Introduction to Artificial Intelligence. Agnieszka Nowak - Brzezińska

Size: px

Start display at page:

Download "Introduction to Artificial Intelligence. Agnieszka Nowak - Brzezińska"

Owen Barker
6 years ago
Views:

1 Intrductin t Artificial Intelligence Agnieszka Nwak - Brzezińska

2 Gals f this Curse This class is a brad intrductin t artificial intelligence (AI) AI is a very brad field with many subareas We will cver many f the primary cncepts/ideas But in 15 weeks we can t cver everything Class Web page

3 Tday s Lecture What is intelligence? What is artificial intelligence? A very brief histry f AI Mdern successes: Stanley the driving rbt An AI screcard Hw much prgress has been made in different aspects f AI AI in practice Successful applicatins The ratinal agent view f AI

5 What is Artificial Intelligence?

6 Sme Definitins (I) The exciting new effrt t make cmputers think machines with minds, in the full literal sense. (excited but nt really useful) Haugeland, 1985

7 Sme Definitins (II) The study f mental faculties thrugh the use f cmputatinal mdels. Charniak and McDermtt, 1985 A field f study that seeks t explain and emulate intelligent behavir in terms f cmputatinal prcesses. Schalkff, 1990 (Applied psychlgy & philsphy?)

8 Sme Definitins (III) The study f hw t make cmputers d things at which, at the mment, peple are better. (I can almst understand this ne). Rich & Knight, 1991

9 Outline f the Curse Knwledge representatin: prpsitinal lgic and first-rder lgic inference in Expert Systems Fuzzy lgic Rugh set Machine learning: classificatin trees Neural netwrks Ohers?

10 What is Intelligence? Intelligence: the capacity t learn and slve prblems (Websters dictinary) in particular, the ability t slve nvel prblems the ability t act ratinally the ability t act like humans Artificial Intelligence build and understand intelligent entities r agents 2 main appraches: engineering versus cgnitive mdeling

11 What is Artificial Intelligence? (Jhn McCarthy, Stanfrd University) What is artificial intelligence? It is the science and engineering f making intelligent machines, especially intelligent cmputer prgrams. It is related t the similar task f using cmputers t understand human intelligence, but AI des nt have t cnfine itself t methds that are bilgically bservable. Yes, but what is intelligence? Intelligence is the cmputatinal part f the ability t achieve gals in the wrld. Varying kinds and degrees f intelligence ccur in peple, many animals and sme machines. Isn't there a slid definitin f intelligence that desn't depend n relating it t human intelligence? Nt yet. The prblem is that we cannt yet characterize in general what kinds f cmputatinal prcedures we want t call intelligent. We understand sme f the mechanisms f intelligence and nt thers. Mre in:

12 What s invlved in Intelligence? Ability t interact with the real wrld t perceive, understand, and act e.g., speech recgnitin and understanding and synthesis e.g., image understanding e.g., ability t take actins, have an effect Reasning and Planning mdeling the external wrld, given input slving new prblems, planning, and making decisins ability t deal with unexpected prblems, uncertainties Learning and Adaptatin we are cntinuusly learning and adapting ur internal mdels are always being updated e.g., a baby learning t categrize and recgnize animals

13 Academic Disciplines imprtant t AI. Philsphy Lgic, methds f reasning, mind as physical system, fundatins f learning, language, ratinality. Mathematics Frmal representatin and prf, algrithms, cmputatin, (un)decidability, (in)tractability, prbability. Ecnmics utility, decisin thery, ratinal ecnmic agents Neurscience neurns as infrmatin prcessing units. Psychlgy/ hw d peple behave, perceive, prcess Cgnitive Science infrmatin, represent knwledge. Cmputer building fast cmputers engineering Cntrl thery design systems that maximize an bjective functin ver time Linguistics knwledge representatin, grammar

14 Histry f AI 1943: early beginnings McCullch & Pitts: Blean circuit mdel f brain 1950: Turing Turing's "Cmputing Machinery and Intelligence 1956: birth f AI Dartmuth meeting: "Artificial Intelligence name adpted 1950s: initial prmise Early AI prgrams, including Samuel's checkers prgram Newell & Simn's Lgic Therist : great enthusiasm Newell and Simn: GPS, general prblem slver Gelertner: Gemetry Therem Prver McCarthy: inventin f LISP

Histry f AI 1966 73: Reality dawns Realizatin that many AI prblems are intractable Limitatins f existing neural netwrk methds identified Neural netwrk research almst disappears 1969 85: Adding dmain

15 Histry f AI : Reality dawns Realizatin that many AI prblems are intractable Limitatins f existing neural netwrk methds identified Neural netwrk research almst disappears : Adding dmain knwledge Develpment f knwledge-based systems Success f rule-based expert systems, E.g., DENDRAL, MYCIN But were brittle and did nt scale well in practice Rise f machine learning Neural netwrks return t ppularity Majr advances in machine learning algrithms and applicatins Rle f uncertainty Bayesian netwrks as a knwledge representatin framewrk AI as Science Integratin f learning, reasning, knwledge representatin AI methds used in visin, language, data mining, etc

16 Histry f AI 1943 McCullch & Pitts: Blean circuit mdel f brain 1950 Turing's "Cmputing Machinery and Intelligence" 1956 adpted Dartmuth meeting: "Artificial Intelligence" 1950s Early AI prgrams, including Samuel's checkers prgram, Newell & Simn's Lgic Therist, Gelernter's Gemetry Engine 1965 reasning Rbinsn's cmplete algrithm fr lgical AI discvers cmputatinal cmplexity Neural netwrk research almst disappears Early develpment f knwledge-based systems AI becmes an industry Neural netwrks return t ppularity AI becmes a science The emergence f intelligent agents

17 Different Types f Artificial Intelligence 1. Mdeling exactly hw humans actually think 2. Mdeling exactly hw humans actually act 3. Mdeling hw ideal agents shuld think 4. Mdeling hw ideal agents shuld act Mdern AI fcuses n the last definitin we will als fcus n this engineering apprach success is judged by hw well the agent perfrms

18 What s invlved in Intelligence? Ability t interact with the real wrld t perceive, understand, and act e.g., speech recgnitin and understanding and synthesis e.g., image understanding e.g., ability t take actins, have an effect Reasning and Planning mdeling the external wrld, given input slving new prblems, planning, and making decisins ability t deal with unexpected prblems, uncertainties Learning and Adaptatin we are cntinuusly learning and adapting ur internal mdels are always being updated e.g., a baby learning t categrize and recgnize animals

19 What s invlved in Intelligence? (again) Perceiving, recgnizing, understanding the real wrld Reasning and planning abut the external wrld Learning and adaptatin S what general principles shuld we use t achieve these gals?

20 Acting humanly: Turing test Turing (1950) "Cmputing machinery and intelligence "Can machines think?" "Can machines behave intelligently? Operatinal test fr intelligent behavir: the Imitatin Game Suggests majr cmpnents required fr AI: - knwledge representatin - reasning, - language/image understanding, - learning * Questin: is it imprtant that an intelligent system act like a human?

21 The Origins f AI 1950 Alan Turing s paper, Cmputing Machinery and Intelligence, described what is nw called The Turing Test. Turing predicted that in abut fifty years "an average interrgatr will nt have mre than a 70 percent chance f making the right identificatin after five minutes f questining" Newell and Simn predicted that "Within ten years a cmputer will be the wrld's chess champin."

22 AI system Turing test Experimenter Cntrl

Eliza, 1965 Patient: Yu are like my father in sme ways. Dctr: What resemblance d yu see? Patient : Yu are nt very aggressive. Dctr : What makes yu think I am nt very aggressive?

24 Eliza, 1965 Patient: Yu are like my father in sme ways. Dctr: What resemblance d yu see? Patient : Yu are nt very aggressive. Dctr : What makes yu think I am nt very aggressive? Patient : Yu dn t argue with me. Dctr : Why d yu think I dn t argue with yu? Patient : Yu are afraid f me. Dctr : Des it please yu t believe I am afraid f yu? Patient : My father is afraid f everybdy. Dctr : What else cmes t mind when yu think f yur father? Patient : Bullies.

26 The Chinese Rm She des nt knw Chinese Chinese Writing is given t the persn Set f rules, in English, fr transfrming phrases Crrect Respnses

27 The Chinese Rm S imagine an individual is lcked in a rm and given a batch f Chinese writing. The persn lcked in the rm des nt understand Chinese. Next he is given mre Chinese writing and a set f rules (in English which he understands) n hw t cllate the first set f Chinese characters with the secnd set f Chinese characters. Suppse the persn gets s gd at manipulating the Chinese symbls and the rules are s gd, that t thse utside the rm it appears that the persn understands Chinese. Searle's pint is that, he desn't really understand Chinese, it really nly fllwing a set f rules. Fllwing this argument, a cmputer culd never be truly intelligent, it is nly manipulating symbls that it really desn't understand the semantic cntext.

28 The Chinese Rm She des nt knw Chinese Chinese Writing is given t the persn Set f rules, in English, fr transfrming phrases Crrect Respnses

29 Newell and Simn Predictin In 1997, Deep Blue beat Gary Kasparv.

30 Pints Ratings Can Cmputers beat Humans at Chess? Chess Playing is a classic AI prblem well-defined prblem very cmplex: difficult fr humans t play well Human Wrld Champin Deep Blue Deep Thught Cnclusin: YES: tday s cmputers can beat even the best human Ratings Deep Blue defeated the reigning wrld chess champin Garry Kasparv in 1997

t play well 3000 2800 2600 2400 2200 2000 1800 1600 1400 1200 Garry Kasparv (current Wrld

31 Pints Ratings Can Cmputers play Humans at Chess? Chess Playing is a classic AI prblem well-defined prblem very cmplex: difficult fr humans t play well Garry Kasparv (current Wrld Champin) Deep Thught Deep Blue Ratings Cnclusin: YES: tday s cmputers can beat even the best human

32 Can Cmputers Talk? This is knwn as speech synthesis translate text t phnetic frm e.g., fictitius -> fik-tish-es use prnunciatin rules t map phnemes t actual sund e.g., tish -> sequence f basic audi sunds Difficulties sunds made by this lkup apprach sund unnatural sunds are nt independent e.g., act and actin mdern systems (e.g., at AT&T) can handle this pretty well a harder prblem is emphasis, emtin, etc humans understand what they are saying machines dn t: s they sund unnatural Cnclusin: NO, fr cmplete sentences YES, fr individual wrds

33 Can Cmputers Recgnize Speech? Speech Recgnitin: mapping sunds frm a micrphne int a list f wrds classic prblem in AI, very difficult Recgnizing single wrds frm a small vcabulary systems can d this with high accuracy (rder f 99%) e.g., directry inquiries limited vcabulary (area cdes, city names) cmputer tries t recgnize yu first, if unsuccessful hands yu ver t a human peratr saves millins f dllars a year fr the phne cmpanies

Recgnizing human speech (ctd.) Recgnizing nrmal speech is much mre difficult speech is cntinuus: where are the bundaries between wrds? e.g., Jhn s car has a flat tire large vcabularies can be many thusands f pssible wrds we can use cntext t help figure ut what smene said e.

34 Recgnizing human speech (ctd.) Recgnizing nrmal speech is much mre difficult speech is cntinuus: where are the bundaries between wrds? e.g., Jhn s car has a flat tire large vcabularies can be many thusands f pssible wrds we can use cntext t help figure ut what smene said e.g., hypthesize and test try telling a waiter in a restaurant: I wuld like sme dream and sugar in my cffee backgrund nise, ther speakers, accents, clds, etc n nrmal speech, mdern systems are nly abut 60-70% accurate Cnclusin: NO, nrmal speech is t cmplex t accurately recgnize YES, fr restricted prblems (small vcabulary, single speaker)

35 Can Cmputers Learn and Adapt? Learning and Adaptatin cnsider a cmputer learning t drive n the freeway we culd cde lts f rules abut what t d and/r we culd have it learn frm experience Darpa s Grand Challenge. Stanfrd s Stanley drve 150 withut supervisin in the Majve dessert machine learning allws cmputers t learn t d things withut explicit prgramming Cnclusin: YES, cmputers can learn and adapt, when presented with infrmatin in the apprpriate way

37 Can Cmputers see? Recgnitin v. Understanding (like Speech) Recgnitin and Understanding f Objects in a scene lk arund this rm yu can effrtlessly recgnize bjects human brain can map 2d visual image t 3d map Why is visual recgnitin a hard prblem? Cnclusin: mstly NO: cmputers can nly see certain types f bjects under limited circumstances YES fr certain cnstrained prblems (e.g., face recgnitin)

38 Can Cmputers plan and make decisins? Intelligence invlves slving prblems and making decisins and plans e.g., yu want t visit yur cusin in Bstn yu need t decide n dates, flights yu need t get t the airprt, etc invlves a sequence f decisins, plans, and actins What makes planning hard? the wrld is nt predictable: yur flight is canceled r there s a backup n the 405 there is a ptentially huge number f details d yu cnsider all flights? all dates? n: cmmnsense cnstrains yur slutins AI systems are nly successful in cnstrained planning prblems Cnclusin: NO, real-wrld planning and decisinmaking is still beynd the capabilities f mdern cmputers exceptin: very well-defined, cnstrained prblems: missin planning fr satelites.

39 Summary f State f AI Systems in Practice Speech synthesis, recgnitin and understanding very useful fr limited vcabulary applicatins uncnstrained speech understanding is still t hard Cmputer visin wrks fr cnstrained prblems (hand-written zip-cdes) understanding real-wrld, natural scenes is still t hard Learning adaptive systems are used in many applicatins: have their limits Planning and Reasning nly wrks fr cnstrained prblems: e.g., chess real-wrld is t cmplex fr general systems Overall: many cmpnents f intelligent systems are dable there are many interesting research prblems remaining

40 Intelligent Systems in Yur Everyday Life Pst Office autmatic address recgnitin and srting f mail Banks autmatic check readers, signature verificatin systems autmated lan applicatin classificatin Telephne Cmpanies autmatic vice recgnitin fr directry inquiries Credit Card Cmpanies autmated fraud detectin Cmputer Cmpanies autmated diagnsis fr help-desk applicatins Netflix: mvie recmmendatin Ggle: Search Technlgy

41 AI Applicatins: Cnsumer Marketing Have yu ever used any kind f credit/atm/stre card while shpping? if s, yu have very likely been input t an AI algrithm All f this infrmatin is recrded digitally Cmpanies like Nielsen gather this infrmatin weekly and search fr patterns general changes in cnsumer behavir tracking respnses t new prducts identifying custmer segments: targeted marketing, e.g., they find ut that cnsumers with sprts cars wh buy textbks respnd well t ffers f new credit cards. Currently a very ht area in marketing Hw d they d this? Algrithms ( data mining ) search data fr patterns based n mathematical theries f learning cmpletely impractical t d manually

42 AI Applicatins: Identificatin Technlgies ID cards e.g., ATM cards can be a nuisance and security risk: cards can be lst, stlen, passwrds frgtten, etc Bimetric Identificatin walk up t a lcked dr camera fingerprint device micrphne iris scan cmputer uses yur bimetric signature fr identificatin face, eyes, fingerprints, vice pattern, iris pattern

43 AI Applicatins: Predicting the Stck Market Value f the Stck?? The Predictin Prblem time in days given the past, predict the future very difficult prblem! we can use learning algrithms t learn a predictive mdel frm histrical data prb(increase at day t+1 values at day t, t-1,t-2...,t-k) such mdels are rutinely used by banks and financial traders t manage prtflis wrth millins f dllars

44 AI-Applicatins: Machine Translatin Language prblems in internatinal business e.g., at a meeting f Japanese, Krean, Vietnamese and Swedish investrs, n cmmn language r: yu are shipping yur sftware manuals t 127 cuntries slutin; hire translatrs t translate wuld be much cheaper if a machine culd d this! Hw hard is autmated translatin very difficult! e.g., English t Russian The spirit is willing but the flesh is weak (English) the vdka is gd but the meat is rtten (Russian) nt nly must the wrds be translated, but their meaning als! Nnetheless... cmmercial systems can d a lt f the wrk very well (e.g., restricted vcabularies in sftware dcumentatin) algrithms which cmbine dictinaries, grammar mdels, etc. see fr example babelfish.altavista.cm

45 The agenda f AI class: 1. Fuzzy lgic 2. Prepsitinal lgic prlg expert systems with inference algrithms 3. Rugh set thery 4. Decisin trees, knn, Naive Bayes 5. Neural netwrk

Introduction to Artificial Intelligence

Introduction to Artificial Intelligence By Budditha Hettige Sources: Based on An Introduction to Multi-agent Systems by Michael Wooldridge, John Wiley & Sons, 2002 Artificial Intelligence A Modern Approach,