Lecture S2: Artificial Intelligence

LectureS2:ArtificialIntelligence
LectureS2:ArtificialIntelligence
3
Overview
AwhirlwindtourofArtificialIntelligence.
Wespendjustonelecture,butthereare:
￿
EntirecourseonAI.COS302
￿
NewAIProfessoratPrinceton.RobSchapire
Today'slevelofaspiration.
￿
Aquicksurveyofseveralimportanttopics.
4
Origins
Ideaofprogrammingcomputersfor"intelligent"behavior.
￿
FirstsuggestedbyAlanTuring,1950.
Term"artificialintelligence"coinedbyJohnMcCarthyin1955.
Dartmouthsummerconference,1956.
￿
Gatheringoffield'sluminaries.
￿
Veryoptimistic!
Optimisticpredictionsverycommonin1950sand1960s.
￿Actualprogressmuchslowerthanpredicted.
￿Somestrikingsuccesses;stilllotstodo.
J.McCarthy,1951
"Everyaspectoflearningoranyotherfeatureof
intelligencecaninprinciplebesoprecisely
describedthatamachinecanbemadetosimulateit."
5
8-puzzle
Slidetilesuntilthey'reinnumericalorder.
Ofcoursethereis:8-puzzleontheWeb.
Whatstrategytousetosolvepuzzle?
￿
Noobviousalgorithm.
Fromthisposition,
￿
Threesuccessorpositions.
￿
Fromeachofthese,
–
Two,three,orfoursuccessors.
–
Andsoon.
Soundslikeatree?
￿
Yup.
6
2
8
4
1
3
5
7
6
8-puzzle:Breadth-FirstSearch
SearchtreeforGoalnode,onelevelatatime.
7
8-puzzle:Depth-FirstSearch
SearchforGoaldowntosomedepth(thesearch"horizon").
8
8-puzzle:HeuristicSearch
Breadth-FirstandDepth-Firstare"blind"searches.
￿
Exhaustivemethodsforfinding
pathtogoal.
￿
Ofteninfeasiblebecausetoomany
nodesexpanded.
￿
Success(eventually)guaranteed.
"Heuristic"search.
￿
Uses"evaluationfunction"torank
successornodes;pickbest.
￿
Noguaranteeofsuccess.
￿
Exampleusesdistancefromstart
plusnumberoftilesoutofplace.
￿
Manyotherfunctionspossible.
￿
Note:only6nodesexpanded.
9
GameTrees
Whatifyouhaveanopponent?
"Minimax"methods.
￿
Youchoosebestmoveassumingopponentchoosestheirbestresponse.
￿
Opponentchoosestheirbestmoveassumingyoushouldbestresponse.
Youchooseyour
bestmove.(max)
Opponentchooses
bestreply.(min)
Youchoose...
-6
24
5
-3
8
2
-6
-3
2
2
10
Tic-Tac-Toe
Partialgametree.
11
Tic-Tac-Toe
Reducetreebyrecognizingsymmetries.
12
SizesofGameTrees
8-puzzle.
￿
Firsttilecanbeinanyoneof9squares.
￿
Secondtileinoneof8.
￿
Totalnumberofnodes=9!/2=181,440
Tic-tac-toe(ignoringsymmetries).
￿
Firstmove:oneof9spots.
￿
Secondmove:oneof8.
￿
Somegamesterminatebefore9movesmade.
￿
Totalnumberofnodes<9!=362,880
Bothnumberssmall,soexhaustivesearchfeasible.
Butwhataboutsomebiggergame,forinstance...
14
Chess
AfavoritetargetofAIresearcherssince1950's.
Howbigisgametree?
￿
20possiblefirstmoves.
￿
35possiblenextmovesonaverage.
–
called"branchingfactor"
￿
Supposegamesaverage40moves(eachside).
￿
Completegametreethereforehas:
–
3580
≥
≥≥
≥10120
nodes!
Soifeachelectronintheuniversewereasupercomputer,etc.,etc.
Anycomputer(andanyperson)cansearchonlythetiniestfractionof
thecompletegametree.
15
Chess
Lotsofeffort,andsteadyimprovements,since1950's.
DeepBluechessmachinedevelopedatIBM.
￿
Hsu,Campbell,Tan,Hoane,Brody,Benjamin,1990's.
￿
32-processorparallelcomputer.
￿
512specialchesschips.
￿
Evaluates200millionpositions/sec.
￿
Hugeamountofchessknowledgeprogrammedin.
￿
Usescleverheuristicfunctionsandminimaxtreesearch.
GarryKasparovvs.DeepBlue(1997)
VladimirKramnikvs.DeepFritz(2002)
GarryKasparovvs.DeepJunior(2003)
16
Backgammon
Backgammon.
￿
Movesuserollof2dice.
–
21differentoutcomes
￿
Manypossiblemovesperdiceroll.
–
about20onaverage
￿
Branchingfactorofabout400.
–
toobig!
Searchesofevenmoderatedepthcomputationallyprohibitive.
￿
GoodheuristicfunctionVERYimportant.
Somethingbettermustbedone...
17
Uses
neuralnetwork
and
reinforcementlearning.
G.Tesauro,1995.
￿
Eachconnectionhasaweight.
￿
Eachnodeoutputsweightedsumofinputs.
￿
Weightsadjustedateachstep.
Inputpatterns.(198)
￿
Describeboardposition.
Outputpatterns.(1)
￿
Probabilityofwinningfromgiven
boardposition.
Weightadjustmentfunction.
￿
Triestomakesuccessivepredictionsagree.
￿
Hence:temporaldifference.
TD-Gammon
22
LoebnerContestDialogs
"Hello,Judge.MynameisElla.Pleasetellmeyourfirstname,or
whateverIshouldcallyou."
"CallmeLisa.AmImaleorfemale?"
"Okay,Iwillcallyou'Lisa.AmIMaleOrFemale?'Iamhererecently
fromTainjin,China,sopleasebepatientwithme."
"I'mnotapatientbecauseIamnotill.Areyoufeelingwelltoday?
"Whatdoyouthinkitmeansthatyouarenotill?"
"Haveyoueverbeenill?"
"'Haveyoueverbeenill?'yousay?That'sit,LisaAmIMaleOr
Female!Iwillscheduleawhippingforyou."
23
LoebnerContestDialogs
"WhatdoyouthinkofReading?"
"I'msomewhatcrazyaboutreading.Iliketoreadmysterynovels."
"Imeantthetown."
"Joneslivesinthecityandisawriterofchildren'sbooks."
"DoesJonesliveinReading,then?"
"Tomorrowisanotherday.Wecanonlylivetodaywell."
"Doyoulive,then?"
"Ilivethenalot.Somedaywewillliveonanotherplanet."
"Ithinkyouarealreadythere."
24
LoebnerContestDialogs
"Whatisit?"
"Whatiswhat?"
"Ahigh-yield,perennial,four-lobedhybridofwheatandryewhose
rootgrain,triticale,cantraceitsancestrytotwentieth-century
Canada."
"Atlast.Anotherclassicgrainfreak."
"Lewis.JimLewis.Wouldyoulikeformetoshowyousomepoetry?"
"No,thanks."
"Ididn'twanttoshowyouthestupidpoetryanyway."
25
ChineseRoomExperiment
Imaginethatyou:
￿
Don'tunderstandChinese.
￿
Arealoneinaroomthathaspaperslotslabeledinputandoutput.
￿
HaveabigbookofChinesewriting.
￿
HaveEnglishinstructions(notranslations)thattellyouwhatto
writeonyouroutputpaperinresponsetovariousinputs.
Andthen:
￿
Chinesespeakersoutsidetheroompassinpiecesofpaperwith
Chinesewriting.Theyknowthesearequestions(butyoudon't).
￿
Youconsultyourmanualofinstructions,figureouttheproper
Chineseresponse,copyitdown,andpassitout.
TheBigQuestion.Searle,1980.
￿
ThefolksoutsidethinkyouunderstandChinese.Doyou?
￿
Ifacomputerdidthesame,woulditunderstandChinese?
26
ChineseRoomExperiment
"Weak"AIhypothesis.
￿
MachinescanbeprogrammedtoEXHIBITintelligentbehavior.
￿
Surelytrue:DeepBlue,TD-Gammon,others.
￿
Programsusemethodsverydifferentfromhumans.
–
performance(oftask)vs.simulation(ofhumanmethods).
"Strong"AIhypothesis.
￿
MachinescanbeprogrammedtoPOSSESSintelligence.
￿
Musttheyusebrain-likemethods(e.g.,neuralnets)?
–
"Connectionism"
SearleusedChineseRoomasabsoluterefutationofthepossibilityof
strongAI.
Butmanydisagree!
27
"Reverse"TuringTest
StandardTuringTest:judgeishuman.
ReverseTuringTest:judgeiscomputer!
Why?
￿
Yahooallowseachuser15MbytesofWebstorage.
–
Youwritea"bot"totosignup1millionusers.
–
Congratulations.Younowhave15Terabytesofstorage!
￿
PayPal
offers$5foreachuserwhoopensanewaccount.
–
Youwriteabottosignup1billionusers.
–
Congratulations.Younowhave$5,000,000,000!
￿
Onlinepolls.
￿
Spamfiltering.
￿
Allneedtodistinguishrealhumansfrombots(programs).
How?captcha.net
28
AIQuotes
"JustastheWrightbrothersatKittyHawkin1903wereontheright
tracktothe747,sotooisAI,withitsattemptstoformalize
commonsenseunderstanding,onitswaytofullyintelligent
machines."PatrickWinston
"Believingthatwritingthesetypesofprogramswillbringuscloserto
realartificialintelligenceislikebelievingthatsomeoneclimbingatree
ismakingprogresstowardreachingthemoon."HubertDreyfus
"Thebrainhappenstobeameatmachine."MarvinMinsky
"Eitherartificialintelligence
is
possible...or
we're
not."HerbSimon
"AIisanythinginsoftwarethatwedon'tknowhowtodoyet."
"Thequestionofwhetheracomputercanthinkisnomoreinteresting
thanthequestionofwhetherasubmarinecanswim."E.W.Dijkstra
29
AIinEverydayLife
ManyexamplesofAImethodsatworkintherealworld.
MicrosoftOffice'shelpfultalkingpaperclip.
￿
R.I.P.?
Speechrecognition.
￿
Speakslowlyandclearlytothetelephonerobot.
Opticalcharacterrecognition(OCR).
￿
MakesU.S.PostalServicehappy.
Controlofspacecraft!
￿
AIsystemgivencontrolof
￿
DeepSpace1for2daysinMay1999.
30
OmittedTopics
Knowledgerepresentation.
Reasoning.
Expertsystems.
Naturallanguageunderstanding.
Speechrecognition.
Computervision.
Anddozensmore...
(Buthey,weonlyhadtheonelecture.)