View
2
Download
0
Category
Preview:
Citation preview
DecisionMakingin Robots and Autonomous Agents
BehaviouralIssues:
Heuris6cs&Biases,etc.
SubramanianRamamoorthySchoolofInforma6cs
16March,2018
TheRa'onalAnimalTheGreeks(Aristotle)thoughtthathumansarera'onalanimals.
– Manyancientphilosophersagreed– Clearly,evenAristotlewasawarethatpeople’s
judgments,decisions&behaviorarenotalwaysra'onal.– Amoreplausiblerevisedclaim:allnormalhumanshave
thecompetencetobera'onal.ArgumentthusfarinDMRhasbeeninthisspirit:
– “correct”principlesofreasoning&decisionmakingareinourminds,evenifwedon’talwaysusethem
– “Excep'ons”werethingslikeBernoulli’sobserva'onthatwebecomeprogressivelymoreindifferenttolargergains
16/03/2018 2
16/03/2018 3
Heuris'csandBiases
• TverskyandKahnemandrama'callyextendedBernoulli’saccount,andprovidedextensivepsychologicalevidencefortheirmodel.
• Whattheyshowedisthathumansdepartinmul'plewaysfromaclassicalpictureofra'onalbehavior.
• ManyhaveinterpretedtheKahneman&TverskyprogramasshowingthatAristotlewaswrong.– mostpeopledonothavethecorrectprinciplesforreasoning&decisionmaking
– wegetbywithmuchsimplerprincipleswhichonlysome/mesgettherightanswer
Amos Tversky1937 - 1996
Daniel Kahneman
AnIntui'onfromGigerenzer
Which US city has more inhabitants, San Diego or San Antonio?• 62%ofAmericansgotthiscorrect• 100%ofGermansgotthiscorrect
TheRecogni'onHeuris'c:Ifoneoftwoobjectsisrecognizedandtheotherisnot,theninferthattherecognizedobjecthasthehighervalue.
EcologicalRa'onality:Heuris'cworkswhenignoranceissystema'c(non–random),i.e.,whenlackofrecogni'oncorrelateswiththecriterion.
16/03/2018 4Goldstein & Gigerenzer, 2002, Psychological Review
ExperimentalStudiesofHumanReasoning
• TheConjunc'onFallacy• BaseRateNeglect• Overconfidence• Framing
16/03/2018 5
Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations. Please rank the following statements by their probability, using 1 for the most probable and 8 for the least probable.
(a) Linda is a teacher in elementary school. (b) Linda works in a bookstore and takes Yoga classes. (c) Linda is active in the feminist movement. (d) Linda is a psychiatric social worker. (e) Linda is a member of the League of Women Voters. (f) Linda is a bank teller. (g) Linda is an insurance sales person. (h) Linda is a bank teller & is active in the feminist movement.
Naïve subjects: (h) > (f) 89%
Sophisticated subjects: 85% 16/03/2018 6
Johnis19,wearsglasses,isalieleshyunlessyoutalktohimaboutStarTrekorLordoftheRings,andstaysuplatemostnightsplayingvideogames.Whichismorelikely?1. JohnisaCSmajorwholovescontemporarysculpture
andgolf,or2. Johnlovescontemporarysculptureandgolf?
Onproblemslikethis,peopletendtopick(1)above(2),incontradic'ontoprobabilitytheory.
16/03/2018 7
ExperimentalStudiesofHumanReasoning
• TheConjunc'onFallacy• BaseRateNeglect• Overconfidence• Framing
16/03/2018 8
A panel of psychologists have interviewed and administered personality tests to 30 engineers and 70 lawyers [70 engineers and 30 lawyers], all successful in their respective fields. On the basis of this information, thumbnail descriptions of the 30 engineers and 70 lawyers [70 engineers and 30 lawyers], have been written. You will find on your forms five descriptions, chosen at random from the 100 available descriptions. For each description, please indicate your probability that the person described is an engineer, on a scale from 0 to 100.
16/03/2018 9
Jack is a 45-year-old man. He is married and has four children. He is generally conservative, careful and ambitious. He shows no interest in political and social issues and spends most of his free time on his many hobbies which include home carpentry, sailing, and mathematical puzzles.
Dick is a 30-year-old man. He is married with no children. A man of high ability and high motivation, he promises to be quite successful in his field. He is well liked by his colleagues.
16/03/2018 10
BaseRateandRepresenta'veness
Withnopersonalitysketch,simplyaskedfortheprobabilitythatanunknownindividualwasanengineer– subjectscorrectlygavetheresponses0.7and0.3
Whenpresentedwithatotallyuninforma'vedescrip'on
– thesubjectsgavetheprobabilitytobe0.5
Kahneman&Tverskyconcludedthatwhennospecificevidenceisgiven,priorprobabili'esareusedproperly;whenworthlessevidenceisgiven,priorprobabili'esareignored.
16/03/2018 11
If a test to detect a disease whose prevalence is 1/1000 has a false positive rate of 5%, what is the chance that a person found to have a positive result actually has the disease, assuming that you know nothing about the person’s symptoms or signs?
____%
Harvard Medical School: “2%” = 18%; “95%” = 45%
16/03/2018 12
16/03/2018 13
Whyisthecorrectanswer2%?
• Thinkofapopula'onof10,000people.• Wewouldexpectjust10peopleinthispopula'ontohavethe
disease(1/1000x10,000=10)• Ifyoutesteverybodyinthepopula'onthenfalseposi'verate
meansthat,inaddi'ontothe10peoplewhodohavethedisease,another500(5%)willbewronglydiagnosedashavingit.
• Inotherwordsonlyabout2%ofthepeoplediagnosedposi've(10/510)actuallyhavethedisease.
• Whenpeoplegiveahighanswerlike95%theyareignoringtheverylowprobability(i.e.rarity)ofhavingthedisease.Incomparison,probabilityofafalseposi'vetestisrela'velyhigh.
ExperimentalStudiesofHumanReasoning
• TheConjunc'onFallacy• BaseRateNeglect• Overconfidence• Framing
16/03/2018 14
In each of the following pairs, which city has more inhabitants?
§ (a) Las Vegas (b) Miami How confident are you that your answer is correct? 50% 60% 70% 80% 90% 100%
§ (a) Sydney (b) Melbourne How confident are you that your answer is correct? 50% 60% 70% 80% 90% 100%
§ (a) Hyderabad (b) Islamabad How confident are you that your answer is correct? 50% 60% 70% 80% 90% 100%
16/03/2018 15
Overconfidenceeffect
• Person'ssubjec'veconfidenceinhisorherjudgementsisreliablygreaterthantheobjec'veaccuracyofthosejudgements– especiallywhenconfidenceishigh
• Manifestedindifferentways:– overes/ma/onofone'sactualperformance– overplacementofone'sperformancerela'vetoothers– overprecisioninexpressingunwarrantedcertaintyintheaccuracyofone'sbeliefs.
16/03/2018 16
ExperimentalStudiesofHumanReasoning
• TheConjunc'onFallacy• BaseRateNeglect• Overconfidence• Framing
16/03/2018 17
Imagine the U.S. is preparing for the outbreak of anunusual Asian disease which is expected to kill 600people. Two alterna've programs to combat thedisease have been proposed. Assume that the exactscien'fices'matesoftheconsequencesoftheprogramareasfollows:
16/03/2018 18
– IfprogramAisadopted,200peoplewillbesaved.
– IfprogramBisadopted,thereisa1/3probabilitythat600peoplewillbesavedanda2/3probabilitythatnopeoplewillbesaved.
16/03/2018 19
– IfprogramAisadopted,200peoplewillbesaved.
– IfprogramBisadopted,thereisa1/3probabilitythat600peoplewillbesavedanda2/3probabilitythatnopeoplewillbesaved.
16/03/2018 20
– IfprogramCisadopted,400peoplewilldie.
– IfprogramDisadopted,thereisa1/3probabilitythatnobodywilldieanda2/3probabilitythat600peoplewilldie.
16/03/2018 21
– IfprogramCisadopted,400peoplewilldie.
– IfprogramDisadopted,thereisa1/3probabilitythatnobodywilldieanda2/3probabilitythat600peoplewilldie.
16/03/2018 22
– IfprogramAisadopted,200peoplewillbesaved.
– IfprogramBisadopted,thereisa1/3probabilitythat600peoplewillbesavedanda2/3probabilitythatnopeoplewillbesaved.
– IfprogramCisadopted,400peoplewilldie.
– IfprogramDisadopted,thereisa1/3probabilitythatnobodywilldieanda2/3probabilitythat600peoplewilldie.
16/03/2018 23
ProgramAversusB
• TverskyandKahneman(1981)foundthatthemajoritychoice(72%)forananalogueofproblem1usingananonymous“Asiandisease”(notSARS)wasanswerA.
• Theprospectofsaving200liveswithcertaintywasmorepromisingthantheprobabilityofaone-in-threechanceofsaving600lives.
16/03/2018 24
AandBhavetheSameU'lity
• Thestandardwayofcalcula'ngexpectedu'lity:• ΣiP(oi)xU(oi)• whereeachoiisapossibleoutcome,P(oi)isits
probabilityandU(oi)isitsu'lityorvalue.
• Onthisbasis,op'onB(ariskyprospect)wouldbeofequalexpectedvaluetothefirstprospectA.
• Sointhiscasesubjectsarenotsimplyusingexpectedu'lity:theyareriskaverse.
16/03/2018 25
ProgramCversusD
• Themajorityofrespondentsinthesecondproblem(78%)chosetheriskierop'on.
• Thecertaindeathof400peopleisapparentlylessacceptablethanthetwo-in-threechancethat600peoplewilldie.
• Onceagain,thereisnodifferencebetweentheop'ons(CandD)instandardexpectedu'lity.
• Wesayinthiscasethatpeopleareriskseeking.
16/03/2018 26
Framing
• TverskyandKahneman’sinten'onwasthatProblem1andProblem2areunderlyinglyiden'cal,butpresentedindifferentways.
• Totheextentthattheyareright,wesaythattheyaredifferentframingsofthesameproblem:thedifferentresponsesforthetwoproblemsarewhatwecallframingeffects.
• Yetanotherwayinwhichpeople’sdecisionsdifferfromwhatwouldbepredictedonthestandardviewofexpectedu'lity.
16/03/2018 27
16/03/2018 28
ProspectTheory:S-curveofValue
Actual gain
Perceivedgain
Actual loss
PerceivedLoss
• Small gain good,• Big gain not much
better,• Small loss terrible,• Big loss not much
worse.
Thus we are generally:• Loss averse,• Risk averse for gains,
but• Risk seeking for losses.
Reference point
16/03/2018 29
Es'ma'onofRisk
Estimates of Probabilities of Death From Various Causes:
Cause Stanford graduate estimates Statistical estimatesHeart Disease 0.22 0.34Cancer 0.18 0.23Other Natural Causes 0.33 0.35All Natural Causes 0.73 0.92
Accident 0.32 0.05Homicide 0.10 0.01Other Unnatural 0.11 0.02All Unnatural 0.53 0.08
(BasedonastudybyAmosTversky,ThayerWatkin’sreport)
BigandSmallProbabili'es
• Smallprobabilityeventsareoverrated.
• Moregenerally,behaviordepartsfromclassicalra'onalitysubstan'allyasregardsverysmallandverylargeprobabili'es.
• Evenwhentoldthefrequencyofevents,subjectstendtoviewthedifferencebetween0and0.001(orthatbetween0.999and1)asmoresignificantthanthatbetween.5and.5001.
16/03/2018 30
The“BleakImplica'ons”Hypothesis
• These results have“bleak implica'ons” for the ra'onalityofordinarypeople(Nisbeeetal.)
• “Itappearsthatpeoplelackthecorrectprogramsformanyimportant judgmental tasks…. We have not had theopportunity to evolve an intellect capable of dealingconceptually with uncertainty.” (Slovic, Fischhoff andLichtenstein,1976)
16/03/2018 31
The“BleakImplica'ons”Hypothesis
“I am par'cularly fond of [the Linda] example, because Iknow that the [conjunc'on] is least probable, yet a lielehomunculus ismy head con'nues to jump up and down,shou'ngatme–“butshecan’t justbeabankteller;readthe descrip'on.” … Why do we consistently make thissimple logical error? Tversky and Kahneman argue,correctlyIthink,thatourmindsarenotbuilt(forwhateverreason) towork by the rules of probability.” (Stephen J.Gould,1992,p.469)
16/03/2018 32
ChallengefromEvolu'onaryPsychology
• The“frequen'st”hypothesis– Usefulinforma'onaboutprobabili'eswasavailableto
ourancestorsintheformoffrequencies• e.g.,3ofthelast12huntsnearriversweresuccessful
– Informa'onabouttheprobabili'esofsingleeventswasnotavailable
– Soperhapsweevolvedamentalcapacitythatisgoodatdealingwithprobabilis'cinforma'on,butonlywhenthatinforma'onispresentedinafrequencyformat.
16/03/2018 33
MakingtheConjunc'onFallacy“Disappear”…
16/03/2018 34
Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations. Please rank the following statements by their probability, using 1 for the most probable and 8 for the least probable.
(a) Linda is a teacher in elementary school. (b) Linda works in a bookstore and takes Yoga classes. (c) Linda is active in the feminist movement. (d) Linda is a psychiatric social worker. (e) Linda is a member of the League of Women Voters. (f) Linda is a bank teller. (g) Linda is an insurance sales person. (h) Linda is a bank teller and is active in the feminist movement.
16/03/2018 35
Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations.
There are 100 people who fit the description above. How many of them are:
…
(f) bank tellers?
(h) bank tellers and active in the feminist movement?
...
(h) > (f) = 13%
16/03/2018 36
Level-kReasoning
Thecen'pedegame:
• Intheonlyequilibriuminvolvingra'onalplayers,player1choosesdowninthefirst'memove
• Whatexactlydorealpeopledo?– Humansubjectsdon’tchoosethedownop'onrightaway– Howdoweunderstandtheirbehaviour?
16/03/2018 37
HideandSeekGames
• Youropponenthashiddenaprizeinoneoffourboxesarrangedinarow.
• Theboxesaremarkedasshownbelow:A,B,A,A• Yourgoalis,ofcourse,tofindtheprize.• Hisgoalisthatyouwillnotfindit.• Youareallowedtoopenonlyonebox.• Whichboxareyougoingtoopen?
16/03/2018 38
FolkWisdom
• “…inLakeWobegon,thecorrectanswerisusually‘c’.” –GarrisonKeillor(1997)onmul'ple-choicetests
• CommentonpoisoningofUkrainian’spresiden'alcandidate:“Anygovernmentwan'ngtokillanopponent…wouldnottryitatamee'ngwithgovernmentofficials.” –ViktorYushchenko(2004)
16/03/2018 39
UniqueEquilibriumofHideandSeekGame
• Ifbothplayersrandomizeuniformly:
Expectedpayoffs:Hider3/4,Seeker1/4
16/03/2018 40
BehaviouralExperimentinHideandSeek
• CentralA(or3)ismostprevalentforbothHidersandSeekers• CentralAisevenmoreprevalentforSeekers
– Asaresult,Seekersdobeeerthaninequilibrium• Shouldn’tHidersrealizethatSeekerswillbejustastemptedto
lookthere?• “Thefindingthatbothchoosersandguessersselectedtheleast
salientalterna'vesuggestslieleornostrategicthinking.”
16/03/2018 41
p-BeautyContestGames(Keynes,1936)
“…professionalinvestmentmaybelikenedtothosenewspapercompe''onsinwhichthecompe'torshavetopickoutthesixprevestfacesfromahundredphotographs,
theprizebeingawardedtothecompe'torwhosechoicemostnearlycorrespondstotheaveragepreferencesofthecompe'torsasawhole….
Itisnotacaseofchoosingthose[faces]that,tothebestofone’sjudgment,arereallytheprevest,noreventhosethataverageopiniongenuinelythinkstheprevest.
16/03/2018 42
p-BeautyContestGames
• Wehavereachedthethirddegreewherewedevoteourintelligencesto…an'cipa'ngwhataverageopinionexpectstheaverageopiniontobe.
• Andtherearesome,Ibelieve,whoprac'cethefourth,fiHhandhigherdegrees.”
–Keynes,GeneralTheory,1936,pp.155-56
16/03/2018 43
Level-kReasoning
Onecouldplayatvaryinglevelsofdepthofreasoning:• Level-0:Randomplay• Level-1:BRtoRandomplay• Level-2:BRtoLevel-1• Nash:PlayNashEquilibrium(level-∞?)• Worldly:BRtodistribu'onofLevel-0,Level-1andNashtypes
StahlandWilson(GEB1995)
16/03/2018 44
Level-kforHideandSeek
• Level-k:EachroleisfilledbyLktypes:L0,L1,L2,L3,orL4(probabili'estobees'mated…) –Note:InHideandSeekthetypescycleaxerL4…
• HightypesanchorbeliefsinanaïveL0typeandadjustswith
iteratedbestresponses: –L1bestrespondstoL0(withuniformerrors) –L2bestrespondstoL1(withuniformerrors) –LkbestrespondstoLk-1(withuniformerrors)
16/03/2018 45
Level-kforHideandSeek
• L0HidersandSeekersaresymmetric–Favorsalientloca'onsequally1.Favor“B”:choosewithprobabilityq>1/42.Favor“endA”:choosewithprob.p/2>1/4 –Choiceprobabili'es:(p/2,q,1-p-q,p/2)
• Note:Specifica'onofAnchoringTypeL0isthekeytomodel’sexplanatorypower –Can’tuseuniformL0(coincidewithequilibrium)…
16/03/2018 46
ModernPerspec've:DiversityinModesofThinking
• Automa'cSystem(Type1)– Fast,unconscious– Parallel,associa've– Lowenergy– “Doer”
• Reflec'veSystem(Type2)– Slow,conscious– Serial– Expensive– “Planner”
16/03/2018 47
TrainingtoGoBetweenModesExample:BoardMemoryinChess
heps://www.youtube.com/watch?v=rWuJqCw|jc
16/03/2018 48
U'lisingTypeIThinking
• frombehavioralscience• arguesthatposi've
reinforcementandindirectsugges'onstotrytoachievenon-forcedcompliancecaninfluencemo'ves,incen'vesanddecisionmaking
• atleastaseffec'vely–ifnotmoreeffec'vely-thandirectinstruc'on,legisla'on,orenforcement
16/03/2018 49
Applica'onExample:PersuasiveTechnology(Beeminder)
16/03/2018 50
Applica'onExample:ChangingDrivingBehaviour-LoeeryIncen've
• Enterpeopleinaloeeryiftheycometocampusoff-peak• Observethatthissmallprobabilityofearningsactuallycauses
non-essen'altravellerstoaltertheirbehaviour
16/03/2018 51
Summary
• Muchofthecourseprogressedundertheassump'onofafairlystrictformofra'onality– Nobodyactuallybelievesthatthisiswhatpeoplealwaysdo– Instead,animplicitclaimisthatpeoplehavethecapacityforthiskind
ofreasoning
• Inthislecture,welookedathowpeoplesystema'callydeviatefromsuchano'onofra'onality
• Thishasmajorimplica'onsforcomputa'onalagents– Intheend,youwantthemtoworkwithrealpeople– Modelling“realpeople”ischallenging,butul'matelytheraisond’etre
16/03/2018 52
Acknowledgements
ManyoftheseslidesareadaptedfromJurafskyetal.’sSymbolicSystemscourseatStanford
16/03/2018 53
Recommended