Graduate AI. Lecture 1: Intro. Teachers: Martial Hebert Ariel Procaccia (this time)

Transcription

1 Graduate AI Lecture 1: Intr Teachers: Martial Hebert Ariel Prcaccia (this time)

2 Curse website 2

3 Bureaucracy TAs: Jhn Dickersn and Felipe Trevizan Useful but nt required bk: Russel and Nrvig, Intrductin t Artificial Intelligence, 3 rd editin 5 hmewrks (35%), midterm exam (20%), prject (35%), participatin (10%) Prject: prpsal, interim reprt, final reprt, shrt presentatin, pster 3

4 Bureaucracy Let us knw if yu did nt receive an frm me n Mnday Ddle pll n recitatins Audits: nly d final prject Jhn is rganizing an AI reading grup 4

5 Style Prbably mre mathematical than the average Grad AI (nntrivial prfs n blackbard) Backgrund: algrithms+cmputatinal cmplexity, basic prbability thery, basic linear algebra, mathematical maturity Schedule shws teachers preferences: MAS / game thery / scial chice (me), perceptin and mtin planning (Martial) Fun (especially this talk) 5

6 AI timeline (NYT 2011) 6

7 Perfrmed fr 84 years Defeated Naplen and Franklin Amazn Mechanical Turk: artificial artificial intelligence 7

8 Amazn Mechanical Turk 8

9 Big questin: can machines think? Mre cncrete questin: can machines d well in the imitatin game? Judge cmmunicates via text channel with cmputer and human, must reliably identify the cmputer 9

10 The Chinese Rm Thught experiment prpsed by Searle Suppse AI has prduced a prgram that can pass the Turing Test in Chinese Yu have a handbk with its pseudcde Yu re in a clsed rm and receive Chinese characters thrugh a slt Yur run the prgram s cde manually and return the utput Des this mean yu understand Chinese? 10

11 Cunterarguments Finding the mind: the whle system understands Chinese, the persn is just a part f the system Redesigning the experiment: suppse the prgram simulates the actins f every neurn in the brain f a Chinese speaker 11

12 Cunterarguments Dubts abut the intuitin Brain perfrms 100 billin peratins per secnd, s it wuld take the persn millins f years t simulate a simple answer Churchland s Luminus Rm: suppse yu are standing in a dark rm and quickly mving a magnet up and dwn, then by Maxwell s thery f artificial luminance it will be luminus. Hwever, this requires 450 trillin mvements per secnd 12

13 Audrey culd recgnize digits spken by a single vice In 1962 IBM demnstrated Shebx, which culd understand 16 wrds Biggest milestne in the Seventies: CMU s Harpy system, which culd understand 1011 wrds vcabulary f three-year-ld 13

14 Samuel s prgram was based n alpha-beta pruning Actually nly cmpeted at respectable amateur level By the Nineties checkers prgrams were beating the best human players Checkers was slved by Jnathan Schaeffer in 2007 after 18 years f calculatin 14

15 Shakey = first mbile rbt t visually interpret envirnment Can lcate items, navigate arund them, and reasn abut its actins I (4:08) 15

16 Started as ChipTest at CMU, fllwed by Deep Thught After graduatin, develpers were hired by IBM Defeated Kasparv in 1997 Kasparv played anti-cmputer pening mves t get Deep Blue ut f its pening bk Kasparv accused IBM f cheating 16

17 Advanced Step in Innvative Mbility (resemblance t Asimv is a cincidence) Can recgnize mving bjects, pstures, gestures, its surrunding envirnment, sunds and faces, which enables it t interact with humans DfW4 17

18 DARPA Urban Challenge 96 km urban area curse, t be cmpleted < 6 hurs, tk place in 2007 Tartan Racing (CMU+GM) claimed the $2 millin prize Challenge invlves missin planning, mtin planning, behavir generatin, perceptin, wrld mdeling k0 18

19 Watsn defeated the tw greatest-ever Jepardy! champins Invlves natural language prcessing, infrmatin retrieval, knwledge representatin and reasning, and machine learning _9c 19

20 The future 20

21 The technlgical singularity Emergence f superhuman intelligence Key idea: self-imprvement Surce f name: analgy between inability t predict events after the develpment f a superintelligence, and the spacetime singularity beynd the event hrizn f a black hle Sme predict: this century Others argue: never 21

22 What is AI? Simplest (but self-referential) answer: lk at the call fr papers f the Internatinal Jint Cnference n Artificial Intelligence 22

23 IJCAI 11 call fr papers Agent-based and multiagent systems The visin: electrnic cmmerce, supply chains, defense systems managed by autnmus sftware agents Cperatin and crdinatin Emergent behavir Cmputatinal game thery and cmputatinal scial chice 23

24 Can GT enable classic AI? Classic AI seeks t design intelligent rbts/agents GT distills ratinality Ratinality is perceived as intelligence Game-theretic MAS may be perceived as intelligent GT enables AI n the multiagent level! 24

25 IJCAI 11 call fr papers Search and cnstraints Uninfrmed search Infrmed search Cnstraint satisfactin can be seen as searching thrugh assignments Cnditins fr backtrack-free search Existence f satisfying assignments 25

26 IJCAI 11 call fr papers Planning (and scheduling) Classical planning Mtin/path planning D* = cmpute path under assumptins, recalculate when new infrmatin is discvered 26

27 IJCAI 11 call fr papers Uncertainty in AI Bayesian netwrks Graphical mdels Markv decisin prcesses (Rbtics and) Visin Object recgnitin Scene understanding 27

28 Object recgnitin Frm Martial Hebert 28

29 Nt cvered Machine learning Natural Language Prcessing Knwledge representatin and reasning Web and knwledge-based infrmatin systems 29

30 Themes Cmplex envirnments representatin Cmplex envirnments large size heuristics Mimic natural wrld / human scieties Synergies between different subdisciplines 30

31 rbcup Rbtic sccer cmpetitin Official gal: By mid-21st century, a team f fully autnmus humanid rbt sccer players shall win the sccer game, cmplying with the fficial rule f the FIFA, against the winner f the mst recent Wrld Cup Cmpnents are similar t autnmus driving, with a multiagent twist (1:40) 31

32 Recmmender systems Step 1: find users with similar ratings Step 2: recmmend item based n their ratings Cmmn mdel fr cllabrative filtering 32

33 Kidney exchange In US, 50,000/yr are diagnsed with lethal kidney disease Cure = transplantatin, but cadaver kidney have a lng waiting list (2-5 yrs) Ptential dnrs may be incmpatible with patient Pairs f incmpatible dnr-patient pairs can smetimes exchange kidneys 33

34 Optimizing kidney exchange Cycle cver = ptimize matched vertices under cycles f length L Prblem is NPhard Regularly slved by ur TA!

35 Challenges Uncertainty: peple enter and leave the pl, matches fail stchastic ptimizatin Hspitals manipulate by hiding sme f their patients and matching internally game thery / mechanism design 35