Cusaitan) Yd» Machine Lear nin 9g eis (Machine Learning) Learning problems : perspectives g Mssues - concept hearning = Vetbian Spaces & candidab @liminations- Frductive Bras — Cectsion “Tree hearning a Representations = Algostithrn ~ Heuststtc Space Search,@ ' om => A computer Program us Said to Learn fr Y Some class 4 Experience “E° with veSpect to 2 Y a T m ‘p th (tS Tasks 1” and Performance ea sore 2 h Performance at tasks to T, a& measuted by 'P’ , tmpstoves usith Experience E. => ue must identify these features # the class dg +ask3 * Re measure Ppertormance to be Tonpsioved * the source 4 Expertence . O checkers Learning problem + Task T: playing checkers * Petbormance measuve pi percent $ games a5en eet ©pponents Taaining Experience Ge) :- playing paacetce games against ftseip. Bandon Tecrgnition Le a ning Problem: * Task Tt Tecognising and Classifying hand- ~Gritten usosids ith In Tages Pettermance moeasure p+ percent o& words ee recely class’Pied° Teratng Ex pere A database ds handenritten wosd, wiih § given classi fications A Robot Brit ving Learning problem ask TP OHettving on Public fovr-bane highwags USing Version Sensox4 Rerbormance measure p: Average distance traveled before an cxror e Cas judged by human Overseer) “raining expertence! e+ A Sequence dy mages and Steering Command, Seconded ashile obsesving a human driver. fical estan 4 ‘the checkers Learning pragram Experiment New pr0blem Generator 2 CEattial game Hypothesis board) ) Sotution hace Thaining examples Cgame history) Cebit Main lO) > Abo , Viva (Bs > 5" aahSe ‘ ial f Rerspectives and fssves i machioe Learn a 129 ape => One uSeful Perspective On machine Learning us that tt tovoltues Bearching a very Large Space o possible hypothesis +0 determine One that best fits the obsesved data and GnGe PIES” Knowledge held by the Learner Pat “= Gonsider the Space 4 hypotheses that could is in Princcple be Ovtput by the @bove checker Kearner. thts hypothests space Consisk % 2! Cvaluation functions that can be ye presented by Some chotce og values for the weights Wo through U6. => The learner'S task uu thus +0 Search tbvough this vast Space t0 docate the hyPothes: eo that ds most Consistent With the available training examples. = The LMs algost tho for fitting weights achieves this goa! by \tewatively toning the wcetghis , adding a Correction to each coetght @ach -+time the hy pothesiged Evaluation Funckon predicts a value tha disfert from he training Wales. att defioes a Conti nucsly Parameterized Space & Potential hy po thesisThis pere i i : S$ perspective 4; learning a& a Search 1 * P9toblem in order +0 Characterize Learning Methods by thety Search Strategies and by the underlying Structure o the Search Space *y they explore — we Can analyge the velationship between the Size a the hypothesis Space to be Searched, the no. training examples available, and the confidence. => we Can have thet a hypo thesis consistent worth the raining data will correctly generalise to unseen examples Qesues % Machine Learning algositthrs exctst Por Learntog gpreral — uohat Larget Punctom% from Specific raining byampies = on ephad | Settings coll particwar Gesired function, abgort thn Coverage +o tre wen Sufpicient Eraining dat perform best for eohich tyPes oh Peoblemy and wepresenta tions 3 ald coltch algor thy => How much training data ds Sufficient? corat general bounds can be found to ¥elat the Confidence es Learned Yypothesis to the Amount troining experionce ese? es novan ceremonies S' Qn ae: : A the character d the Learners 6 hypotheses espace ? => when and how can Prior knotuledge held by the Learner gutde the procem ay generaltaing fom @tamples’ Can prior Knowledge be helpful even when tt % Only Approximately Covyeck? =} What ds the best Strategy Por choosing i how does ity a useful mext training experience - an the chotce qd, this Strateqy alter the complex a the pearing. pxoblem 2 =} what “us ‘the best way +o meduce the Learniag task +0 one ©») move Funct on approximation pxoblems? put another “way , cohat Speci fic functions Should the System te attempt to learn? Can thts psocess itself be automated? — Hou can the Learner automatically alter it8 sepsesentation to impsove tts ability 40 Stepresent and Learn the tayget function?Concept Learnic 4 q LnPerttng a boolean- Valued function from chaining examples dy tté input and output: A concept hearning Tasks : qj: ae -S congider the example +aSk 4 Learning Fhe target Concept: = she given tables ceprooents by a Set ab atti butes = the attribute Enjoy sport tndi cates ushether ® fo») Not Aldo enjoys his fovesib oater Spomt on thts day. =s the +aak, a to learn +0 predict the | Value ds Enjoysport for an arbitary day, fpaSed on the values 45 at$ Other attibuty e = let us begin by Considering a Stmple representation ia which each hypothesis ets dy a Conjunction co Conatraintt consi attributes . on the instance => let each hypothesis be a Vector d Six Constraints + Specifying the value o the Six attributes Sky » Ainjemp, Humidity , wind » water, fovea!4+ Qndicate by a "fT that ony volue up 9 acceptable for this attribute. > Spect fy a Single sequired Value Cogiteaims Por the attribute, ©y yx tadicate by "g’ that no value wt acupiats => rn Bome tmtance “x Satisfies all the Conshaints od hypothesis hs then h classifies Pa’ an the Posi tive. => the hypo thesis that Aldo” enjoys hrs faves Sposit only on cold dlauys udith high humidity ssion ws sepsesented by the express . C%rcold> high » %+252) Jumid Te has Humidity | tlted | avatar fae Enjoy Strong] Warm yes Strong) Wate) Same yes Strong] WOT) chagel pig Strong | worm Cherngel Yes rgaie Positive and negaaive straining examples For the target Concept Enjoy sport=> The Most Jeneral “hypothesis - that everyday du a positive example - us sepsesented by CTs) => the most SpectPic possible hypothesis — that no day ds a positive exampie- cs sepsesenitd by CB $99.96). => Sn general any concept fearing ta&k Con be described by the Set % instances @ver which the target function ws defined, the t function » the Set 4 canaidate targe the Learner, hypothesis Considesed by and the et 4 available Haining exomply Alotattan @n general »c can be any beolean- valued function defined ever the smstances x 5 thot WA €: x >of. => the target Concopt Cowesponds +o the Nalue ¢ a attribute Enjoy spoxt” Cie, ccx=t tf Enjoyspost ~ yes > and ccx) = 0 th Enfoyspot = No.)nae ve e , yctive learning Hypo thy Any hypothesis found 40 approximate the a Sufficient Large ast also approximate they unobsesved target un ction well over U Set + 4r1aining examples | he target function well over examples Concept learning as Search > Corcepi learning can be viewed an the task Searching through a Large gpace 4 hypothests — (mplietty depined by the hypothesis BepseSentation important +0 note that by Selecting => Tt os the designer Q gpa eres wepresentati on, Learning algostthm amplicitly defines q, the the Space 4; all hypotnesis that Ene psogram con ever represent @ theresove can ever leayo => Consider» fhe example » the pnerantes "yond hypothesis "Hy" wn the Enjoysport learning ark. ~ = Given that the attribute "Sky has +tinee possible values hye r ” “Aiviemp" 0 Homily Cach have Two \ w 0” hod, water é O.SSi bl de iG) por=> The instance Space "y¥’ containg exactly 3.9.9.9.%.9 = 96 adbtinet ingionces. = A similar calculation Shows tat there ate 5.y.U.4.4.y= lao Syntaceically dPAticc! hypothesis wot4hio ae => Learning os a Bearch pasoblem , then tE Us natural thot owt Study 4 learning algosithm will examine different strategies Pov Searching the hypothesis Space - General ~ £0 - Specific ordering of hypotheses —_ Many algo t thes for Concept Learning orgariae| the. Search Ehrough the hypothest S Space by relying on o very useful Structure that exists Por Any Conapt learning — Peoblem. wy TO ‘illustrate the general - to~ Speci fic ondering, Consider the -wo hypotheses hi = {Bonny, 9,7, Str9ng, 224 ho= 8h Sunny 3.99.97 5 Def (er hy and hy be~ boolean— valued functions defined over x. Then hy ds mone general than -(©1) equal tO hy Curitten hy > hx)@ i4 and only “fe — (M4, € ¥) [Theo =") iat ( hyo = o]} Gastances *y’ Hypotheses H Speuif x1 = he % = i a= < Sunny» worms high, ha = < Sunny 99275 leght 3 warm, 8ames ha = the Limitation $ £ addvre Sses Several H the = the Key (dea in the candidate —eliminatio a olescription % the gon thm aw to ovipat consistent usith the | Gok oF jolleengpoupeses training ex ample: 4 this descré ptio? dy tl —s TAIS Computes the enumerating all Ser ith out explt city members 5 "fin Can be, applied "ito "Peele Such a leaning megularities tn chemi cod mass Spectroscopy cmitchell 194% =, hearning cmtso| mules fod heuristic Search C mitchell 1932 Representa tion* ( deft nition) A “ho hypothests “h’ ds consistent twtth @ Set od training examples "D’ th and only ik heads ccx) Por each example (x, 60x) tao Consistent (hi 0) = (va, etx) & p) heros ce) |é . | I , Lythe Vesston Space , denoted % va othe wD Yespoct to hypothesis Space "WY and training Jexampies "S “H cansisient | J=> The Subset 4 hypotheses from With the training examples 40.70 VS yo {he HI consistent (hid) 4 alaoxithm oe oe the List -athen= eliminate ln \gowi thm fisot ce to contain all 5 any hypothesis aining example AP the est - Thon — elimioate a tortialiges the vession 8pa hypotheses “Hy” y then ejiminate found yaconststent —Uotth any tr aulgosn then 1} Vergion Space Ce 0. lest, containing every hypothesis in “H 2+ for each traicing example » (x) ¢) remove from VertionSpace OY hypothests “Ko Ror which hex C6xd 3 output the List 4 hypothests [0 Veraton Space.St 5 x Tesunng, Bhd 1>> 22, warm, 1:4 > Ae £19> Version space with § Specific bourdary sets Definker the general bourclary &, with vespect to hypothesis Space -H ond training data 0, uw the St F maxi mally general members ad H Consistent with Oo Ge $qeu| @nsistent (4:0) A (mage) Tota, 9JA Comsiatent( g's0)} } 3wapmuion Te oe Hepinition + the Specific boundary 5, with syegpect 46 hypothesis Space UW ara graining ws ithe Set 4 minienally genera | tH) Comiz bet data 0D, (ie. mari mally Sporigic) members A, | ustth ‘D S= {Seu Comsistent (5.0) A CaS! € HD TiGsas men Gonsistent (80) } vepre sent aston thediare tary set 4 stance booloan - valued Verxston Space het >x be an arbi bo a Set & | ord let rs mix let rx font hypotheses defined ove Concept defined over X> be an anbitary target tary Bet 4 raining Mary jet ep bec arbi examples fsx eens } + H/C and for all x ph) idg S G Gau, saa! cle fine al . 507 {he HVC © 6 C4 66) (g=4h 2383| elimination algorithm Compal, all hypothe srs an — the candidat containing the Version Space f comsiétent from cH’ that are woth sequence + training onanples: mnipialiging the vension Spare observed thet Us <7 Jt bas by 4 the set H aul by gartiali 3f04 the “G boundary containing the >» and enctleltaing the °S ae get ap entato the ‘most Spect fi ¢ Cleaat gree) hypotrest* So” $0d9 FP )}Candidate elimination algaaithm * | Candidate - @lioviaati ONGORU i 4o the set 4 maximally general . Gnitialige Gi io A hypothe ges —> Qnitialige ‘si to the Set o& maximally gpeci bic hypotheses in => for example glint e th d ts a positive example * Remove from “Gi any hypo thesis socomistent with “d’ * For each hypothesis “s' in S Hot ds snot consistent with ‘d e Romove s from S * Add to S all minimal general hs Such that “KR us (emsrsied with d! > and sone rgattons member a. Gi Gs more general er, @ Remove from iS any hypothesis that us more genera) tran another hypothesis to's a iP d Bw a negative example ° Remove from S$ any hypothesis 20 Cora with @ | * For each hypothesis gto Gt thad aa | Mot @misten wth ol+ Remove @ from Gr | + Add to “G all anini mal Speer igati ond { h 4 4 Such that - © fh us Gomidstent with 4d, and Some member dy G wy more Specific than h | | thod 34 | a} G « Remove grom och any hypothesi4 frypo thes # less grec than another Training example So* warm, f: Sy Sa° Go GIeOa te AS Tet Wt ot Zan paling euample Z Sunny swarms wormal » Sttong» warm, Same >, enjoy sport = YES ZB8uany, Lam , Htgh, Strong » WaIm, Same> » a | _ Eojsoy Sport= Yes@ é.\ |§- Sunny , Warm, 9 strong? < Sunn y2,2,St100g,.7 7 < Sonny swarms 27229 > BMY : motes in < 4g, , waren 5 9, »St70g» 2-29 1 VAS 1€Y, warm, 51;2,D} £4: the -firal veraton pace for ‘the Enjoy Sport Conapt warning pxoblern and’ training exanpls desuribed earlier. Prductive Bias* — the Candidate - elimination algontthm will Coverage towara the “tue target Concept tt ds given acwiak training puourded examples and provided 44S Initial hypothos' mw the rere Concept. Space contaiA Biased tlypo thesis pace — Consider again the Enjoyspoxt example in which we restricted the hypothests Space +0 faclude only Conjune Hons d attribute Values. Bomause qh this Mestriction » the hypothest # Spore ws unable to smpresent even Bienple dtsjanctive target Concepts Such os * sky = Sunny @y) Sky = cloudy". ag Example} Sk Aig 4 Humid enter | fore Y | wind + Emoy ' Sv0 Fe PY | warm | osmal | St09 |cool Saal yes » Cloudy | usar | wormal | strong | cool | charge} Yer 3 Ratny warm | norm) | Strong | Col change AO. oes => Consistent with the fiTét +00 example and sepptesontable th the ge hypothesis space H TS Sot <1 -warm , Normal , Strang, cool, charge >| An Unbtased Learner, + ye cs Capable dg; otepresenting every Possible Bici & te anstanes x+ In general » the Set ch atl Subsets 4 2 Set x sae power Set 4 *- SSNs 7-7.92,7>V < Cloudy, 259,2,9,7>| the paailtty d Bias free learning @ pehinvior | Consider a Concept learning al gost the ‘LL Sor the a Af Tene, | eens Cr beec AOR concept | defined over x, and let De ~ {dream >} bo an Fasbitary Set a4 training examples qc. fet LG)» Oc) | cbnok the classification assigned 10 the tostane |X, by L after training on the pata De: —> The inductive bi 4b ds ang roinimel Set dy aseertions “8 Such tab for ang target Concept |-¢ and Ccoxstesponding training examples De C¥ ape xD PBA AT TEL Gia) Gnductive bias -4 candidat - eliminator algosiithm The target Concept “cds Gntained tn the giver hupottesi¢ Space “H. ; +> the inductive Candidah elimination algo’ ihm at the top 4 the figure takes two Ups© Grductive Systero Naining ertampla} Candidate Eliaination Classt fication Pageant ds new New instance 4instancecor) USing Ay pothesrs Space H " don’t know” Equivalent deductive Systm Traiaing Examples New instance | Classi fication 4 new instance, Assertion (or) "Onn know! ” y wntaim the tagpt cencept” £4 modeling ‘inductive Systems by equiva Deduc tive Systerns Decision Tee learning SDectsion tree fearning th o method for approxicatig dicrets Valued target functions, ta eahich the Learned -function® tu sepresented by a cecisio? tree. = learned trees Gn ako be cye- vepresenkd as Set g& ib-then yale bo 1psore human mea dability,Dectsion Tree yesentatian ) Hs Decision trees classizy instances by Sovting © them down the tree Prom the soot +O Sore leak node, udhich provides the classi fication qd the instance. Ks Each node (a the -tree Spec'fia a test 4 Same atttibute 4 the instance , and each branch descending Om +that node corresponds, to one 4 the possible Values Por. thir attribute iain! Sonny osha yes (Horst) htgh oral SHonpia ta Ped | J 0 i. 3 No yes £19 + Decteton Wee for the Cnieegh ply Tennt 3Fox the given ( Corvosponds 40 ( ovtlook = Sunny A V( Outlook = VC owtloor = Rato Lor => Most atgont thins fox learning decision trees thm thos employs & Top- eSpace % on a core algosi - down, greedy possible Oectsion talbtch Adtri bute ae = we will define @ Co \lbda tnformat(on how well a given tyalmng examples cla Si fication =+ (D3 vses this taformation garn measurex +o Select among the candidak attributes at each step While growing the tree. erample, ihe Search through th expression Houmidtty = Normal ) overcast J A wlind = weak) ‘ aece Learning that have been developed “Trees. the best classifier ? grartstical Poopenty y qain > that measures attribute Separates the ing to their target acord the doertsion tee af ther axe variatiom%En tyof eaauyes homoqenattt 4 Examples =3Jn order to desing toformation in prectselys we begin by defining 9 measure commonly ErOpY, USed fo to formation theory , called Entropy . an that chata cteviges the Cim) puTlty a arbitary Collection 4 examples - = Given a collection S; Containing positive, and mega tive example 8orne target Concept. The Entropy 4 °S Sielative to this boolean Classification ds Eroops.cs) = — faleg.to ~ Bloale example Suppose °S' ca a Collection 4 ty example 4 Some boolean Conupt, including q positive & 5 repoie Entropy (a4, 5-] = PR) ty) - Sly, (540 Entropy = 440 LTOs Cexam ples , Target ~attribuk , Attribuda) ¥% CyYCalie a Root Node for the yee ¥% Uh all Examples ave Positive » Retu7 the Single - node tree Root, wrtth label = + * ib all Exampies are Tegative , Reluro the Bingle -node tree Root , with label = - » Yh Attnbutes ek empty, Retusn the Single: = most comrno) ta Exampes: node tree Root, with Label Value “rorget - attribute * Otherualke Begin © A < The attribute from Attributes tha beet” classifies Exarnples * the dectsion attribute fox Root . mpl Target a pute a t ‘oudes (Ay *End * Rewin Root JA yorthm ad Summary t Ha apect u & Go2 5 Me Sf I 1_te ae Geni ; loan ~ Valusd unc oo 77 Of Oo Po to & fig the Entvopy function yelative boolean classitj cate? ar the pre example waved betwaro FO! por tion Pay » PositiveGain CS, A) = Entropy(S) — Da Fc} entops(S Hypetheses Space Sxuch fn Dect ston tees => %p3 Gn & Choractetiged a Searching = @ Space 4 hypothests for one that fits the training examples: =— 7 the hypothests Space Searched by G03 ds the Set d possible Dec’ston trees . Gnductive Bias to decision teee learning Approumat *qductive bias 4 203% Shontes trees axe cre persed over larger trees i Qssua in Ceci sion T¥ee Leasning- x Avoiding Over Fttti 09. the Oala * Ruduced Ess08 peuntag % Rule post- psuning x Ancorporating continuous - valued Atte bale oy Alzernative Mmeasyses Sos electing Attributes # Hand! eng +rratning Example with missing @rttrtbute Valuet- ee. Pe tort Cudering