You are on page 1of 31
NIT- 1 ; = SNTRODUCTION - Well Defined lia Designirg 0 Leomirg System, Qaues th Mochine Leoyniy The Concept Learning task = General do speufic ovduving of-hypottesie, Find-S. List athon ebiminattr algorithm , uandidate ebininadion algoritten , Wduchive Bias. “0 Nochine Learning . - Hor raid do be subset of auhiPiiat urdettigence. ~ H ub a fied of AT trod ablows systems to Leow and improve from exPovionce without bering expbio'tly progrormmed. de: Npchine leorning enables a mathine vty automat leom from data, dinbrove Performance from experiences ond predict chaos without Being erpuicituy progrommed- Well Posed Learning Froblem : DeBinition ; cA stomPaton Rogrom ab Acid do Leorn from exbevenn's’ with enect +o sorne class of darks T ond performance measwu P, JE dts performance o Ons fn T, a» measured by P, improves with expouience €. Qn General wall defined learning froblems ; we identify the clam of dark, cthe meanwu of porforrrane Checkers Leorni P eiTask T 2 Playing checkers © RerPrmonce Meas P : percent of games won agai opponents - . Training expovonce E+ playing practice gome Htedlf. A. Handtorlting recognition Leorning Problem: eTaak T : Fecqgnizing and claysi fying Hordern words within tages. + Performance measwu P> percent of words eorrectyy clansiPied.- . Foining expowence E> a dujehesc aes o data boar of handwritten words with giver clamsiptcadions - A Robot Oriiving Leorning Problem: Task T : Ortving on public four lane Highooays using vision + Periformance Mearurt’ P's average distance brovlled before an error Co qudged By Human overseer ) sTrainteg ExpouenaE’; a sequence of dmmggee and. steering emmands Treeor oile observing a Human driver. SerMpors - of deorning Aystom + Lela chiscuss some opie ovning Jn wompulov machine learning ? : 7 = reladiomhibs ond xorrdasions con be hidden within lange amount of data . Machine Leaming / Datta, Mining may Be able th Pind ctheat relationshify - (Dats. Analysince Vistabizabion) need of automation, amount of Knowledge available about cevuin dark might be too ange for Humons Ceg. medical dignosh'c ) Solving comfatex probloms and dec'rion malig. Rabid Snevement un thu production of datie 1 1 » How Learning un d¥ffeunt un wombuteu wrt Human ’- — machine bared knowledge fan exceeds thr Human brain as fan ax understands’ » memorizing ond cepmbrehedin Knavledy« ub oat 2 “fi 4 = leorning oh womButer ur done By explicit progrommirg © 30. Fry wan not reBtoce Human. beings. (mingmum Human, dntewention) - learning din. eomputtr. ub not Siared 5 while fumens Pave tha tendenvy of diarness — leorning un tomButin ub sdf oganiung ard adephie- « Types of LEARNING. (A) Superviloed Leomigg: + Thu ag class of learning trod uoes a model Jo deoon sth mabbing uo infu and daryer variables. + Sh on. gasuforvteed leoming problem the modet crite to Learn. Qy HHseP and Beengnize patterns ond exctract othe reladinnhiB omong dint data, © thus ub no subrvuision or teothun a drive atin modal /Leorming proceas- + Goal ip da find underlying pattonns fa thy. doy . Wainipg exomple / Oiti se sod Js unlabeled Tyba Cuatodng SN Aisotiastion 3) Reenfreernuat earning ¢ oH dha type A prsBlem; cohun thuve do a agent RE frderods ith Environment and revs Sthor rewards punishment (rewards cam Be ther positive or megabive) + thou aw no draining daty sets in cthis Leornipg . The System sith deoan By its oton exbouerce. Bi A Robohe Dog: which Learn to move ams on its oow A Méthine Leorntng system wit Learn fiom ‘Alatoric data. , Builds tthe Prediction models, and whenever it recloven new dota, , predicts ste 2 St con Learn form tht past dette ¢ improve 3 St oh a date driven technolagy 4. Hub mu similarto date mining an Tt aloo deals with fuge amount 6 Data, - eo ee vnadicably Now; Lets comeTo tu Phares f Machine Leaorning systero Design There ae vesrious stages dn th machine Learning procera: - Stoge 1: Data Letleetion and Intperations —this up chi first adage Whou the dita un eeected ond Untpgrded. ~ data wetlected acts an fnput ot ctu model ond Unpa au tle fu Gy Data Preprocessing = cleoning date may Be required; that Jncludes Ternoving dubliante Correcting errots ,ded with missing value , nomalizadon ond data. lyfar enveroion © Uoheather dtu training experionce provides direct or Undine Peedhack Teganding an made oy porformance system./ Learner . for example dn chedsens game: tthe System might Lear from. Direed Trafning examples |. stonaishing of hdividual checkou foand Makes ond athe ceorrert move for each. stomaisting & Move sequences and firal “The Dyes 4 cohich earner. tontrals thu sequence of training ne i he joe, * chu Learner might debends on cthe teathin to seleck Board tater ond provide ithe correct move for each. * The Learner might itself propose, Board states and ark tthe tedther for the corned move. * The leowmenr Rave uemblez eontrot oven Board states ond Ghdfrect) training tlaxificahons, as it dows open tt Learns by playing itself. @ How axl Jt represents exomple sky Ahi TemB Humidity wind) water} foreaant| Exjoy shot 1 Serey | Wasim Normal | Sheng | wanm |tome yes 2 Sunny | wan High Shong |danm | Sore Nes 3 Rainy | coid Hag Shreng | Wanm [change | nto 4 Stony Wanm Huge Shong evd, Change Yes ia] Hew daak, ub tm “ Jearm tt predict the valur of ctu enjoy sport for on arbitony day (new uhwtance of the doy) » Saned on ting values of it attybute 4b. Giron example . eh porsiBle :- so tien pte for ahs WOE cule ax nat going to deocR alk xthear oncefats ito machine on it wht he eraRuatie ond not frasible ; thoufere onyadthe conaintent eoncefsts- cou need tobGcRovsanvs 7a Leatning ar o task : » Let enc Bypothenis be a vector of sic weordyaints . specifying tm valu. of six athidutes ; Sky » Arr temB . Humidity, wind, walin ond Forecnat » For cock attiButi, the Aypothesis witt eithur Be ; 1) Undicaded 8y 69” which means that any value un acceptable fos Dhak othiButt: il) shdicased By "0" wich means Dhak no value dd aceefatabte iil) SPeerfy a single required value ; cither sunny or Tainy for as = BE he Porson Cy Ris favourite sports eat = Hef Rormtaibyea etait by bt oe es C2, told, High. 2.2, 2d Notations: - 42 he set of items oven WAI cthe ewrceht uo defined ula called. Set of Jratonces and gonvrally given Sy a4; when d up cthe not of feature (in care iP coc Beciae ox - us gate cane eh uh gonial 5 dit do given 6 smultipulcadion 6f noe of valuus & for Oe atts bait. BY the cemeebt or function to Be Jeoted. cb called the danger tencept , whic ud denoted by Cm “5X = fo, 13 “(x)=i vf enyoysport "ves ~ positive exomBle £G)= 0 uf eryousPost = "NO" ~ negative 3Y tne gymbad H denotes cthe ack of all passible Aypothenis Bhat cthe deamner may xensiden. regarcling the Adontity of athe

nggabie 203 a. ¢ xcemmon and retorife the pomie values W fis Temp Humidity wind walt, —ferecoat a Saas See OULD sy 9 = 486 oe { Dade ™ * General do speriBe owdoring of Mupottenis: Considur ithe tao Rybothents Ae = (Sunny 2. 7, Sheng, 2,2) Ra = CSunny, 2, ene) =I Fy pothesis tlami fies a utvatonce positive asin athe above tepbotrens Agwill Uaraity more uinatar Qn positive (09 You tom gust a Hat any A positive By fs win, “Ar can never be move bran Rij a it lanai fies dew Hatonces ap paattive thon dhe Ry. Ddefition: “Let AZ ond Ry bec Then 83 B move general 4P and onby Gp —— ol f cto ret of X of. alk diate si : Pee ae m ' ee) storreap ord fo some gubsek aP x (dhe subset op drctarces rat ot clanifies positive ) ~The onows seonnedtiog Rybothals refrenents athe =mure- - than Teed, with the arrow Rotating towards he on jou gata Aypotbenis. NOTES the subset of stivatances Ranocterized by Ry subsumes the subsdt Ranactenized by Ra. Bence fig ih move “geneva straw, Re. Fitd= S_ Algoritmns. Fibdipg a moximally aperifirc Rybothesia :- Steps 1. Gritialize A ty tne most .aperiPic Aypothenis in H. 2% For eoch positive brainieg Saatonce x Fox cock attributt conctraint a; sv fy GP whe wonatrairt a ub satisfied ty © Then do_ noth fe mplan ay dn fA ty cthe next move genal conatraint Satisfied, ay x 5 oa “Pigt uctep of FINO-$ ub to JfiHaline A op othe mext Shee PreKypothests Ja A: R-0G% % 0 0 O O) * Consfden. one Ficgt training exomBte Xa = ¢ Sunny, wom, Normat, strong, wonrn, Same >. + We con aledy see Hod Pupbothesis A. uo to aBeeiPic . None ef the ‘D’ conabalits sR AR wu satisfied by this example, s0 ent is replaced. by othe next mat general eevchrumt that fits the example. y undeurtording: we tam o£e the atep- 2 of dhe algosittrm aa’ v “e For eae eveltomp for cocB abbibits Apr RypoHicinvalue) > Ygnroe. a u Prplaxeeinetiw meet poral MypotRuals, Shou fre 5 Ra = < sumy, worm, Nonrel, Strong. warm, same y + Gngiden sthe next straining exomble,: (24 tromble) as Sunny, warm, High: . Hhrmg, worm same 7, + stompo. th with the preaviour fypethuais As. and gowwlizz it Fe ‘ fe : ven an ning ¢ ran ple A A Sunny, worm, high. shrong » cont, rary a | Siribanly Jt ext Be gpnerabited to Ry Mas < Sumy, worm. ? Stung cae) output * Gumanteed Jo owlhur the moat SPeei fic hypothesis esithin H Trak ub wenalatent With wnt positive training examples. © dts frat Auypotheois itt olan Be conaiatent with dhe ne examples provided the ovest target eoncefat uo tentained. fr H Ord provided Hae trainiey exomblen au. correct: ~ Yrromuen by find 5;- 1 far the Leaner tenverged ote ttn torrent donge concept 2 9% thy prefer the meat apecific Aypotheois 2 5 Fre the boning examples eewaistort 2 4. Whos uP Www au aoverol reantrnaty Apeetfic wenafatent Aypottesis 2 “F NO THE CAN OT DATE The Key dea din this abgorithrn wb ate oubpub a description of cine set ef all ypothets Konaistort ajith dhe -trainfay examples Refrresonteation: on fb ds eoralatont ith a aah of balnbey examBles D JP and only af A(x)= £(x) for each exorple (x, e(xy) un D.” Conststent (4.0) = (4x, (xy > € D) oREX)= CC) Note difference Bekveen definitions of tonsistert and sadis fies 1 An example x dn waid to sobishy Aypottusis A when (x)= +, Tegardless cf whdhu. x sh positive or negative exompl: of tu derget roncept- Q An exmmble = de sald te be onaistent with Aypotheris fe SEE Al(x)= C(x). DePinition s version Bow :- The version face, denoted Gy USy,> with reapoct do fypothenis apace H ord tainieg example D, ws the subseh oP Aypothesis from H tonaistentt With De taining examples tn D. VSuro = FREH | Comistent (8.0) 5 ( NENA (GroRsT version sBace dd contain Outpud othe Wat of Rypothenis a» veraion abace Nore: © rwwion Bare need ts Be FAKE @ Not feasiBle ; since TH requires exhaustive enumenation (. t. Ihating ) ab Aybothests fn Bote s . (ey Sart Consider a Rypothettca example tn descriBe Roo Ust = Then adj ovithrm GapRC chy. ov. 6 Rybothuris given an:- di, Ae od Aa, fs, He First we Uns thin; ® @® 6 Checkif 81% C(x) thenremove R, ftom eck elae Reb it ® &) aaa + C3) © QO&H ion Al g jovithm,, ' computes the version sBace centainipg al AyPothesis from H strat Gre consistent with an observed training examen. — 3 B considered an extension of” Find - $” agovithmn, astt consider both positive a» uel ar negosive exom fue « for positive exomble Jt cootts dike “Find: $" alyortim while for negative exomBles : f+ i ; i Fen Es Greif. oc Teetae Step4. Qnitaline Gi cto the ret of maximally gerrat Pybotbenia dn H- Snihatize S$ to the rt of madimally sPecifie Rypotresis i H- StebQ for eo trainieg example d Hd positive Ymake spel fie os ines (ae, f%, 003) tar (ue d bd re) Take gonuas do sbeesfic « Gleb3 Owifuk the Bybothesix Ge 3. Exomjole Shy Arle Humidity wind wat forecaat sot aiay ai Sunny Worm Nosrrat Strong Wanm = Same i 2 Sunny Wasim High Bar san te Wig® Strong worm Rogge = NO re For eoR training exaple d (ae 1.) sam gy a ‘Aypothes is Genvrad sto sBeeific un cans af ve exomples ond SPeehfic sty General dn cane of tve exoniple. (b fonder athe Piast ae etomples x1, be) Xu Sumy, axmM, Norma), Shreng, whom, some 7, +Vve Since tt positive in Grutone make jDetoendadie 5; Algorithm and leare 2 Gas Go {unto OE a : Sys , +e ‘gain gts Positive make 3} ic be Gg= Gi § Teran unchanged 3 &= . <2 wom? 222, Steps gate ve} 90 make sPeeifiests gonnat dun chy fast training erornPle > hy “etua) ; wy sumpy. warm . Righ 7 story cool chonge y - +ve Gaz Gz {we will cteat th vin the fir ateB 3 Sq= Gury , worm, 2 Breet ae Owput Gig & S¢ ar firal Pypothenis: Sqe

You might also like