002)fordiscussionofsomeofthem.Here,letus
considerseveralfurtherconstraintsonpriorsthathavebeenproposed.
tautologies,andprobability0onlytocontradictionstoallothersentencesitassigns
intermediatevalues.Itseemstobeanepistemologicaldesideratumthatapriorbe
regular,reflectinganopenmindednessappropriatetoanagentwhoisatabularasa.
11/23/08 6
logicalpossibilitywouldbetopretendthatonesevidencewasstrongerthanitreally
evidence in your credences: you are required to give equal probability to all
possibilitiesamongwhichyourevidencedoesnotdiscriminate(andinastateoftotal
evidentialinnocence,thatisallofthem).Asophisticatedversionoftheprincipleof
indifference,favouredbysocalledobjectiveBayesians,hasbeenexploredbyJaynes
Williamson(2005).
Thentherearetwoprinciplesthataremeanttocodifyonesepistemiccommitment
toaligningonescredencestocertainprobabilistichypotheses.Inthefirst,Lewiss
PrincipalPrinciple(heresimplified),thehypothesesconcerntheobjectivechanceof
therelevantpropositions(Lewis1980):
C0(a|cht(a)=x)=x(forallaandtforwhichthisisdefined).
HereC0issomereasonableprior,aanarbitraryproposition,andcht(a)=xtheclaim
thatthechanceattimetofaisx.Theideaisthatoneshouldalignonescredences
withwhatonetakesthecorrespondingobjectivechancestobe,wherethelatterare
genuineprobabilitiesintheworld.Inthesecond,vanFraassensReflectionPrinciple
(van Fraassen 1984), the hypotheses concern ones own future credences for the
relevantpropositions
HereCtisonesprobabilityfunctionattimet,andCt onesfunctionatlatertimet .
11/23/08 7
Theideaisthatrationalityrequiresacertaincommitmenttoonesfutureopinions;
when all is going well, ones future selves are betterinformed versions of ones
currentself.
3.ContrastsbetweenTraditionalEpistemologyandBayesianEpistemology
WecannowbringoutseveralpointsofcontrastbetweentraditionalandBayesian
epistemology.Wehavenotedthatknowledgeandbeliefarebinarynotions,tobe
contrastedwiththepotentiallyinfinitelymanydegreesofcredence(corresponding
toalltherealnumbersinthe[0,1]interval).Knowledgeisnotmerelyjustifiedtrue
fourthconditionwillcompletetheanalysissomekindofconditionthatrulesout
casesinwhichonehasajustifiedtruebeliefbyluck,orforsomeanomalousreason.
Someepistemologists(e.g.Sosa(1999))advocateversionsofsafetyasaconditionon
knowledgeroughly,attheclosestworldsinwhichagivenagentbelieves p,p is
true.Others(e.g.Nozick(1981))advocatesensitivityroughly,attheclosestworlds
similarityofworldsthathastakencentrestageintheanalysisofknowledge.
Noticethattruth,justification,andtheseantiluckconditionsmaybecharacterized
as at least partially objective, with belief providing the only purely subjective
analyzes this doxastic notion, but which has no clear analogue of the objective
conditions.Mostimportantly,Bayesianismapparentlyhasnothingthatcorrespondsto
11/23/08 8
thefactivityofknowledge:thatonecanonlyknowtruths.Andevenwhenourbeliefs
fallshortofknowledge,stillitisadesideratumthattheybetrue;buttheBayesian
seemstohavenocorrespondingdesideratumforintermediatecredences,whichareits
stockintrade.Whenyouassign,forexample,probability0.3toitrainingtomorrow,
proposal)oranyantiluckconditiononknowledgemoreonthisshortly.
parentheticalreferencesto P. Manyauthorssuppressthesereferences,encouraging
onetoforgettheirinherentsubjectivity(andevenwesecretedthemawayinslightly
disingenuousparentheses!).Traditionalepistemologists,bycontrast,conductmuchof
reliable,andsoonwithoutanyqualification.Andagain,truth,justification,andanti
luckconditionsaretypicallysupposedtoholdornotindependentlyofwhethersome
agentthinksthattheydo,orwhethersomemodelsaystheydo.
Thesynchronicrequirementthatanagentscredencesobeytheprobabilitycalculus
epistemologythatonesbeliefsshouldbeconsistent.Thediachronicrequirementof
conditionalizationisreminiscentoftheQuineanprincipleofminimummutilation
(change beliefs as little as the evidence allows)a conservative recipe for belief
indifferencecorrespondsveryroughlytoaCartesianadmonitiontosuspendjudgment
whenonesevidenceislacking,butitisfarmorespecific.Andtherearenotraditional
analoguesofthevariousadditionalconstraintsonpriors.Goingintheotherdirection,
Bayesianismissilentaboutsomeofthecornerstonesandmorerecentconcernsof
traditionalepistemologywewilldiscussthisatgreaterlengthattheend.
GiventhestrikingdifferencesbetweentraditionalandBayesianepistemology,are
therereasonstopreferonetotheother?
4.Thesis:BayesianEpistemologyIsSuperiortoTraditionalEpistemology
Jeffrey, a famous Bayesian, suggests two main benefits accrued by the Bayesian
frameworkinhis(1992):
1.Subjectiveprobabilitiesfigureindecisiontheory,anaccountofhowouropinions
andourdesiresconspiretodictatewhatweshoulddo.Thedesirabilityofeachofour
possibleactionsismeasuredbyitsexpectedutility,aprobabilityweightedsumofthe
utilitiesassociatedwiththataction.TocompleteJeffreysargument,weshouldadd
Pascal).Theanalysisofrationalactionsurelyneedstoadverttomorefinegrained
mentalstatesthanbinarybeliefandknowledge.(SeeErikssonandHjek2007for
morediscussionofwhytheintermediatecredencesthatarenecessaryforthatanalysis
cannotbereducedtothesebinarynotions.)
2.Observationsrarelydelivercertaintiesrather,theireffectistypicallytoraiseour
11/23/08 10
probabilitiesforcertainpropositions(andtodropourprobabilitiesforothers),without
anyreachingtheextremesof1or0.Traditionalepistemologyapparentlyhasnoway
conditionalizationistailormadetodoso.
We may continue the list that Jeffrey has started of putative advantages of
Bayesianismovertraditionalepistemologyatsomelength:
3.Knowledgeisunforgiving.Itsstandardsaresohighthattheycanrarelybemet,at
leastincertaincontexts.(Thisisrelatedtothefactthatknowledgedoesnotcomein
degreesnearknowledgeisnotknowledgeatall.)Thisinturnplaysintothehands
of skeptics. But it is harder for skeptical arguments to get a toehold against the
underminesaclaimofknowledgeregardingX,butitisinnocuousfromaprobabilistic
pointofview:anagentcansimplyassign X somesuitableprobabilitylessthan1.
Indeed,evenanassignmentofprobability1isconsistentwiththepossibilityoferror
underplausibleassumptions,itcanbeshownthatadartthrownatrandomata
representationofthe[0,1]intervalhasprobability1ofhittinganirrationalnumber,
eventhoughitmightfailtodoso.
4.Moreover,itisaplatitudethatdoxasticstatescomeindegrees,andthecategories
believe,amongotherthings,that2+2=4,thatyouhaveahand,thatLondonisin
England, and (say) that Khartoum is in Sudan. But you do not have the same
11/23/08 11
confidenceinallthesepropositions,aswecaneasilyrevealinyourbettingbehavior
propositionsforwhichyouhavelessconfidencethatthiscoinwilllandheads,that
itwillraintomorrowinNovosibirsk,andsoon.Wemayconflateyourattitudesto
them all as suspensions of belief (as Descartes would), but that belies their
underlyingstructure.Suchattitudesarebetterunderstoodassubjectiveprobabilities.
comparativelylittleofourreasoningcanbecapturedbyit,eitherinscienceorindaily
life(pace PopperandHempel).Afterall,whetherwelikeitornot,ourepistemic
practicesconstantlybetrayourcommitmenttorelationsofsupportthatfallshortof
entailment(OaksfordandChater2007).Wethinkthatitwouldbeirrationaltodeny
thatthesunwillrisetomorrow,toprojectgrueratherthangreeninourinductions,
andtocommitthegamblers fallacy.Probabilitytheoryhelpsustounderstandsuch
relations.
century ofworkinprobabilitytheoryandstatistics.Traditionalepistemologymay
comparabletotheformidableformalmachinerythatwefindintheBayesianstool
kit.
7. Bayesian methods, in turn, have much wider application than any formal
11/23/08 12
engineering,andartificialintelligenceifyouneedanyconvincingofthis.
8.TherearemanyargumentsforBayesianism,whichcollectivelyprovideakindof
defenseofthethesisthatrationalcredencesareprobabilities.Anagentscredences
areidentifiedwithherbettingprices;itisthenshownthatsheissusceptibletosure
losses iff these prices do not conform to Kolmogorovs axioms. There are also
Savage1954,Joyce1999),fromcalibration(vanFraassen1984),fromgradational
accuracyorminimizationofdiscrepancyfromtruth(Joyce1998),fromqualitative
constraintsonreasonableopinion(seeEarman1996foradiscussionofresultsofCox
and others), and so on. Moreover, there are various arguments in support of
(Armendt 1980, Lewis 1999) and arguments from minimal revision of ones
traditionalepistemology.
9. Finally,apragmaticargumentforBayesianismcomesfromanevaluationofits
fruits.Asweshowatgreaterlengthinsection7,Bayesianismishighlyexplanatory
withminimalresourcesasimple,fecundtheoryifevertherewasone.Traditional
epistemologyishardpressedtoofferthesamerewards.Forexample,wewillseehow
analysis,andsomeerroneousintuitionscanbecorrected.Itseemsthatnoanalysis
11/23/08 13
couchedpurelyintermsofknowledgeandbeliefcouldpaysuchdividends.
epistemology.Butthisdoesnottellthewholestory.Forstarters,thetriumphsof
Bayesian confirmation theory just touted are supposedly offset by the socalled
confirmanythingbyBayesianlights:inthatcase,P(h|e)=P(he)/P(e)=P(h).Yet
weoftenthinkthatsucholdevidencecanbeconfirmatory.Considertheevidenceof
theadvanceoftheperihelionofMercury,whichwasknowntoEinsteinatthetime
thatheformulatedgeneralrelativitytheory,andthus(wemayassume)wasassigned
confirmatoryofgeneralrelativitytheory.ThechallengeforBayesiansistoaccount
forthis.(SeeZynda(1995)fordiscussion.)
Bayesianism,then,isnotwithoutproblemsofitsown.Soletusrevisitthecontest
betweentraditionalepistemologyandBayesianism,thistimelookingatarguments
thatinclineinfavoroftheformer.
Epistemology
1.Bayesiansintroduceanewtechnicalterm,degreeofbelief,buttheystruggleto
explicateit.Tobesure,theliteratureisfullofnodstobettinginterpretations,but
behaviorismthatfocusessolelyontheratherpeculiarkindofbehaviorthatismostly
foundatracetracksandcasinos.Othercharacterizationsofdegreeofbeliefthatfall
outofdecisiontheoreticrepresentationtheoremsarealsoproblematic.(SeeEriksson
11/23/08 14
explication.
2.Recalltheabsenceofanynotionoftruthofanintermediatedegreeofbelief.Yet
truthistheveryaimofbelief.Itisusuallythoughttoconsistincorrespondencetothe
waythingsare.Moreover,wewantourmethodsforacquiringbeliefstobereliable,in
correspondence, and notion of reliability for the Bayesian? The terms of her
example,onecanassignveryhighprobabilitytotheperiodattheendofthissentence
beingthecreatoroftheuniversewithoutincurringanyBayesiansanction:onecando
sowhileassigningcorrespondinglylowprobabilitytotheperiodnotbeingthecreator,
andwhiledutifullyconditionalizingonalltheevidencethatcomesin.Traditional
epistemologyisnotsotolerant,andrightlynot.
3. Relatedly, the Bayesian does not answer the skeptic, but merely ignores him.
Bayesianismdoesntmakeskepticalpositionsgoaway;itmerelymakesthemharder
tostate.
4.TheBayesiansimilarlylacksanotionofjustificationortotheextentthatshe
hasone,itistoopermissive.AtleastonwhatwehavecalledBayesianorthodoxy,any
priorisasuitablestartingpointforaBayesianodyssey,yetmereconformitytothe
probabilitycalculusisscantjustification.
Now,theBayesianwillbequicktoanswerthisandthepreviousobjectionsina
11/23/08 15
singlestroke.Shewillappealtovariousconvergencetheorems.Forexample:
If observations are precise then the form and properties of the prior
distribution have negligible influence on the posterior distribution. From a
practicalpointofview,then,theuntrammeledsubjectivityofopinionceases
toapplyassoonasmuchdatabecomesavailable.Moregenerally,twopeople
withwidelydivergentprioropinionsbutreasonablyopenmindswillbeforced
intoarbitrarilycloseagreementaboutfutureobservationsbyasufficientamount
ofdata.(Edwardsetal.1963:201)
conditionalizingthepeoplespriorsonthedata.GaifmanandSnir(1982)similarly
showthatforeachsuitablyopenmindedagent,thereisadatasetsufficientlyrichto
forceherarbitrarilyclosetoassigningprobability1tothetruememberofapartition
parlay these theorems into providing surrogates for that fourth condition for
knowledge, insisting that such convergences do not happen by luck, or for some
anomalousreason,butareprobabilisticallyguaranteed.
Thesearebeautifultheorems,butoneshouldnotoverstatetheirepistemological
significance. They are glass halffull theorems, but a simple reversal of the
quantifiersturnsthemintoglasshalfemptytheorems.Foreachdataset,thereisa
convergence:afterconditionalizingherprioronthedataset,sheisstillnowherenear
assigningprobability1tothetruehypothesis,andstillnowherenearagreementwith
suitablyopenmindedagentandsufficientlyrichdataset.Nodataset,however
rich,willdriveanywhereatalladogmaticagentwhoconcentratesallcredenceona
singleworld(maximallyspecifichypothesis).Worse,anagentwithawackyenough
11/23/08 16
priorwillbedrivenawayfromthetruth.Considersomeonewhostartsbygivinglow
probabilitytobeingabraininavat,butwhosepriorregardsalltheevidencethatshe
actually gets as confirming that she is. And we can always come up with rival
irresolvableconflictbetweenanatheistandacreationistwhoseesGodshandiwork
ineverything.
5.AproponentofBayesianismmaydescribeitasfecund(andwedid);butan
opponentmaydescribeitasempty.Withsolittleconstraintonpriors,itisnot
surprisingthatBayesianismaccountsforsomuch.Indeeditisirrationaltodenythat
thesunwillrisetomorrow,toprojectgrueratherthangreeninourinductions,and
tocommitthegamblers fallacy.ThetroubleisthatBayesianismcondonesallsuch
inferentialpracticesforallarelicensedbysuitablyperversepriors.
6.ThetraditionalepistemologistmayprotestthatBayesiansdistancethemselvesfrom
the world. Recall our discussion of the relativization of Bayesian claims to the
subjectiveprobabilityfunctionsofagents.Ratherthanhookingupdirectlywiththe
world,thetermsoftheirepistemologyareallinternaltoprobabilisticmodelsofthe
world.Moreover,theBayesianapparentlydoesnothavemuchofastoryaboutwhat
makesamodelgood,oronemodelbetterthananother.Thisisrelatedtotheconcern
that the Bayesian does not do justice to truth, justification, and the antiluck
conditions.
WhatarewetomakeoftheseconflictingconsiderationsforandagainstBayesian
epistemology?Atthisstageofthedialectic,anygoodHegelianwillinsistthatits
11/23/08 17
timefora:
6.Synthesis
Shouldwereallypreferoneapproachtoepistemologyovertheother?Shouldone
ofthetwoapproachesbejettisoned?Wewillarguethatweshouldnotregardthemas
incompetition.Infact,thetwoapproachescomplementeachotherinbothsubject
matterandmethod.
Traditionalepistemologistssometimesstressthatphilosophydiffersfromscience
andinsistthatphilosophyhasitsowndistinctmethodofenquiry,namelyconceptual
analysis.Bayesians,ontheotherhand,typicallyconsidertheirworkmoreinlinewith
scientifictheorizing.ThisisreflectedinthemanyconnectionsBayesianepistemology
haswithBayesianstatistics,decisiontheoryandtheliteratureoncausaldiscoveryin
artificialintelligence.ItisalsoreflectedintheimportanceBayesiansgivetosolving
successfulinthisrespect.Giventhesesuccesses,Bayesiansshouldhardlybeexpected
to give up their framework just because it is not a complete panacea for all
mechanicseveninthefaceofsometheoreticalproblems(forexample,itscommitment
toactionatadistance).AnddespiteitsbeingsupersededbyRelativityTheory,we
understandpreciselyhowNewtonianmechanicsisstillapproximatelytrue.Webelieve
thatthesuccessorofBayesianismwillstandinasimilarrelationtoBayesianismas
RelativityTheorystandstoNewtonianmechanics.
SoletusrevisitthechargesleveledatBayesianismintheprevioussection,keeping
inmindthisviewofitasaworkinprogressthatnonethelessisclearlyearningits
11/23/08 18
theoretical keeplike any good scientific theory. We may happily take its
fundamentalconcept,degreeofbeliefasaprimitiveintheabsenceofasuccessful
analysisofit.Ithasearneditstheoreticalkeepbyitscontributiontoavirtuoustotal
reductionofonetotheother.Indeed,theprospectsforsuchreductionstrikeusas
unpromising.(SeeErikssonandHjek2007.)
AgainstthechargeofBayesianismbeingempty,itcanpleadthegoodcompanyof
deductivelogic.Tobesure,crazysetsofbeliefcanbeconsistent,andinferences
fromabsurdpremisescanbevalidtheslogangarbagein,garbageoutisastruein
epistemologyasitisincomputerscience.Bayesianism,likelogic,cannevertheless
playasalutaryroleinkeepingourdegreesofbelief,likeourbeliefs,inharmony,and
inpolicingourelicitinferences.Afterall,deductivelogicisneverregardedasa
completesetofconstraintsonbelief;similarly,theBayesianconstraintsondegrees
ofbeliefshouldnotberegardedascomplete.Someadditionalconstraintsmaywell
findtheirinspirationintraditionalepistemology.
AgainstthechargeofBayesianismsverdictsbeingmodelrelative,itcanplead
thegoodcompanyofscience.Afterall,ourbestmethodsofenquiryinthephysical
andsocialsciencesworklikethis.Arguablyweshouldnotexpectepistemologytobe
different.
NorneedansweringskepticalchallengesbepartofBayesianismsjobdescription,
rationaldecisionmaking,confirmation,andtheuseofprobabilisticandstatistical
methodsinthesciences.Thereisnoharmintheirlabourbeingdivided.Theyaretwo
11/23/08 19
differentwaystoapproachepistemologyandtheyoftenanswerdifferentquestions.
Andwheretheirquestionsareshared,theirapproaches canbecomplementary
ratherthanmutuallyexclusive.WeseethatBayesianepistemologyhelpstoaddress
somequestionsthatwefindintraditionalepistemological debates.TheBayesian
treatmentofissuessuchastestimonyandthecoherencetheoryofjustification(see
section7)arecasesinpoint.ItturnsoutthattheformalmachineryofBayesianismis
wellsuitedtomakecertainquestionsmorepreciseandtoprovideanswerswhenour
intuitionsdontgiveclearverdicts.
NowletusseewhatonecandowiththemachineryofBayesianism.Itshouldbe
assessed by the problems it solves and how much it unifies, for example, the
methodologyofscience.
7.AchievementsofBayesianEpistemology
Many of the Bayesian success stories are from confirmation theory. But
Bayesianismhasmuchmoretoofferasitsdomainofapplicabilityalsoincludesother
partsofepistemologyandphilosophyofscience.Herearefivehighlights.
1.ConfirmationTheory.Aswesawinsection2,Bayesiansbeginwiththeideathat
intuitionsaboutconfirmationcanbevindicated.Supposethathentailse,sothatP(e|
h)=1.ThentheposteriorprobabilityofhisP(h|e)=P(h)/P(e).Hence,forafixed
priorprobabilityofh,theposteriorprobabilityofhincreasesifP(e)decreases.From
thiswecanimmediatelyaccountforthemethodologicalinsightthatmoresurprising
evidence confirms better. Similarly, Bayesians have provided a rationale for the
varietyofevidence thesisthe more varied the evidence is, the betterand have
11/23/08 20
providedilluminatingdiscussionsoftheDuhemQuinethesis(Earman1996,ch.3).
Toaddresstheseissues,severalmodelassumptionshavetobemade.Inthecaseof
thevarietyofevidencethesis,forexample,morevariedhastobeexplicatedin
probabilisticterms.Thiscanbedoneindifferentways.ManyBayesiandiscussionsof
thevarietyofevidencethesisassumethattheevidenceiscertain.Butwesawalready
thatthisisnotalwaysthecase.ItspeaksinfavoroftheBayesianframeworkthatit
provides the tools to model more complicated testing scenarios. See Bovens and
Hartmann(2003)formorerealisticBayesianmodelsofthevarietyofevidencethesis
andtheDuhemQuinethesis.
Bayesianconfirmationtheoryconnectsnaturallywithempiricalpsychologyasa
wealthofworkinthepsychologyofreasoningunderuncertaintydemonstrates.See,
for example, Chater and Oaksford (2008) and Oaksford and Chater (2007) for
sophisticatedBayesianmodelsthataccountforempiricalfindings.Crupietal.(2008)
show how the presence of the conjunction fallacy (i.e. that experimental subjects
assignahigherprobabilitytoaconjunctionthantooneoftheconjuncts),asfamously
demonstratedinpsychologicalexperimentsbyTverskyandKahneman(1983),canbe
explainedinconfirmationtheoreticalterms.Crupietal.(2007)argueonnormative
andexperimentalgroundsforaspecificmeasureofevidentialsupport(thesocalled
rationallyreconstructspecificepisodesfromthehistoryofscience.Whiletraditional
epistemology does not have the resources to study such episodes, the Bayesian
frameworkisideallysuitedforthesepurposes(Franklin1990).Itisalsobettersuited
thansomesystemofepistemicordoxasticlogicimaginetryingtoilluminatesome
scientificepisodesolelywithKandBoperators!
11/23/08 21
2.DynamicsofBelief.Traditionalepistemology,withitsfocusontheanalysisof
aboutbeliefchange,itisgenerallyassumedthatittakesplaceonthebasisoflearned
evidencethatiscertain.Traditionalepistemologysharesthisassumptionwithlogical
theories of belief revision such as the AGM theory (Grdenfors and Rott 1995).
However, Jeffrey taught us that learning often does not come in the form of
certainties.Toaddressthesecasesoflearningandbeliefchange,philosophersaswell
asresearchersinartificialintelligencehaveformulatednewupdatingrules(suchas
Bayesiannetworks(Neapolitan2003).
powerfulalgorithmshavebeendevelopedtolearncausalrelationsfromprobabilistic
data(KorbandNicholson2004,Pearl2000,Spirtesetal.2001).Thesealgorithmsuse
thetheoryofBayesiannetworks.ABayesiannetworkorganizesasetofvariablesinto
betweensomeofthenodes.Theonlyconstraintisthattherearenoclosedpaths
node.Eachnoderepresentsapropositionalvariable,whichcantakeanynumberof
mutuallyexclusiveandexhaustivevalues.TomakeaDAGintoaBayesiannetwork,
onemorestepisrequired:weneedtospecifythepriorprobabilitiesforthevariables
intherootnodesandtheconditionalprobabilitiesforthevariablesinallothernodes,
11/23/08 22
givenanycombinationofvaluesofthevariablesintheirrespectiveparentnodes.The
betweenthevariablesinthenetwork.ThisinformationisexpressedbytheParental
independentofallvariablesrepresentedbyitsnondescendentnodes,conditionalon
allvariablesrepresentedbyitsparentnodes. Inthecausalmodelingliterature,this
conditioniscalledtheCausalMarkovCondition.
4.TheCoherenceTheoryofJustification.ConfrontedwiththeCartesianskeptic,
coherentists pointoutthatwhenourbeliefsystemshangtogetherwell,withtheir
differentpartssupportingeachother,thenthisisanindicationofthetruthofthe
systems(BonJour1985).However,thecorrespondingtheorythecoherencetheory
of justificationsuffers from several problems. Here are two. First, the theory is
vague,asitisdifficulttomakeprecisewhatcoherenceis.Second,coherenceisnot
necessarilytruthconducive.Forexample,fairytalesaremadeup,althoughthestories
theytellmaybehighlycoherent.Hence,thecoherenceofasetofpropositionsisat
besttruthconduciveceterisparibus.Butwhatgoesintheceterisparibusclause?This
questionishardtoaddressifweonlyhavethetoolboxoftraditionalepistemology.
Bayesians can be of real help here. They have proposed and analyzed various
coherenceistruthconducive(BovensandHartmann2003,DouvenandMeijs2007,
Olsson2009).
knowledge and belief such as our senses, memory and testimony. All three have
inspired Bayesian modelbuilding. First, the uncertainty of the evidence from our
conditionalizationJeffreyconditionalization.Second,conditionalizationrepresents
anidealizedversionoftheepistemologicalroleofmemoryonewhoupdatesonly
byconditionalizationneverforgetswhileBayesianmodelsofboundedrationality
allowformemoryloss(Mehtaetal.2004).Third,Bayesianshavetheresourcesto
model the effect of combining the testimony of several witnesses (Bovens and
location,asexemplifiedbytheSleepingBeautyproblem(Elga2000),andwhether
propositionsconcerninghowtheworldis.
8.AvenuesforFutureResearch
Inthisclosingsection,webrieflypointtosometopicsthatwewouldliketosee
addressedinfutureresearch.WhilesomeofthemconcerntherelationofBayesian
internaltotheBayesianprogram.
Thereisnoharminlabourbeingdividedbetweenthetwokindsofepistemology,as
wehavearguedbutitwouldbeallthebetteriftheycouldbecomemorecooperative
enterprises.Thinkofsomeofthetimehonoreddebatesintraditionalepistemology:
skepticism,theanalysisofknowledge,reliabilism,internalismvs.externalism.Think
contrastivism,relativism,luminosity,knowledgehow(asopposedtoknowledge
that),knowledgewh_(who,where,when,which),Wherearethecounterpart
debatesinBayesianepistemology?Goingintheotherdirection,thinkofsomeofthe
timehonoreddebatesinBayesianepistemology:constraintsonpriors,updatingrules,
currentlyhottopics:credencesaboutchances(asinthePrincipalPrinciple),credences
aboutonesfuturecredences(asintheReflectionPrinciple),updatingcredenceson
traditionalepistemology?Eachofthesetopicssuggestsabridgewaitingtobebuilt.
Tosomeextent,suchprogressawaitsabetterunderstandingoftherelationship
betweentraditionalclaimsaboutbelief/knowledgeandBayesianclaimsaboutdegrees
ofbelief,whichisstillcontroversial.Aswehavesaid,wearenotsanguineaboutthe
prospectsofareductionineitherdirection,althoughreductionissurelynottheonly
waytoofferillumination.Andevenifthetwoepistemologiescontinueonseparate
tracks,stilldevelopmentsinonecanprovideinspirationorheuristicguidanceforthe
other.
Bayesianismstartedasaconfirmationtheory.Andindeed,theformalmachineryto
addressconfirmationtheoreticalquestionsishighlydeveloped.Ourambitiousgoal,
however,shouldbetodevelopafullfledgedBayesianphilosophyofscience.Hereis
anincompletelistofquestionsthatshouldbeaddressedtothisend:
(i)WhichstancedoesBayesianismtakeintherealismdebate?Isitneutraltothe
debate,ordoesitfavoraversionofscientificrealismorantirealism?(Douven2005,
11/23/08 25
Earman1996)
(ii)CanscientifictheorychangebeunderstoodinBayesianterms?(Earman1996)
epistemicallyadvantageoustoaimforunifiedtheories?(Myrvold2003)
(iv)CanaBayesianreadingofInferencetotheBestExplanationbegiven?Lipton
(2004)arguesthatexplanatoryconsiderationsareencodedinthelikelihoods,andnot
inthepriors.
(v)HowcanscientificidealizationsbeunderstoodinBayesianterms?Addressingthis
questionisimportantasidealizationsareubiquitousinscience.Thetroublebegins
whenweattachapriorprobabilityofzerotoanidealized(hencefalse)statement.The
posterioristhenarguablyalsozero,whicharguablyrendersBayesianismuseless.
3.BayesianSocialEpistemology.Bayesianepistemology,aswehavepresenteditso
far,sharesoneimportantfeaturewithtraditionalepistemology:itisindividualistic,
i.e.itisconcernedwithoneagent,whohasbeliefsandwhoupdatesherbeliefsinthe
light of new evidence. However, the doxastic unit could well be a community
comprising several individuals, or more. Kuhn (1962) argued that it is the entire
scientificcommunitythatacceptsorrejectsaparadigm.Orthinkofajurythathasto
comeupwithaconsensualverdictinamurdercase.Inlightofexamplessuchas
Kitcher1993).Whilemuchworkinthisfieldisinformal,formaltoolshaverecently
beendevelopedthataddressissuesinsocialepistemology.Especiallynoteworthyis
the work on judgment aggregation (List and Puppe 2009) that comprises
investigationsinspiredbythediscursivedilemma.BovensandRabinowicz(2006)
11/23/08 26
and Hartmann and Sprenger (2009) have given epistemic analyses of various
aggregationrulesstudiedinthiscontext.Itishopedthatthisworkwilleventually
developintoafullBayesianaccountofgroupjudgmentandgroupdecisionmaking.
Other topics of current interest include the debate about rational disagreement
(Feldman and Warfield 2009) and, related to this, theories about consensus and
compromiseformation.
Tosumup:Bayesianepistemologyisanexcitingandthrivingresearchprogram.
TheresplentymoreworkforBayesianstodo!1
References
Armendt,B.(1980).IsThereaDutchBookArgumentforProbabilityKinematics?
PhilosophyofScience47:583588.
BonJour,L.(1985).TheStructureofEmpiricalKnowledge.Cambridge,MA:Harvard
UniversityPress.
UniversityPress.
Bovens,L.andW.Rabinowicz(2006).DemocraticAnswerstoComplexQuestions:
AnEpistemicPerspective.Synthese150:131153.
BayesianCognitiveScience.Oxford:OxfordUniversityPress.
Crupi,V.,K.TentoriandM.Gonzalez(2007).OnBayesianMeasuresofEvidential
1
WethankJohnCusbert,KevinKorb,RalphMiles,JonahSchupbach,Declan
SmithiesandJanSprengerforveryhelpfulcomments.
11/23/08 27
229252.
Crupi, V., B. Fitelson and K. Tentori (2008). Probability, Confirmation and the
ConjunctionFallacy.ThinkingandReasoning14(2):182199.
Diaconis,P.andS.L.Zabell(1982).UpdatingSubjectiveProbability.Journalofthe
AmericanStatisticalAssociation77(380):822830.
Metascience14:338343.
Douven,I.andW.Meijs(2007).MeasuringCoherence.Synthese156:405425.
Earman,J.(1996).BayesorBust?ACriticalExaminationofBayesianConfirmation
Theory.Cambridge,Mass.:MITPress.
Edwards,W.,Lindman,H.,andSavage,J.(1963).BayesianStatisticalInferencefor
PsychologicalResearch.PsychologicalReview70:193242.
Eells,E.andB.Fitelson(2000).MeasuringConfirmationandEvidence. Journalof
Philosophy97:663672.
Elga, A. (2000). SelfLocating Belief and the Sleeping Beauty Problem. Analysis
60(2):14347.
SpecialIssueFormalEpistemologyI,B.Fitelson(ed.),183213.
Feldman,R.andT.Warfield(eds.)(2009).Disagreement.Oxford:OxfordUniversity
Press.
S36278.
Franklin,A.(1990).Experiment,RightorWrong.Cambridge:CambridgeUniversity
11/23/08 28
Press.
SymbolicLogic47:495548.
Grdenfors,P.andH.Rott(1995).BeliefRevision,inHandbookofLogicinArtificial
IntelligenceandLogicProgramming.VolumeIV:EpistemicandTemporal
Reasoning, eds. D.M. Gabbay, C.J. Hogger und J.A. Robinson. Oxford:
OxfordUniversityPress,35132.
Glymour,C.(1980).TheoryandEvidence.Princeton:PrincetonUniversityPress.
Goldman,A.(1999).KnowledgeinaSocialWorld.Oxford:ClarendonPress.
Good,I.J.(1971).LettertoTheAmericanStatistician25:6263.
Hjek,A.andN.Hall(2002).InductionandProbability.InTheBlackwellGuideto
thePhilosophyofScience, P.MachamerandM.Silberstein(eds.).Oxford:
Blackwell,149172.
TrackingtheTruth.Synthese.
Howson,C.andP.Urbach(2007).ScientificReasoning:TheBayesianApproach
.
Jaynes,E.(2003).ProbabilityTheory:TheLogicofScience
.Cambridge:Cambridge
UniversityPress.
Jeffrey,R.(1983). TheLogicofDecision.2nded.Chicago:UniversityofChicago
Press.
Cambridge:CambridgeUniversityPress.
Joyce,J.(1998).ANonpragmaticVindicationofProbabilism.PhilosophyofScience
65:575603.
11/23/08 29
Joyce,J.(1999).TheFoundationsofCausalDecisionTheory.Cambridge:Cambridge
UniversityPress.
Joyce,J.(2008).Bayes'Theorem. TheStanfordEncyclopediaofPhilosophy(Fall
2008Edition),E.Zalta(ed.).
Kitcher,P.(1993).TheAdvancementofScience.SciencewithoutLegend,Objectivity
withoutIllusions.NewYork:OxfordUniversityPress.
Kolmogorov,A.N.(1933).GrundbegriffederWahrscheinlichkeitsrechnung.Berlin:
Springer.Englishtranslation:FoundationsoftheTheoryofProbability.New
York:Chelsea1950.
Korb,K.andA.Nicholson(2004).BayesianArtificialIntelligence.BocaRaton:CRC
Press.
ChicagoPress.
Lewis,D.(1980).ASubjectivist'sGuidetoObjectiveChance.InR.Jeffrey(ed.),
StudiesinInductiveLogicandProbability,VolII.,Berkeley:Universityof
CaliforniaPress,263293.
Lewis,D.(1999).WhyConditionalize?InPapersinMetaphysicsandEpistemology.
Cambridge:CambridgeUniversityPress.
Lipton,P.(2004).InferencetotheBestExplanation.London:Routledge.
List,C.andC.Puppe(2009). JudgmentAggregation:ASurvey.In:PAnand,C.
Choice.Oxford:OxfordUniversityPress.
Mehta,N.,S.RajivandK.Srinivasan(2004).RoleofForgettinginMemoryBased
Economics2(2),107140.
Myrvold,W.(2003).ABayesianAccountoftheVirtueofUnification.Philosophyof
Science70:399423.
Neapolitan,R.(2003).LearningBayesianNetworks.PrenticeHall.
Nozick,R.(1981).PhilosophicalExplanations.Cambridge,MA:HarvardUniversity
Press.
ApproachtoHumanReasoning.Oxford:OxfordUniversityPress.
Olsson,E.(2009).AgainstCoherence:Truth,Probability,andJustification.Oxford:
OxfordUniversityPress.
Pearl,J.(2000).Causality.NewYork:CambridgeUniversityPress.
Ramsey,F.P.(1931).TruthandProbability.InR.B.Braithwaite(ed.),Foundations
ofMathematicsandotherEssays.London:Routledge,15698;reprintedinD.
H.Mellor(ed.),F.P.Ramsey:PhilosophicalPapers,Cambridge:Cambridge
UniversityPress1990.
Savage,L.J.(1954).TheFoundationsofStatistics.JohnWiley.
Fallacy.Synthese.
Sosa,E.(1999).HowtoDefeatOppositiontoMoore.PhilosophicalPerspectives13:
141153.
Spirtes,P.,C.Glymour,andR.Scheines(2001).Causation,PredictionandSearch.
Cambridge,Mass.:MITPress.
Tversky,A.andD.Kahneman(1983).ExtensionalversusIntuitiveReasoning:The
293315.
vanFraassen,B.(1984).BeliefandtheWill.JournalofPhilosophy81:235256.
ComputationalFoundations.Oxford:OxfordUniversityPress.
Zynda,L.(1995).OldEvidenceandNewTheories.PhilosophicalStudies77:6795.