Ar\ficial Intelligence

Ar#ficialIntelligence
Dr.QaiserAbbas
DepartmentofComputerScience&IT,
UniversityofSargodha,Sargodha,40100,Pakistan
[email protected]
Saturday,23April16
1
1.WhatisAI?
thoughtprocesses
andreasoning
ThinkingHumanly
“TheexciOngnewefforttomake
computersthink...machineswith
minds,inthefullandliteral
sense.”(Haugeland,1985)
“[TheautomaOonof]acOviOesthatwe
associatewithhumanthinking,
acOviOessuchasdecision-making,
problemsolving,
learning...”(Bellman,1978)
ThinkingRa#onally
“ThestudyofmentalfaculOes
(inherentmentalorphysicalpower)
throughtheuseofcomputaOonal
models.”
(CharniakandMcDermo_,1985)
“ThestudyofthecomputaOonsthat
makeitpossibletoperceive,reason,
andact.”(Winston,1992)
behavior
Ac#ngHumanly
“TheartofcreaOngmachinesthatper-
formfuncOonsthatrequire
intelligencewhenperformedby
people.”(Kurzweil,1990)
“Thestudyofhowtomakecomputers
dothingsatwhich,atthemoment,
peoplearebe_er.”(RichandKnight,
1991)
Ac#ngRa#onally
“ComputaOonalIntelligenceisthe
studyofthedesignofintelligent
agents.”(Pooleetal.,1998)
“AI...isconcernedwithintelligent
behaviorinarOfacts.”(Nilsson,1998)
Lee:successintermsof
fidelitytohumanperformance
Right:anidealperformance
measure,calledra#onality
Saturday,23April16
2
1.1.Ac#nghumanly:TheTuringTest
approach
•  TheTuringTest,proposedbyAlanTuring(1950),wasdesignedto
provideasaOsfactoryoperaOonaldefiniOonofintelligence.
•  Acomputerpassesthetestifahumaninterrogator,aeerposing
somewri_enquesOons,cannottellwhetherthewri_enresponses
comefromapersonorfromacomputer.(SeeCh26fordetails)
•  Programmingacomputertopassarigorously(extremelythorough)
appliedtestprovidesplentytoworkon.Thecomputerwouldneed
topossessthefollowingcapabiliOes:
–  naturallanguageprocessingtoenableittocommunicatesuccessfully
inEnglishorothernaturallanguages;
–  knowledgerepresenta#ontostorewhatitknowsorhears;
–  automatedreasoningtousethestoredinformaOontoanswer
quesOonsandtodrawnewconclusions;
–  machinelearningtoadapttonewcircumstancesandtodetectand
extrapolate(extendexisOngmethods)pa_erns.
Saturday,23April16
3
1.1.Ac#nghumanly:TheTuringTest
approach
•  TopassthetotalTuringTest,thecomputer
willneed
–  Computervisiontoperceiveobjects,and
–  Robo#cstomanipulateobjectsandmoveabout.
•  ThesesixdisciplinescomposemostofAI,and
Turingdeservescreditfordesigningatestthat
remainsrelevant60yearslater.
Saturday,23April16
4
1.2.Thinkinghumanly:Thecogni#ve
modelingapproach
•  Ifwearegoingtosaythatagivenprogramthinkslikeahuman,we
musthavesomewayofdetermininghowhumansthink.Weneed
togetinsidetheactualworkingsofhumanminds.
•  Therearethreewaystodothis:
–  throughintrospecOon—tryingtocatchourownthoughtsastheygo
by;
–  throughpsychologicalexperiments—observingapersoninacOon;
–  andthroughbrainimaging—observingthebraininacOon.
•  Oncewehaveasufficientlyprecisetheoryofthemind,itbecomes
possibletoexpressthetheoryasacomputerprogram.Ifthe
program’sinput–outputbehaviormatchescorrespondinghuman
behavior,thatisevidencethatsomeoftheprogram’smechanisms
couldalsobeoperaOnginhumans.
Saturday,23April16
5
1.2.Thinkinghumanly:Thecogni#ve
modelingapproach
•  Forexample,AllenNewellandHerbertSimon,who
developedGPS,the“GeneralProblemSolver”(Newelland
Simon,1961),werenotcontent(willingtoaccept
something)merelytohavetheirprogramsolveproblems
correctly.Theyweremoreconcernedwithcomparingthe
trace(follow)ofitsreasoningstepstotracesofhuman
subjectssolvingthesameproblems.
•  Theinterdisciplinaryfieldofcogni#vesciencebrings
togethercomputermodelsfromAIandexperimental
techniquesfrompsychologytoconstructpreciseand
testabletheoriesofthehumanmind.
•  RealcogniOvescience,however,isnecessarilybasedon
experimentalinvesOgaOonofactualhumansoranimals.
Saturday,23April16
6
1.3.Thinkingra#onally:The“lawsof
thought”approach
•  TheGreekphilosopherAristotlewasoneofthefirsttoa_empttocodify
“rightthinking,”thatis,irrefutable(impossibletodenyordisprove)
reasoningprocesses.
•  Hissyllogismsprovidedpa_ernsforargumentstructuresthatalways
yieldedcorrectconclusionswhengivencorrectpremises:
–  forexample,“Socratesisaman;allmenaremortal;therefore,Socratesis
mortal.”TheselawsofthoughtweresupposedtogoverntheoperaOonofthe
mind;theirstudyiniOatedthefieldcalledlogic.
•  Logiciansinthe19thcenturydevelopedaprecisenotaOonforstatements
aboutallkindsofobjectsintheworldandtherelaOonsamongthem.
(ContrastthiswithordinaryarithmeOcnotaOon,whichprovidesonlyfor
statementsaboutnumbers.)
•  By1965,programsexistedthatcould,inprinciple,solveanysolvable
problemdescribedinlogicalnotaOon.(AlthoughifnosoluOonexists,the
programmightloopforever.)
•  Theso-calledlogicisttradiOonwithinarOficialintelligencehopestobuild
onsuchprogramstocreateintelligentsystems.
Saturday,23April16
7
1.3.Thinkingra#onally:The“lawsof
thought”approach
•  Therearetwomainobstaclestothisapproach.
•  First,itisnoteasytotakeinformalknowledge
andstateitintheformaltermsrequiredby
logicalnotaOon,parOcularlywhentheknowledge
islessthan100%certain.
•  Second,thereisabigdifferencebetweensolving
aproblem“inprinciple”andsolvingitinpracOce.
Evenproblemswithjustafewhundredfactscan
exhaustthecomputaOonalresourcesofany
computerunlessithassomeguidanceasto
whichreasoningstepstotryfirst.
Saturday,23April16
8
1.4.Ac#ngra#onally:Thera#onal
agentapproach
• 
• 
• 
• 
• 
• 
Anagentisjustsomethingthatacts(agentcomesfromtheLaOnagere,todo).
Ofcourse,allcomputerprogramsdosomething,butcomputeragentsare
expectedtodomore:operateautonomously,perceivetheirenvironment,persist
overaprolongedOmeperiod,adapttochange,andcreateandpursuegoals.
Ara#onalagentisonethatactssoastoachievethebestoutcomeor,whenthere
isuncertainty,thebestexpectedoutcome.
Inthe“lawsofthought”approachtoAI,theemphasiswasoncorrectinferences.
MakingcorrectinferencesissomeOmespartofbeingaraOonalagent,because
onewaytoactraOonallyistoreasonlogicallytotheconclusion.
Ontheotherhand,correctinferenceisnotallofraOonality;insomesituaOons,
thereisnoprovablycorrectthingtodo,butsomethingmustsOllbedone.
TherearealsowaysofacOngraOonallythatcannotbesaidtoinvolveinference.
Forexample,recoiling(springbackinfear)fromahotstoveisareflexacOonthat
isusuallymoresuccessfulthanasloweracOontakenaeercarefuldeliberaOon
(longandcarefulconsideraOonordiscussion).
Saturday,23April16
9
1.4.Ac#ngra#onally:Thera#onal
agentapproach
•  KnowledgerepresentaOonandreasoningenableagentstoreachgood
decisions.
•  Weneedtobeabletogeneratecomprehensiblesentencesinnatural
languagetogetbyinacomplexsociety.
•  WeneedlearningnotonlyforerudiOon(thequalityofhavingorshowing
greatknowledgeorlearning;),butalsobecauseitimprovesourabilityto
generateeffecOvebehavior.
•  TheraOonal-agentapproachhastwoadvantagesovertheother
approaches.
–  First,itismoregeneralthanthe“lawsofthought”approachbecausecorrect
inferenceisjustoneofseveralpossiblemechanismsforachievingraOonality.
–  Second,itismoreamenable(capableofbeingacteduponinaparOcularway)
toscienOficdevelopmentthanareapproachesbasedonhumanbehavioror
humanthought.
•  ThestandardofraOonalityismathemaOcallywelldefinedandcompletely
general,andcanbe“unpacked”togenerateagentdesignsthatprovably
achieveit.
Saturday,23April16
10
2.THEFOUNDATIONSOFARTIFICIAL
INTELLIGENCE
•  Wewillstudyabriefhistoryofthedisciplines
thatcontributedideas,viewpoints,and
techniquestoAI.
Saturday,23April16
11
2.1.Philosophy
•  Canformalrulesbeusedtodrawvalidconclusions?
•  Aristotle(384–322B.C.),wasthefirsttoformulateaprecisesetoflaws
governingtheraOonalpartofthemind.Hedevelopedaninformalsystem
ofsyllogismsforproperreasoning,whichinprincipleallowedoneto
generateconclusionsmechanically,giveniniOalpremises.
•  Lateron,RamonLull(1315),ThomasHobbes(1588–1679),alsopresents
reasoningwithadirecOontonumericalcomputa#on.Thenaround1500,
LeonardodaVinci(1452–1519)designedbutdidnotbuildamechanical
calculator;recentreconstrucOonshaveshownthedesigntobefuncOonal.
•  ThefirstknowncalculaOngmachinewasconstructedaround1623bythe
GermanscienOstWilhelmSchickard(1592–1635),althoughthePascaline,
builtin1642byBlaisePascal(1623–1662),ismorefamous.
Saturday,23April16
12
2.1.Philosophy
• 
Go^riedWilhelmLeibniz(1646–1716)builtamechanicaldeviceintendedtocarryoutoperaOons
onconceptsratherthannumbers,butitsscopewasratherlimited.LeibnizdidsurpassPascalby
buildingacalculatorthatcouldadd,subtract,mulOply,andtakeroots,whereasthePascalinecould
onlyaddandsubtract.
Somespeculated(formatheoryorconjecture)thatmachinesmightnotjustdocalcula<onsbut
actuallybeabletothinkandactontheirown.
Thisleadtoanotherdiscussionthatmachinesaresimplyama?eroramind?
• 
Howdoesthemindarisefromaphysicalbrain?
• 
It’sonethingtosaythatthemindoperates,atleastinpart,accordingtologicalrules,andtobuild
physicalsystemsthatemulatesomeofthoserules;it’sanothertosaythattheminditselfissucha
physicalsystem.
RenéDescartes(1596–1650),wasastrongadvocateofthepowerofreasoninginunderstanding
theworld,aphilosophynowcalledra#onalism,andonethatcountsAristotleandLeibnitzas
members.
Descarteswasalsoaproponent(advocate)ofdualism.Heheldthatthereisapartofthehuman
mind(orsoulorspirit)thatisoutsideofnature,exemptfromphysicallaws.Animals,ontheother
hand,didnotpossessthisdualquality;theycouldbetreatedasmachines.
• 
• 
• 
• 
Saturday,23April16
13
2.1.Philosophy
• 
AnalternaOvetodualismismaterialism,whichholdsthatthebrain’soperaOon
accordingtothelawsofphysicscons<tutesthemind.
• 
Givenaphysicalmindthatmanipulatesknowledge,thenextproblemisto
establishthesourceofknowledge.
• 
Wheredoesknowledgecomefrom?
• 
Theempiricismmovement,starOngwithFrancisBacon’s(1561–1626)is
characterizedbyadictum(formalannouncement)ofJohnLocke(1632–1704):
“Nothingisintheunderstanding,whichwasnotfirstinthesenses.”
DavidHume’s(1711–1776)proposedinduc#onthatgeneralrulesareacquiredby
exposuretorepeatedassociaOonsbetweentheirelements.
LudwigWi_genstein(1889–1951)BertrandRussell(1872–1970),RudolfCarnap
(1891–1970),developedthedoctrine(abelieforsetofbeliefs)oflogical
posi#vism.Thisdoctrineholdsthatallknowledgecanbecharacterizedbylogical
theoriesconnectedtoobserva#onsentencesthatcorrespondtosensoryinputs.
• 
• 
Saturday,23April16
14
2.1.Philosophy
• 
• 
Theconfirma#ontheoryofCarnapandCarlHempel(1905–1997)a_emptedto
analyzetheacquisiOonofknowledgefromexperience.Itdefinedanexplicit
computaOonalprocedureforextracOngknowledgefromelementaryexperiences
andwasprobablythefirsttheoryofmindasacomputaOonalprocess.
ThefinalelementinthephilosophicalpictureofthemindistheconnecOon
betweenknowledgeandacOon.
• 
HowdoesknowledgeleadtoacOon?
• 
ThisquesOonisvital(necessary)toAIbecauseintelligencerequiresacOonaswell
asreasoning.Aristotleargued(inDeMotuAnimalium)thatacOonsarejusOfiedby
alogicalconnecOonbetweengoalsandknowledgeoftheacOon’soutcome
Aristotle’salgorithmwasimplemented2300yearslaterbyNewellandSimonin
theirGPSprogram.Wewouldnowcallitaregressionplanningsystem(see
Chapter10).
• 
Saturday,23April16
15
2.2Mathema#cs
•  Whataretheformalrulestodrawvalidconclusions?(Seenin2.1)
•  PhilosophersstakedoutsomeofthefundamentalideasofAI,but
theleaptoaformalsciencerequiredalevelofmathemaOcal
formalizaOoninthreefundamentalareas:logic,computaOon,and
probability.
•  Theideaofformallogiccanbetracedbacktothephilosophersof
ancientGreece,butitsmathemaOcaldevelopmentreallybegan
withtheworkofGeorgeBoole(1815–1864),whoworkedoutthe
detailsofproposiOonalorBooleanlogic.
•  In1879,Go_lobFrege(1848–1925)extendedBoole’slogicto
includeobjectsandrelaOons,creaOngthefirst-orderlogicthatis
usedtoday.
•  AlfredTarski(1902–1983)introducedatheoryofreferencethat
showshowtorelatetheobjectsinalogictoobjectsinthereal
world.
Saturday,23April16
16
2.2Mathema#cs
• 
• 
• 
• 
• 
• 
• 
• 
• 
Whatcanbecomputed?
ThenextstepwastodeterminethelimitsofwhatcouldbedonewithlogicandcomputaOon.
Thefirstnontrivial(significant)algorithmisthoughttobeEuclid’salgorithmforcompuOnggreatest
commondivisors.
Thewordalgorithm(andtheideaofstudyingthem)comesfromal-Khowarazmi,aPersian
mathemaOcianofthe9thcentury,whosewriOngsalsointroducedArabicnumeralsandalgebrato
Europe.
BooleandothersdiscussedalgorithmsforlogicaldeducOon.
In1930,KurtGödel(1906–1978)showedthatfirst-orderlogiccouldnotcapturetheprincipleof
mathemaOcalinducOonneededtocharacterizethenaturalnumbers.
In1931,GödelshowedthatlimitsondeducOondoexistandhepresentedtheincompleteness
theorem,accordingtowhichtherearetruestatementsthatareundecidableinthesensethatthey
havenoproofwithinthetheory.
AlanTuring(1912–1954)triedtocharacterizeexactlywhichfuncOonsarecomputableandwhich
arenot.
TheChurch–TuringthesisstatesthattheTuringmachine(Turing,1936)iscapableofcompuOngany
computablefuncOon,andtherearesomefuncOonsthatnoTuringmachinecancompute.
Saturday,23April16
17
2.2Mathema#cs
•  AproblemiscalledintractableiftheOmerequiredtosolveinstancesof
theproblemgrowsexponenOallywiththesizeoftheinstances.
•  ThedisOncOonbetweenpolynomial(nk)andexponenOal(2n)growthin
complexitywasfirstemphasizedinthemid-1960s(Cobham,1964;
Edmonds,1965).
•  Therefore,oneshouldstrivetodividetheoverallproblemofgeneraOng
intelligentbehaviorintotractablesubproblemsratherthanintractable
ones.
•  Howcanonerecognizeanintractableproblem?ThetheoryofNPcompleteness,pioneeredbyStevenCook(1971)andRichardKarp(1972),
providesamethod.
•  AlthoughithasnotbeenprovedthatNP-completeproblemsare
necessarilyintractable,mosttheoreOciansbelieveit.
•  WorkinAIhashelpedexplainwhysomeinstancesofNP-complete
problemsarehard,yetothersareeasy(Cheesemanetal.,1991).
Saturday,23April16
18
2.2Mathema#cs
•  HowdowereasonwithuncertaininformaOon?
•  BesideslogicandcomputaOon,thethirdgreatcontribuOonof
mathemaOcstoAIisthetheoryofprobability.TheItalianGerolamo
Cardano(1501–1576)firstframedtheideaofprobability,describingitin
termsofthepossibleoutcomesofgamblingevents.
•  ProbabilityquicklybecameaninvaluablepartofallthequanOtaOve
sciences,helpingtodealwithuncertainmeasurementsandincomplete
theories.
•  JamesBernoulli(1654–1705),PierreLaplace(1749–1827),andothers
advancedthetheoryandintroducednewstaOsOcalmethods.
•  ThomasBayes(1702–1761),whoappearsonthefrontcoverofthe
textbook,proposedaruleforupdaOngprobabiliOesinthelightofnew
evidence.
•  Bayes’ruleunderliesmostmodernapproachestouncertainreasoningin
AIsystems.
Saturday,23April16
19
2.3Economics(ReaditYourself)
• 
• 
• 
• 
• 
• 
• 
Howshouldwemakedecisionssoastomaximizepayoff?(Readityourself)
WhiletheancientGreeksandothershadmadecontribuOonstoeconomic
thought,Smithwasthefirsttotreatitasascience,usingtheideathateconomies
canbethoughtofasconsisOngofindividualagentsmaximizingtheirown
economicwell-being.
Decisiontheory,whichcombinesprobabilitytheorywithuOlitytheory(preferred
outcomeslikedollarorhamburger),providesaformalandcompleteframework
fordecisions(economicorotherwise)madeunderuncertainty—thatis,incases
whereprobabilisOcdescripOonsappropriatelycapturethedecisionmaker’s
environment.
Howshouldwedothiswhenothersmaynotgoalong?
Decisiontheoryissuitablefor“large”economieswhereeachagentneedpayno
a_enOontotheacOonsofotheragentsasindividuals.
Howshouldwedothiswhenthepayoffmaybefarinthefuture?
Thistopicwaspursuedinthefieldofopera#onsresearch.TheworkofRichard
Bellman(1957)formalizedaclassofsequenOaldecisionproblemscalledMarkov
decisionprocesses,whichwewillstudyinChapters17and21.
Saturday,23April16
20
2.4Neuroscience
•  HowdobrainsprocessinformaOon?
•  Neuroscienceisthestudyofthenervoussystem,parOcularlythe
brain.Theexactwayinwhichthebrainenablesthoughtisoneof
thegreatmysteriesofscience.
•  PaulBroca’s(1824–1880)studyofaphasia(speechdeficit)showed
thatspeechproducOonwaslocalizedtotheporOonofthelee
hemispherenowcalledBroca’sarea.
•  BythatOme,itwasknownthatthebrainconsistedofnervecells,
orneurons,butitwasnotunOl1873thatCamilloGolgi(1843–
1926)developedastaining(markordiscolour)techniqueallowing
theobservaOonofindividualneuronsinthebrain(seeFigure1.2
fromthetextbook).
•  NicolasRashevsky(1936,1938)wasthefirsttoapplymathemaOcal
modelstothestudyofthenervoussystem.
Saturday,23April16
21
2.4Neuroscience
Saturday,23April16
22
2.4Neuroscience
• 
• 
• 
• 
• 
Wenowhavesomedataonthemappingbetweenareasofthebrainandtheparts
ofthebodythattheycontrolorfromwhichtheyreceivesensoryinput.
Moreover,wedonotfullyunderstandhowotherareascantakeoverfuncOons
whenoneareaisdamaged.Thereisalmostnotheoryonhowanindividual
memoryisstored.
Themeasurementofintact(undamaged)brainacOvitybeganin1929withthe
invenOonbyHansBergeroftheelectroencephalograph(EEG).Therecent
developmentoffuncOonalmagneOcresonanceimaging(fMRI)(Ogawaetal.,
1990;CabezaandNyberg,2001)isgivingneuroscienOstsunprecedentedlydetailed
imagesofbrainacOvity.
IndividualneuronscanbesOmulatedelectrically,chemically,orevenopOcally
(HanandBoyden,2007),allowingneuronalinput–outputrelaOonshipstobe
mapped.Despitetheseadvances,wearesOllalongwayfromunderstandinghow
cogniOve(mentalacOonorprocesstoacquireknowledgeandunderstanding)
processesactuallywork.
BrainsanddigitalcomputershavesomewhatdifferentproperOes.Figure1.3
showsthatcomputershaveacycleOmethatisamillionOmesfasterthanabrain.
Saturday,23April16
23
2.4Neuroscience
Saturday,23April16
24
2.5Psychology
•  Howdohumansandanimalsthinkandact?
•  ItisaboutthebehaviorandcogniOveprocess
ofhumansandanimals(Readityourself).
Saturday,23April16
25
2.6Computerengineering
•  Howcanwebuildanefficientcomputer?
•  ForarOficialintelligencetosucceed,weneed
twothings:intelligenceandanarOfact.The
computerhasbeenthearOfactofchoice.
•  EvoluOonofComputeranditssoewaretools
(Readityourself).
Saturday,23April16
26
2.7Controltheoryandcyberne#cs
• 
• 
• 
• 
• 
• 
• 
HowcanarOfactsoperateundertheirowncontrol?
KtesibiosofAlexandria(c.250B.C.)builtthefirstself-controllingmachine:awater
clockwitharegulatorthatmaintainedaconstantflowrate.ThisinvenOon
changedthedefiniOonofwhatanarOfactcoulddo.
Previously,onlylivingthingscouldmodifytheirbehaviorinresponsetochangesin
theenvironment.
Otherexamplesofself-regulaOngfeedbackcontrolsystemsincludethesteam
enginegovernor,createdbyJamesWa_(1736–1819),andthethermostat,
inventedbyCornelisDrebbel(1572–1633),whoalsoinventedthesubmarine.The
mathemaOcaltheoryofstablefeedbacksystemswasdevelopedinthe19th
century.
ThecentralfigureinthecreaOonofwhatisnowcalledcontroltheorywasNorbert
Wiener(1894–1964).
ControltheoryisaninterdisciplinarybranchofengineeringandmathemaOcsthat
dealswiththebehaviorofdynamicalsystemswithinputs,andhowtheirbehavior
ismodifiedbyfeedback.
Cyberne#csisthescienceofcommunicaOonsandautomaOccontrolsystemsin
bothmachinesandlivingthings.
Saturday,23April16
27
2.7Controltheoryandcyberne#cs
•  Ashby’sDesignforaBrain(1948,1952)elaborated
thattheintelligencecouldbecreatedbytheuseof
homeostaOcdevicescontainingappropriatefeedback
loopstoachievestableadapOvebehavior.
•  Homeosta#c:thetendencytowardsarelaOvelystable
equilibriumbetweeninterdependentelements.
•  Moderncontroltheory,especiallythebranchknownas
stochasOcopOmalcontrol,hasasitsgoalthedesignof
systemsthatmaximizeanobjec#vefunc#onover
Ome.ThisroughlymatchesourviewofAI:designing
systemsthatbehaveopOmally.
Saturday,23April16
28
2.8Linguis#cs
•  Howdoeslanguagerelatetothought?
•  BehavioristtheorydidnotaddressthenoOonofcreaOvityinlanguage—it
didnotexplainhowachildcouldunderstandandmakeupsentencesthat
heorshehadneverheardbefore.
•  Chomsky’stheory—basedonsyntacOcmodelsgoingbacktotheIndian
linguistPanini(c.350B.C.)—couldexplainthis,andunlikeprevious
theories,itwasformalenoughthatitcouldinprinciplebeprogrammed.
•  ModernlinguisOcsandAI,then,were“born”ataboutthesameOme,and
grewuptogether,intersecOnginahybridfieldcalledcomputa#onal
linguis#csornaturallanguageprocessing.
•  Understandinglanguagerequiresanunderstandingofthesubjectma_er
andcontext,notjustanunderstandingofthestructureofsentences.
•  Muchoftheearlyworkinknowledgerepresenta#on(thestudyofhowto
putknowledgeintoaformthatacomputercanreasonwith)wasOedto
languageandinformedbyresearchinlinguisOcs.
Saturday,23April16
29
3.TheHistoryOfAr#ficialIntelligence
3.1.Thegesta#on(thedevelopmentoveraperiodofOme)ofar#ficialintelligence(1943–
1955)
•  WarrenMcCullochandWalterPifs(1943):Theydrewonthreesources:
– 
– 
– 
– 
• 
• 
• 
knowledgeofthebasicphysiologyandfuncOonofneuronsinthebrain;
aformalanalysisofproposiOonallogicduetoRussellandWhitehead;
andTuring’stheoryofcomputaOon.
Theyshowed,forexample,thatanycomputablefuncOoncouldbecomputedbysomenetworkof
connectedneurons,andthatallthelogicalconnecOves(and,or,not,etc.)couldbeimplementedby
simplenetstructures.
DonaldHebb(1949):demonstratedaruleformodifyingtheconnecOonstrengths
betweenneurons.Hisrule,nowcalledHebbianlearning,remainsaninfluenOalmodel
tothisday.
MarvinMinskyandDeanEdmonds(1950):builttheSNARC(neuralnetwork),asit
wascalled,used3000vacuumtubesandasurplusautomaOcpilotmechanismfroma
B-24bombertosimulateanetworkof40neurons.
AlanTuring:HeproposedtheChildProgrammeidea,explaining“Insteadoftryingto
produceaprogrammetosimulatetheadultmind,whynotrathertrytoproduceone
whichsimulatedthechild’s?”
Saturday,23April16
30
3.2.Thebirthofar#ficialintelligence
(1956)
•  JohnMcCarthy(1951):movedtoStanfordandthento
DartmouthCollege,whichwastobecometheofficial
birthplaceofthefieldofAI.
•  Forthenext20years,thefieldwouldbedominatedby
somepeopleandtheirstudentsandcolleaguesatMIT,
CMU,Stanford,andIBM.
•  AIfromthestartembracedtheideaofduplicaOnghuman
faculOessuchascreaOvity,self-improvement,andlanguage
use.Noneoftheotherfieldswereaddressingtheseissues.
•  AIistheonlyoneofthesefieldsthatisclearlyabranchof
computerscience,andAIistheonlyfieldtoa_emptto
buildmachinesthatwillfuncOonautonomouslyincomplex,
changingenvironments.
Saturday,23April16
31
3.3.Earlyenthusiasm,great
expecta#ons(1952–1969)
•  NewellandSimon’searlysuccesswasfollowedupwiththeGeneral
ProblemSolver,orGPS.
–  ThesuccessofGPSandsubsequentprogramsasmodelsofcogniOonled
NewellandSimon(1976)toformulatethefamousphysicalsymbolsystem
hypothesis,whichstatesthat“aphysicalsymbolsystemhasthenecessaryand
sufficientmeansforgeneralintelligentacOon.”
•  AtIBM,NathanielRochesterandhiscolleaguesproducedsomeofthe
firstAIprograms.HerbertGelernter(1959)constructedtheGeometry
TheoremProver,whichwasabletoprovetheoremsthatmanystudentsof
mathemaOcswouldfindquitetricky.
•  StarOngin1952,ArthurSamuelwroteaseriesofprogramsforcheckers
(draughts)thateventuallylearnedtoplayatastrongamateurlevel.Along
theway,hedisprovedtheideathatcomputerscandoonlywhattheyare
toldto:hisprogramquicklylearnedtoplayabe_ergamethanitscreator.
•  McCarthydefinedthehigh-levellanguageLisp,whichwastobecomethe
dominantAIprogramminglanguageforthenext30years.
Saturday,23April16
32
3.3.Earlyenthusiasm,great
expecta#ons(1952–1969)
–  Alsoin1958,McCarthypublishedapaperenOtledProgramswithCommon
Sense,inwhichhedescribedtheAdviceTaker,ahypotheOcalprogramthat
canbeseenasthefirstcompleteAIsystem.
•  HisplantouselogictobuildtheulOmateAdviceTakerwasadvancedbyJ.
A.Robinson’sdiscoveryin1965oftheresoluOonmethod(acomplete
theorem-provingalgorithmforfirst-orderlogic;seeChapter9).
•  Minskysupervisedaseriesofstudentswhochoselimitedproblemscalled
microworlds.
–  JamesSlagle’sSAINTprogram(1963)wasabletosolveclosed-formcalculus
integraOonproblemstypicaloffirstyearcollegecourses.
–  TomEvans’sANALOGYprogram(1968)solvedgeometricanalogyproblems
thatappearinIQtests.
–  DanielBobrow’sSTUDENTprogram(1967)solvedalgebrastoryproblems.
–  Themostfamousmicroworldwastheblocksworld(SHRDLU),whichconsists
ofasetofsolidblocksplacedonatabletop(ormoreoeen,asimulaOonofa
tabletop),asdiscussedonthenextslide.
Saturday,23April16
33
3.3.Earlyenthusiasm,great
expecta#ons(1952–1969)
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
ThedialogthatwasusedasaSHRDLUdemo:
Person:PICKUPABIGREDBLOCK.
Computer:OK.(doesit)
Person:GRASPTHEPYRAMID.
Computer:IDON'TUNDERSTANDWHICHPYRAMIDYOUMEAN.
Person:FINDABLOCKWHICHISTALLERTHANTHEONEYOUAREHOLDINGANDPUTITINTOTHEBOX.
Computer:BY"IT",IASSUMEYOUMEANTHEBLOCKWHICHISTALLERTHANTHEONEIAMHOLDING.
Computer:OK.(doesit)
Person:WHATDOESTHEBOXCONTAIN?
Computer:THEBLUEPYRAMIDANDTHEBLUEBLOCK.
Andsoon
Saturday,23April16
34
3.4.Adoseofreality(1966–1973)
•  Fromthebeginning,AIresearcherswerenotshyabout
makingpredicOonsoftheircomingsuccesses.The
followingstatementbyHerbertSimonin1957isoeen
quoted:
“Itisnotmyaimtosurpriseorshockyou—butthe
simplestwayIcansummarizeistosaythattherearenow
intheworldmachinesthatthink,thatlearnandthat
create.Moreover,theirabilitytodothesethingsisgoing
toincreaserapidlyunOl—inavisiblefuture—therangeof
problemstheycanhandlewillbecoextensivewiththe
rangetowhichthehumanmindhasbeenapplied.”
•  Restofliterature,readityourself.
Saturday,23April16
35
Remaining
•  3.5Knowledge-basedsystems:Thekeytopower?
(1969–1979)
•  3.6AIbecomesanindustry(1980–present)
•  3.7Thereturnofneuralnetworks(1986–present)
•  3.8AIadoptsthescien#ficmethod(1987–present)
•  3.9Theemergenceofintelligentagents(1995–
present)
•  3.10Theavailabilityofverylargedatasets(2001–
present)
Saturday,23April16
36
THESTATEOFTHEART
•  WhatcanAIdotoday?Aconciseanswerisdifficultbecausethere
aresomanyacOviOesinsomanysubfields.Herewesampleafew
applicaOons;
•  UnderstandingandProcessingNaturalLanguages:
•  RoboOcvehicles:
•  SpeechrecogniOon:
•  Autonomousplanningandscheduling:
•  Gameplaying:
•  SpamfighOng:
•  LogisOcs(thedetailedorganizaOonandimplementaOonofa
complexoperaOon)planning:
•  RoboOcs:
•  MachineTranslaOon:
Saturday,23April16
37
AssignmentNo.1
Saturday,23April16
38