Artificial Intelligence Introduction to AI

Arya MirAI and Robotics

Nov 26, 2011 (5 years and 8 months ago)

855 views

COMP2039ArtificialIntelligence
IntroductiontoAI
BobDamper
February7,2006
COMP2039ArtificialIntelligenceIntroductiontoAI–p.1/19
Housekeeping
LecturersareSrinandanDasmahaptra,BobDamper
andMichaelLuck.
Coursetext:S.RussellandP.Norvig(2003)Artificial
Intelligence:AModernApproach(SecondEdition),
PrenticeHall,UpperSaddleRiver,NJ.It’sabigbook
andwewillonlybestudyingafractionofitscontent.
Syllabus:
Introduction
Search
Reasoning
Uncertainty
MachineLearning
Applications
PhilosophicalIssues
COMP2039ArtificialIntelligenceIntroductiontoAI–p.2/19
Issues
Whatisintelligence?
Whatisartificialintelligence?
Thecomputationalmetaphor
TheTuringtest
Somebasicdichotomies:
strongversusweakAI
neatversusscruffyAI
symbolicversussub-symbolicAI
knowledge-basedversusdata-drivenAI
embodiedversus‘disembodied’AI
What’sthedifferencebetweenAIandcomputer
science?AIasthedomainofhardproblems.
COMP2039ArtificialIntelligenceIntroductiontoAI–p.3/19
WhatisIntelligence?
AIhasprovednotoriouslydifficulttodefine,reflectingthe
difficultyofdefiningnaturalintelligence:
“In1921,theJournalofEducationalPsychologyasked
fourteenleadingexperts...toprovidetheir
definitions...[they]got14differentanswersback.”(Pfeifer
andScheirer1999,p.6)
Oneinfluentialattemptatdefiningnaturalintelligenceisdue
tovanHeerden(1968):
“Intelligentbehavioris:toberepeatedlysuccessfulin
satisfyingone’spsychologicalneedsindiverse,observably
different,situationsonthebasisofpastexperience.”
Buthowcanonedefine“psychologicalneeds”scientifically?
COMP2039ArtificialIntelligenceIntroductiontoAI–p.4/19
SomeDifferentNotionsofIntelligence
Someofthe14differentnotionswere:
“Theabilitytocarryonabstractthinking”(Terman)
“...abilitytolearntoadjust...totheenvironment”
(Colvin)
“...abilitytolearntoadapt...adequatelytonew
situationsinlife”(Pinter)
“Thecapacitytoacquirecapacity”(Woodrow)
“Thecapacitytolearnorprofitbyexperience”
(Dearborn)
“Abiologicalmechanismbywhichtheeffectsofa
complexityofstimuliarebroughttogether...”
(Peterson)
COMP2039ArtificialIntelligenceIntroductiontoAI–p.5/19
CommonsenseNotionsofIntelligence
PfeiferandScheirer(1999,pp.7–12)identifysome
‘commonsense’characteristicsofintelligentbehaviour:
Graduated...
Thinkingandproblemsolving
Learningandmemory
Language
Intuitionandcreativity
Emotions
Survivinginacomplexworld
Perceptualandmotorabilities
COMP2039ArtificialIntelligenceIntroductiontoAI–p.6/19
SowhatisArtificialIntelligence?
Anearlyattemptatdefinitionis“ArtificialIntelligenceis
thescienceofmakingmachinesdothingsthatwould
requireintelligenceifdonebymen”(Minsky1968,p.v).
Thisviewremainscurrent–Nilsson(1998,p.1)writes
that“Artificialintelligence,broadly(andsomewhat
circularly)defined,isconcernedwithintelligent
behaviorinartifacts”.
Yes,thesearecircular!
Howabout:
“AIistheattempttomechanisethoughtby
technologicalmeans.”?
Butthenwhatis“thought”?!
MORAL:definitionofAIisnoteasy!
COMP2039ArtificialIntelligenceIntroductiontoAI–p.7/19
TheComputationalMetaphor
Theideathatintelligentbehaviourisaformof
computationgoesbacktoTuring(1936,1950).
AsNewell(1990)says:“AIisabranchofcomputer
science”.
Thephysicalsymbolsystem(PSS)hypothesis(Newell
1980)...thephysicalmediumdoesn’tmatter!
“Ifthehypothesisthatthebrainisacomputeristo
makeadefiniteclaim,theword‘computer’musthavea
precisemeaning”(Copeland1993,p.204).
Yet“...despitesome50yearsofstudy...,thereisstill
noconsensusonjustwhataretheessentialelements
ofcomputing”(Pylyshyn1984,p.69).
COMP2039ArtificialIntelligenceIntroductiontoAI–p.8/19
TheChurch-TuringThesis
AccordingtotheC-Tthesis,“computersarecapableof
implementinganyeffectivelydescribedsymbolic
process”(LugerandStubblefield1993,p.30).
Turingproposedessentiallythisthesisintermsofthe
universalTuringmachine...
...andsodidChurchintermsoflambdacalculus.
SoanotherformoftheC-Tthesisisthatalleffective
computationalproceduresareequivalent.This
includesnotonlylambdacalculusandtheUTM,but
cellularautomata(vonNeumann),productionrules
(Post),formalgrammars(Chomsky),etc.
Becauseeffectivecomputationisuniversal,itisagood
candidateforsimulatinghumanintelligence–whichis
yetanotherversionoftheC-Tthesis.
COMP2039ArtificialIntelligenceIntroductiontoAI–p.9/19
OtherMetaphors
Whatothermetaphorsareavailableforintelligent
behaviour?
Dynamicalsystems?
Analoguecomputation?
Agent-basedorcollectivecomputing?
Neuralcomputing?
Artificialevolution?
Learningsystems?
EmbodiedAI?
Butaren’ttheseall(ormostly)Turing-equivalentandso
containedwithintheC-Tthesis?
COMP2039ArtificialIntelligenceIntroductiontoAI–p.10/19
TheTuringTest
Turing,1950:“Iproposetoconsiderthequestion,‘Can
machinesthink?’”...(excepthedidn’t!)
Itisagiventhathumansareintelligent.
Turingproposedan“imitationgame”inwhichaplayer
communicatingviaateletypewithamanandawoman
inremote,hiddenroomswasrequiredtotellwhichwas
which.(Onecouldlie,theotherhadtotellthetruth.)
Nowswapacomputerforoneofthehumans.By
extension,ifwecannottellwhichishumanandwhich
ismachine,themachinemustbeintelligent.
SeveralvariantsoftheTuringtest.
Usualonethesedaysistodetermineifa(single)
conversationalentityisahumanoracomputer.
COMP2039ArtificialIntelligenceIntroductiontoAI–p.11/19
COMP2039ArtificialIntelligenceIntroductiontoAI–p.12/19
ProblemswiththeTuringTest?
Equatesintelligencewithhumanabilities.
Equatesintelligencewithlinguisticabilities.
Observermaybegullibleand/orunwillingtodeclare
“it’sacomputer”forfearofbeingwrong.
Howtolimitthedomainofdiscourse?
Howlongdowecontinuethetest?
Someinterestingreading:
PRO:B.J.Copeland(2000)“TheTuringtest”,Mindsand
Machines,10(4),pp.519–539.
ANTI:R.M.French(2000)“TheTuringtest:Thefirst
50years”,TrendsinCognitiveScience,4(3),
pp.115–122.
COMP2039ArtificialIntelligenceIntroductiontoAI–p.13/19
TheTotalTuringTest
S.Harnad(1989)“Minds,machinesandSearle”,Journalof
TheoreticalandExperimentalArtificialIntelligence,1(1),
pp.5–25.
“WhoistosaythattheTuringTest...couldbe
successivelypassedwithoutoperationsthatdrawon
oursensory,motorandotherhighercognitive
capacitiesaswell”.
Harnaddistinguishesbetweenthe“teletypeTuring
test”andthe“roboticTuringtest”.
TheteletypeTuringtestispurelyverbal(symbols
in/symbolsout).
Therobotic(‘total’)Turingtestrequiresthe‘computer’
tolookandactlikeahuman.
COMP2039ArtificialIntelligenceIntroductiontoAI–p.14/19
HardandEasyProblems
Hardproblemsforhumans:
Fasterror-freearithmetic
Provingtheoremsinmathematics
Playingchess
Easyproblemsforhumans:
Speak,recogniseandunderstandspeech
Graspaglasswithoutbreakingit
Veryoften,easyproblemsforhumansarehard
problemsforcomputers!
AIisthedomainofhardcomputationalproblems,such
asspeechrecognition;imageunderstanding;flexible
graspingandmanipulationofdelicateobjects,etc.
COMP2039ArtificialIntelligenceIntroductiontoAI–p.15/19
ExamplesofAISystems
“GoodoldfashionedAI”orGOFAI(atermduetoJohn
Haugeland)...
SHRDLU
a
,asimulationofarobotmovingina‘blocks
world’controlledbyanaturallanguage‘understanding’
interface.DuetoWinograd(1972)soitisold!
MYCIN:arule-basedexpertsystemtodiagnoseblood
infections.Alsoratherold(Shortliffe1976).
MYCINoutperformedmostdoctors.
So,coulditbesaidtopasstheTuringtest???
a
SecondcolumnofkeysonaLinotypemachine.
COMP2039ArtificialIntelligenceIntroductiontoAI–p.16/19
MoreExamples
...gettingmoreuptodate:
TheNETtalkneuralnetwork...
T.J.SejnowskiandC.R.Rosenberg(1987)“Parallel
networksthatlearntopronounceEnglishtext”,
ComplexSystems,1(1),pp.145–168.
Emailfilters...
M.Sahami,S.Dumais,D.HeckermanandE.Horvitz
(1998)“ABayesianapproachtofilteringjunkemail”,
AAAIWorkshoponLearningforTextCategorization,
Madison,WI.AAAITechnicalReportWS-98-05
TheGooglesearchengine...
COMP2039ArtificialIntelligenceIntroductiontoAI–p.17/19
NETtalk
cat
----
output￿units
hidden￿units
input￿units
/k/
Middleletterof7-letterwindowconvertedtoits
corresponding‘phoneme’(i.e.,abstractsound).
Networktrainedonerrorback-propagationalgorithm.
COMP2039ArtificialIntelligenceIntroductiontoAI–p.18/19
FilteringJunkEmail
AbstractofSahamietal.(1998)
InaddressingthegrowingproblemofjunkemailontheInternet,weexamine
methodsfortheautomatedconstructionoffilterstoeliminatesuchunwanted
messagesfromauser’smailstream.Bycastingthisprobleminadecision
theoreticframework,weareabletomakeuseofprobabilisticlearning
methodsinconjunctionwithanotionofdifferentialmisclassificationcostto
producefilterswhichareespeciallyappropriateforthenuancesofthistask.
Whilethismayappear,atfirst,tobeastraightforwardtextclassification
problem,weshowthatbyconsideringdomain-specificfeaturesofthis
problem,inadditiontotherawtextofE-mailmessages,wecanproduce
muchmoreaccuratefilters.Finally,weshowtheefficacyofsuchfiltersina
realworldusagescenario,arguingthatthistechnologyismatureenoughfor
deployment.
COMP2039ArtificialIntelligenceIntroductiontoAI–p.19/19