UNIT- I : Toboducton —uell- poted teavning roblems datqning a learning system, perspectives and tssucs ta machine leaming. Concept leaming and the qeneval to cperipic ordering. unhoduction , a concert of learning task, concept teaming as search, find-s: findinga manimally Specific hypotheste, version spaces and the candidake elimination algorithm, Yemearks on vevdion spaces and candidate elimination, inductive bias- Decision Tree leaoning- Tnbrodudion , Actision-tree represent abion appropriate Problems for decieion-hee leaming ithe basic decision tee leaning algorithm, | hypothests Space searchin decision tree leaming- inductive bias tn decision tree leannting testes tn decision hee leaming- | WIELL~ POSED LEARNING: PROBLEMS Deyinition t- A computer program that is caid-to leam from experience & with vespeck to some class of tasks T and performance measure py iF Jit performance at tasks int ar measured by P) improves with experience E+ i To general,to have a well-defined learning problem, we Mur | tdenkity these three features t the class of tasks the measure of [Performance ‘to be (mproved and the source of enperience |A checkers leaming problem : je Tosk T; playing checkers: ts 0 en’ |. performance measure Pt Percentage of oe ae agent “ree 4 atselt . Training Cuperience E+ playing prachice qames aqainut tse}A handyoriting vecognition learning problem : * TAT + yecognizing and clartiyying handwritten words within images spofomance measure Pp! percent of woords convectly dassigied. + Training expoience Era database of handusvitten words with qe dlawikeation. A dbo driving learning problem: “Task Ts Awiving on public four ~ lane highways using vicion sensors *Pafermance measure Pr average distance traveled before an error Cas judged by human overseer) - . Training experience E+ a Sequence of images and Steering commands vetorded while observing a human driver. DESIGNING A LEARNING) SYSTEM +- T qek a successful learning system we need tp have a proper Aetqn, to make the desiqn proper we'll follow certain steps tn Ahis Cares desiqning & leaming system fs a Five -step process The steps are, 's choosing the Thaining Experience 2 choosing the Target Function - 3+ choose a Representation for the Target function. s+ choosing a function Approntien ation Algorithm 5+ The Gnal Deitqn A. Choosing the Faining ermpenience. + The first design chotce {5 to choose the type of training experience vohich the system will learn + «The type of training expenence available can have a stqnifeank trom fmpack pn success py failure of the learner.There ave three atbibutes which ‘impact on success oY failure of they learner + tb whether the haining enperi ence povides divect ov tndivect feedback regarding the choices made by the perfor mance systern The degree to which the learner conbrols the sequence of training exoenples 3+ wow well tb represents the dichibution of examples over which the final System poformance Pp must be measured, R. choosing the Target Function t- The nent fmportant step is choosing the target function . Tt means actor ding 49 the knowledge fed to the algpsithm the machine tearnfng will choose Nextmove function which will describe with the type oF legal Moves Should be taken For example : uthile playing chess with the opponent, when Dpponenk wilt play then the machine leaming algorithm will decide ohat be the number of possible teqal moves Laken in ordev-to ge Success « | NextMove: BM | Wis function accepls as input any board fromthe set of legal board states [Band produres a8 output some move from the set of fegal moves m- hn alternabive target function and one that will tun out to be easiey toleam in tht setting % an evaluation -function that ausint a numerical Sore to any given board state. ViBOR Denole that Veaps any leqal board state fom the set Bt» Some veal value. let us therefore define the Aarget value vb) for an anbitvary board State bin B, as follows if bts a final board state thatts won then vib) = 100 - RTF bf a final board stake that fc tort, then Vib) = ~100 - Bif biwa final board statethat t drawn, then v(b) =0°4 Th és nota Final stale tn the qae, then vib)= ¥(b), where bts dhe best Hnal board state that can be achieved stavting from b and Playing optimality until the end of the qome: S- choosing a representation for the target function: we need to choose 4 representation that the leaming algorithm will use to daribe the function NextMove. This function Next move will be caludated asa linear combination of the Folloving board features s + xitdhe number of black gieces on the board "X22 the number of red pieces on the board 4x3 the number of black kings on the board + ay tthe number of red kings on the board « #x51 the number of black pieces threatened by ved [ier WI Captured on red's next tum) x6 tthe number of red ptees threatened by black. hich Can be Newt Move =u0 + utxd +u2x2+ U3 x3+ULXy + U5X5S + UKE. Here uo, ul up te UG are the coeticfents that will be chosen Ueamed) by the teaming algovithro 4 Choose a Function Approntmalion Algorithm Te learn the Laract functionNentmove , we require a set of brining examples each describing a speifte board state b and the Haining value (covert Move) Ytor bs The training algorithm leans] appronimate the Loeiyicienls Uo, Ul upto UG with the help of these training exam ple by estimating and adjusting there weights « » Rule for éslimating training Values. Veain tb) © -V successor Cb) « >» Adjusting the weights 2 : ( Vinein th) - 0 69)” 2b) Virgin lb)> € training enample beLMS weight updale vule tov each -yaining errample /£b3) VygifO?~3 (game history ) ' a ce '. Performance System The performance system fs the module that must solve the given Peformance Lark, in this care playing checkers by using the learned towgek function ts).Tt takes an fastance py a new problem (new gene) a Sapul and produces a race of th solulion (qame history) as out put Q- Crilic The Crikic takes as input the history ov trace of the game and produces 4 output a sek of taining exarnples of the target function. As chow ‘in the diagram, each ating example 7 ‘thes case Corvespon ds to some qame Stale in the race 1 along with an estimate Virgin of the tarqett target funckion value for this enample-3. Grenevalizer The Greneralizer takes as input the raining examples and produces an output hypothesis thal fs 4B eskimate of the target funckion - 24 genevallizer fom the spetifie haining Hramples, hupothesteing @ general functfon thak covers there examples and other cases beypnd the training enamples + y Experiment Gienerator The Experiment Greneraloy lakes as inpuk the current hypothests (cum ently Jeamed function) and output a new problem (te inikial board state) for the performance system to explore Tts vole is to pick new practice problems that will manimize the learning vate oF the overall system: PERSPECTIVES AND ISSUES JN MACHINE LEARNIN& - One Useful perspective on machine learning 1s that ttinvolves Searching bypotheres to Ackewmine one that best fib A very loge Space of possible leaner: the observed data and any prtoy knowledge held by the Tscuts in Machine leaming D whak algorithms entst for teaming general target funckions from Spee tvaining enampla 2 2 How mudh training data fe Suppicient 1 2> when and how can prior knowledge held by the leamer quide the process of generalizing from examples} Wy wlhat % the bel shale qy (or choosing a urelul nent training erpenrente, and how Aves the choice of this shaleqy alter the complenity of the leaning problem } 5) What Gs the best wary to neduce the learning tak to one ov more Function approximation problem t & How can the leamer automaltealiy atier th aeprerentatron to improve th ability to represent and leom the target function }CONCEPT LEARNING: AND GENERAL TO SPECT FIC ORDERTNG # Leaoning Involver acquiring qeneral concepts from specific training eramples. Example t people tonfinually learn general concep or Case doa Such as“ bivd," Seay", * eitualions fn which | should study more in Order to pass the exam, ete. * Gach such concept can be viewed as descsibing some subset of objet ov event defined ovey a larger set . Allemalively cach concept can be thought of asa Boolean-valued Function defined over Hate larget sek CExamplet A function defined over all animals: whoce value % trite for birds and false for other animals) « Definition : Coneepk learning = infening a Boolean—valued -function fom “haining examples of tt {nput and output. A CONCEPT LEARNING TASK Consider the exanaple task oF learning the target concept “Days on which Aldo enjoys his favourite water ¢port ". Humidity! wind | water |fove cast | Enjoy [Example | sky |ity Temp _Sport | | | i ' |Sunny | Worm Noval | strong | warm |same Yes | (2 Seana | warm | High [stiong |uiain | same — {Yess | || 3 | Rainy | fold. Hégh | shong jwsarm Jehange |No- | | 4: jer jwarm | High sheng cool change Yes | Table + positive and neqalive fainting enamples for the target | Concept Enjoysport »The touk is to learn fo predict the value of Enjoysport for an arbitrary day, based on the values of ib other attributes t ql What hypotheds veprerentation is provided to the eee * Let's consider a simple veprcuentalion in which each hypothests tonite oF & tonjunction of constrain on the instante athibutes * (ek cach nypothess be a veelor of sn Conshainks , specifying the Values of the sit abhi bute. Sky, iv Temp, Humidity «wind, water and Fove cast For each athibule , the hypothests will either | tIndicate bya “?" that any value fs acceptable for this athibute. * Specify a single requined value (€-q Warm) for the athibute, oy * tndicate by A“ that no value fs acceptable + TE Some instance m galielies au the Conshatnts of hypothests hy then h classifies » as @ posibive example (hix)et)- The Hypothets that PERSON enjoys HRs favourite sport only on cold ows with high humidity iS seprerented, by the expression. (4, told, nighs2,2,2) The most Yenwal hypothesis - that every day ft a positive example - {5 vepresented by (Lb Lnd) The moct specific possible hypothesis that no day Fe aw positive trample fs reprerented by (hb idibed)\ ~, | © Notation. ey *The set of stems over Which the Concept f& defined ts called the set of ‘Instances, which 1 denoted by X Enample: x ts the set of alt possible days » each vepresented by the athibules Skyy Atv Temp, humidity, Wind, water and Forecast * The concept ov function to be teamed fe called the target concept, which tc denoted by ¢ erx— fon} Enample t- the torqck toncept conesponds to the value of the abbibute Enjoy spor (i-€-1 Cla) =1 iF Enjoy sport = yo pand cla) =0 iF Enjoy sport = No)- “Anstances for which ex) =1 ave called positive examples, or members OF Hhe target concep. * Instances for whith E(x) =0 are cated negative examples, oy non- members of the target concept « “The ovdened pain (x,¢U0) t0 decevibe the Haining example consisting OF the instance x and th towel concept value Cl») + “D40 denote the set of available taining examples “The symbol H to denote the set of ati possible hypotheses that the le omer may consider veqarding the identify of the target foncepl+ Each hypotheses hth H vepresenb a Boolean valued Function defined over x hixs for The goal of the teamey ts 40 find a hypothesis h such that »)| for all min xo Given: P Instances X4 possible days, each described by the attributes: # Sky (with possible values Sunny, Cloudy and Rainy) + * Riv Temp Cutith values warm and cold) + Humidity (with values Novmal and High) + Wind (with values strong and weak)s * Water(with values warm and cool) « Forecast ( with values same and change)- * Hypotheses Ht Each hypother’c te deseribed by a conjunction of forshaink on the atbibukes sky, My Temp, Humidity, wlinds iter Vand Forecast. The tonshainls may be “2" lany value *Sacceptable), "tno value ts acceptable), ora specific value. |" Tavaet Concept c+ Enjoysport ¢ x fon} |" Training eramples Dtpositive and negative examples of the target | function. * Dekermines "A hypothests hin H such that ht») = CL) for all xin x Tablet The Enjoysport Concept leaning task: The inductive leaming hypothesis Any hypothests found 40 apprornmate the target function well over a suffitiently large set of ‘Araining example: will also | Appronimate the target funclion well ovey ather unobserved examples.CONCEPT LEARNING AS SEARCH S concept Teaming can be viewed arthe bark of Searching through @ large space of hypothesis frmplivitly defined by the hypothesis Yepresentation. ° The qoal of thic search %s 40 find the hypothois that best fik The haining exampler+ Example = Consider the instances X and hypotheses H in the Enjoysport learning tok The atbibute sky has Ahver possible values, and AivTemp « Humidity, Wind water, Forecast each have two possible values the tnstance Values. Space x contains enactly. 3-2-2.2.2-2 2496 Afstinch instances - Dee He = 520 syntactically distinct hypothaes within H- Every hypotheste Containing one or more “P* he) tf and only if | [CV nEX) [Uh =) > Ch Trstances & ™~ >) Z| si a uz €Suunny, warm, High «Strong stopl, came > hy = hy = h3= j FH the Fquie, the bow on the lett reprerenb the cet A of all Instances, the bor on the vight the set H of all hypotheses: # Each hypothesis conesponds to some subset of x- the Subset of ‘instances, Ahak at classiqier positive. [# The ayows connecting hypotheses vepretent the move - general - than Ydakion, with the aro pointing toward the ters general hy polherts- © Note the subset of MMstances characterized by ha Subsumet the Subset chavacen zed by hy hence ha ie more~ geneval -than hyFIND-S? FINDING: A MAXIMALLY SPECIFIC HYPOTHESIS FIND-$ Algorithm Ie Thibialize h to the most Specific hypothesis in H- 2. For each positive haining instance Foy each atbibute tonshaint aj inh. Tf the constraint aj is cakiclied bya Tan do nothing Else veplace ai in h by the next move general constraint thak is akistied by a. 3: Output hypotheris h- To illushale thie algorithm, assume the learner is given the Sequence of training example rom the Enjoysport task. tind | Water [Forecast | Enjoy sport | ' sunny | Warm | Normal Shong | warm | same Yu | Example | Sky | Aty Temp Q [swnnd warm | High Strong | Warm | same Yes. 3. [Rainy | cold High — |Shong | warm jehange | NO- 4 sary Warm High shong Coo} | thange | Yes e The Fst slep of FInD~s % totnitialize h to the most specific hypothate in h- 1h, 0 bs pip) © lonsider the first haining example Xi = , + Observing Ahe first Haining example itis cleay that hypothesis h te 4oo specifics Alone of the “P" tonshatnis in h ave calistied by this example, sp each 7s Veplaced by the nent more general constraintthat fil the example hi = ® consider the second taining example My = * comider the thivd training example Xa=, - Bpon encountering the Ahind haining the algorithm makes no change tO hs The Find-$ algorithm simply ‘qnores every negative example: ha = * tovsider the fourth taining example Xq = , The fourth enample leads to 4 further generalization of h hye hypotheses H | oh? || Specific instances ¥% | o “e h : | SP , General Xie dsunmy War normal shong warm same>+ ho = Xe ¥= , + Wye cgunnyy warm 4} Strong’ } >The key propaty of the FIND-S alqovithm id + FIND-$ f¢ quavanteed to output the most specific hypothests within H that fs consistent with the potitive training cram ples - « FIND-S algorithms Final hypothais will also be consistent with the negative eaamples provided the comect target Concept Te Contained tn 4, and provided the training examples ave coneck Unanswered by FIND-S- ' Raw the teamer tonverged Yo the comect target concept t R why prefer the most specific hypothesis} B+ Ave the haining examples consictent} - _ 4 764 We wohat there ave several manimaliy specific tonststent hypothe sd VERSION SPACES AND THE CANDIDATE — ELIMINATION’ ALGORITHM The Key idea fm the CANDIDATE - ELIMINATION algovithm %5 to output @ Adsorption of the set of all hypotheses consictent with the aining examples. Representation Detinition Corutslent- A hypotheste h ie conuistent with a set of haining example: top if and. only ff by = ex) for each erample (ms cix)) in D+ Note difference between delinitions of tonutstent and satislies- e An mample ais said to satisfy hypotheis h when hoy =! Yeqardless of whether » <> a positive oF negative example ofthe target toncept + * An example a fe said to consistenk with hypothesic h iff hor) = cb) Definition + Version space - The Version spaces denoted VSj_ With Nespeck to hypotheds spate H and training example: 0) te the most cubset of hypotheses fom H tonstetent with the taining eramples tn D —————————_--—— Sho = fhen| consistent tmod}| The LIST- THEN - ELIMINATION algorithm The LIST- THEN- ELIMINATE algorithm Aivst tikializes the version Space to contain aly hypothats in H and then eliminates any hypothe found inuonsident with any taining example- 1 Version space € a Vist of containing every hypothats fn H- 2 For each taining example, (2, C09) vemove from version Space any hypothats h for which h(a) CO 3+ Output the list of hypotheses in Version Space The LIST- THEN - ELIMINATE Alqovilhm - Uist Then - gliminake works fn prindple, so tong ax version space Ts finite. eo However, Since ik ts nrequives erhawtive enumeration of au Nypotheses in practice tr fe not feasible-® CANDIDATE - ELIMINATION learning’ Algorithm The CANDIDATE - ELIMINATION algorithm computes the Version space Containing all hypotheres from H that are consistent with an observed Sequence of training examples. Tnikialize Gr to the set of monimally general hypotheses in H- TniKalize § to the cet of montimalty specific hypotheses tn H- For each Haining example di do *Ipdtea positive example * Remove fom é any hypothesis inconsictent with A- * Fev each ypovhests § in S that te nok consistent with A * Remove ¢ fom S. ° Add t0 S al minimal generalizations h of ¢ such that “HE consistent with dy and sone member of 6 ts more general than h- * Remove fom § any hypothesis %& more general than another Hypothars tn Gi. . CANDIDATE - ELUMINATION algovithm wing Version Spaces» An Thustralive Example \7- = i Example Sky |Niv Ternp | oat = Ey Sunny | Warn {smal Strong | Warm | same | Yes 2. Wind _|water |Fevecare |°Nd oy, Sunny | avr | high strong | warm | Same |} yes | a Rainy | told High | Stong [var [change | No | | 4 | sunny | eoaim | righ | Song | cool | thange [yes _|CANDIDATE~ ELIMINATION algoyithm begins by iniKializing the version Space to the cet of all hypotheses fr Hj Tnikializing the G1 boundary set 40 contain the most general hypothesis in He SoeEyuyrl> 1ifte Trifializing the $ boundany sel to contain the most PES Cleast general) hypothe + Soc pi bi Pipi hih> Swhen the first haining example ts presented, the CANDIDATE ELIMINATION Algorithm checks the S boundary and finds that tf # ovenly specific and W fails to covey the positive example * The boundary is therefore revised by moving if to the last more general hypothesis that covers thts new enample- "No update of the 6 boundary is needed in texponse to this training Cromple because bro conectly covers this exam ple- For training example d, | $1 [é gummy, war, Normal, song, warm same’ | Gotu (€9,2.1,2,825] # then the ceeond haining enample {< observed pit har a Sinilay Hfeck of qenerolizing § further to $5 y leaving G again unchanged fey Giz > 6u= Giorn fo) Fer training example d, + pemanaiemin eli cei + [esurmgreteom nora song wan same>| $a E sunny, warm, 2, strong, Wann ,Same> | | I Consider the third Haining example-thic negative example vevelar that the 1é - 7 | ST boundary of the veision Space ts overly general, that te, the hypotherts \8n & Fnconcctly predict that thie new example ts a positive crample- The hypotheale fn the Gr boundary must therefore be Spectalized until it Cov ectly classifies this new negative examples | Foy Haining example 4, < Rainy , wold , High, trong wavrn, change? |< sunny, warm, 1, Strong, ularm, Same> $953 Ga [ Counnytrt BYIDLY, warm, BLVD SPY DL Griven that there are six athibuter that could be Specified to specialize Grr, why ave there only three new hypotheses in 6132 For example, the hypothesis he (212, Nommal,t,,2) fs a minimal Specialization of G1, that tonectly labels the new example as a neqative example, bul it fe not included tn 613. The reason thts hypothesis ie excluded fy Ahak tf ic Invomsistent with the previonsly encountered positive exampler-# consider the Fourth thaining example: For training example A, < sunny, mane tigh Soong cool, change>t 3 [ Sy sunny Wares 1, Shang 11> Gia [Z sunny.?, 9 eran wart, ) Wert, 11,4, sare?) oth ; “ posibive example further generalizes the $ boundany of the Version 5 | PPACE+ TE also vesutte tn vemoving one member of the 61 boundary» | Recourse this matter member faik to cover this new positive example- Alter Processing these four examples, the boundary Seti sy and Gry delimit OF the version Space of all hypotheses consistent with the set of tnevementally obsevved haining enamples. Sy [ €sunny, wam?, shonq, 12> € Sunny 3,1, sheng, 4, 19 L6unny swarm, 1,4, 33929, Warm, 2) Shong, 2.7> K f fy ( 4 No| INDUCTIVE BIAS The fundamental questions for induckive inference: ' What if the target concept t not contained én the hypotheds space 2 2% Can we avoid the AifReutty by using a hypothesis space that | includes every. possible hypothesic 9 F How does the size of this hypothests space éafluence the abiliky of the alqnithm to generalize the unobserved tastancer How doesthe size of the hypothests space tnfluence the Number oF Haining examples that must be observed 2 These fundamental questions are examined ty the Contert of a | CANDIDATE - ELIMINATION algorithm A Biased Hypothesis space * Suppose Une Largeh concept fs not contained tn the hypotheis space Hy then obvious solution ts to ensich the hypothesis Space to inctude every possible hypothests« : * Comider the Enjoysport example in which the hypothesis Space e | Nesbrcted to frelude only conjunctionsot athibute valet aie o Wnts vesbiction the huypotheste Space te unable to represent even Simple disjunctive Larger concepts such as * Sky = Sunny or sky = cloudy * Ssiuncti ae, the \* The following tive haining examples of disjunctive hypotherts » algovithin would find that there are Zero hypotheses in the | version spacesZsunny, warm Normat Shong cool change» .¥ < cloudy warm Normal shong tool change > < Rainy Warm Normal Shong Cool thange> N © Theandidate Elimination algorithm % applied » Hhen it end up with enphy vesion space. After first two haining enample: $= <1 ulam Normal strong cool change? dhe thi © THe new hypothests & ovedly general and st fnconcelly covers “the hird . fiate C+ negative training example | $0 H does not include the @pprop™™ o 7 dived. # In Pris case, & More expressive hypothesis Space '¢ vequi - An Unbiased Leamer. # The Solution to the problem of arsuring that the dargel concept zim the hypothesis space Hs 40 provide a hypother’s space capable of representing every teachable tontept that ts vepresenting every possible subset of the instanus xX. "The set of al Subse oF a set x fe called the power set of X- * Tm the Enjoysport Ieanning task the size of the instance Space x of dos deswuived by the sin athibules ts 46° Instances # Thus, there are g dtetinek target concep that could be delined Over this instance space and leamer Might be calied upon to lean. # The Conjuctive hypotherts space te able to teprerent only 443 of these biowd hypothesis gpace indeed * Let Us Yeformulate the Enjoysport learning task finan unbiaved way t by defining a new hypothesis Space H' that can represent every subse ot instances.@ * The target concept “ skye Sunny or Sky cloudy" could then be desuibed ar Csunnget 253) v (cloudy. ty 12,9 2) The Fubi lity of Bras~ Free leaning Trductive leaming requines Some form of prior assumptions, 0” inductve pias- Definikon Consider a concept teaming algorithm L for the Set of instances X- eet c be an Arbihary concept defined over X- Stee = | OU CO0)} bean arbitrary Set of ha‘ining exampler of ¢- # Let L (Xin Be) denote the clawificabion axsiqned to the fnstance Ki by L after taining on the data De * The inductive bias of L fe any minima set of arsevtions B such that for any target ‘concept cand corresponding training examples De © (VOKHEX) ((BADEAKI) FL (x1 Dd] The below figure explains * Modding inductive systems by equivalent deductive ystems. * The thput- output behaviour of the CANDIDATE - ELIMINATION algorithm Using a hypothets space Ht tdenfical to that of a deductive theorem Prover utilizing the axention “1 contains the target concept?” This assertion {6 therefore called the induchive bias of the CANDIDATE - EUMINATION algorithm.’ chavacteriring inductive systems by Ahety induckive bios alows edt them by thei equivalent dAcductive System This | provide a way to Compare inductive systems aceording to their Policies for eeneralizing beyond the obserned training data. inductive System : Training examples J candidate t c Ti T Elimination ass: Freatton of new instance, Or alaovithm 4 doole eno" New instance | “sing Hypothesis Space W- Equivalent trduckive system As classi freation of neve Training example. : frutance,or “don't * know". Theorem power > New Pnstance Assertion’ contains th target concept’ a —___. inductive bias made enplicit \