MACHINE
LLNS
MD ho MITCH DRE
Ca INDIAN EDITION7
: ONIT=T
———
Introduction
The field @ MLS Concemed with “how 4 conshuck
Coraguber reggromms Hook aukormetically Tmprve uirh experience
\
Succewtl ML apelieations
\ Dakeawining Groqram Hut Lean te detect fraudulent |
credit card Lramactions
2: La fomekon Filtering system jk Learn “users reading _
Preferences"
3 Aukeremow velicles dat beam ty rye on qublic Nighwsou,
e
* ML concerts and is erulty come fom many freloy incheding-
= Stabidics = biology
— AL De cognitive Science
= Philosophy ~ Compukational, Com plexity.
— Leametion Theory
= tealre| Hreory
We gock ¢ this course 1, te Prerent te
“ey Alsorithras awek decory oat form Ha €or op ML.
SME: Ts twherently 4a mulkidiscs plinemy Feld
+ A sHecephd orderstonncting of hod to make compukers Learn eseuld
FORTE many news vies a comple «
+ fk Seems mL wil glay an nerenainghy Central role
in
Computers science:wedonok yer Know how te make computers Learn Nearly ariel
however obqeritnes (alg 5) hove been
geogie Learn
ay
invented Ask we effective Fa certain typer Reanning tasks
EA certain types sp Hearning *
such a> speech recognition NLP’ (Natural Lempuoge Procensing
Paying Games avd Docker Minang, ( crbact trocledye fom Longe des) -
Desa (Gotine op oan)
y
~ Gemgulers Lenawivs from wedical reconds to ove treatment fa |
new AiSeares,
Howes leaning Fram eryerience te optimize eneray costs
bayed on the: Joie rage patterns af their oceuganh,
= ee fearwang Ake interests of Ahety users inorder to igh Wight
velevank stories Frm dhe online motming news papers
D well- gosed Lecaning. problems
Leer sty Computer freqram Rok Improves ts Rertrmance ab
Some tase Arreugh experience «
+ ih general te hae well-defined Leann qotlerns
we must identify 2 peatiney
+ Ate claw fp tasks
+ Ate mearure of YeTfamance to be improved
Ae Sowee op experiences
ae Computer Qrarain Aut Leann to qlay “checkers”
wigek Ampove ths perfamane ay mearured by Its
Isel
if
ability te en ecarouh Paferiente obtairrea & Plering gpme ont |
|
|| ose
Hardwviting Recognition Aesrnivg problem :
Task + feoynizirg acd cownifying handwritten Words within \mages
Rerfoimance: Measure 0 Fercent ap werels correctly claified
Treining experience £5 A Dw a haduritten words with Given
clawification :
|
&x@:
Task Driving on Public Aare highesays vsing Vision Sensors
Ferfrmamce Measure PL Ay distance traveled befre an errrdg
(@ dyed by human overseer)
Theawing Experience E> A sequence 9 Woges ard steering Commend
recorded ubile observing a leman Arwer
(P) Designing a Jecoming System (5)
+ choosing be Areining, experience
: Choosing, Ale Target Function
+ choosing a reprerentation fn Ale Target Function
cheesing 4 Fanction Ppp re ximetion algottnm (2)
he Final design:
BD chering Ha traning Engerione
| + The Pmt design choice is te choose Me type y training
Experience Bam whith sur System will Leann.
~ Th cam bee Sgificant impact on Suce or foilune op the
desrney-
r2| /
= es 7| = one Rey allvibute 15 wheter Ae Lrainirg exepeniene (rbd
Provider — divect cr indirect feedback regarding Abe “Chisices
WY He Perfrmance System:
Er while Leanning to play peckers ,
a the System wig heem from dareck examples camnisting op
tndivdluat checker beard stokes ad Ma covect move Ar each.
~ Auckireet infe, commisting of Ake move sequence mmol final
Sukcomen sf Veriows Fpomes Played .
Note. « Learning for direct breinng feedback 6 easier than
» 6 wdivect feedback,
|
|
|
|
|
+ A second tw albvibute oy dhe FOr is ale |
| ea" examples.
degree to which He Learner controls Mu Sequence Krein
+ Ger
The Aewmer wight Tely m He teacher te select beard shedes
and te pryide Ae Correct prove FA each » Er)
- Te heanner might itself Propose Leeard sheder 4Lt it Fincls
confusing ad ark tte teacher fr dla cored move
Gy)
- We deamner mow have. Complete Conte) over Lot, the
boewd deter ard reiving clavthcation
ae :
rd altvibute ai XP tg hes well tt repreyet dhe
iter cecal,
Mele: qe crucial aysumplion Abt dhe cichabudion op ‘rai. enamples
Ke identical 40 the — Aigwitition og dest enaimgles.
AUS budion ge exampfer +|
GED choosing a Target eunction
Te net design cloie is te determine erocty what tyre "|
Rnoukedge wil be Leamed and hoo Ads will be wicd by
Rerfamomce Program. |
+ Ger checkers~ playing groqram atk can gprerede Als hegat
moves from any beard state:
~ Te. rq weeds ony to hear how 42 choose jhe beat won
From cmong Arese Legal wes: |
| c iiss
[tLe us Call Ais function (fa) 2 ChoosetMove 2 B—>M | |
Ve indicates Adin ty accepts asinqut omy Lert From B-
(® = set # hesal board stote) avd produces ay oakput |
|
Some wove fom M ( ser # leged wwoves) -
+ dmither Larger ty 18 am evaluation ty MePR,
dr denste Abst ov maps amy Legal beard Shle frm & to
Seme veal value R-
- ¥ akign higher Scores fo better board states.
ga darapt Velne VL) as Fallows :
"Lek us define
1. TB & isa final board inte Ht is Wen, then v&) =100
ac ae =e “ » dost, then *VEJ= -tod
|
| a ae a ‘ * drawn then vib) = © ;
STP bis a nok a final sete in Ae ame, then viey=Ve')
Be ete Lat finek beard siete from b ah playing i
optimally unt] dhe end of te epme- |
'| &D Chossing & Yepresentation for dhe ‘Taract Function *
+ we wwst — dhovse a repretectalion tat Are Learning Program will vse
to deacrite Ate fy Qo At ik uf) Leann
pv
| we coud allen dhe prgram te = 0 eittver
ee Lang fable wills iskinet ent fn cach didinct beard tebe ()
~ wing celleckion ap ruler tuk wutdh aspinst board Feshues Gr |
quadratic Polynomial fn ge predefined beard features (67)
~ avkicied
neural network
+ Br Lek wr cheese a Singh regrerertation! EA any given Loord state,
des fa GF with cabeabeted ma bineor combinetion 9 dle
Following board features
H+ Ae morep black pieces om Ake board
tr. te red 0 "
Fa ey «Gack Kings "
Hy t “ ‘ey red “ “
Ris 7 az 0
black pieces AHyreate ned by rvecdcie)
which Can be captired on red ect turn
Xe dhe no: ge yedd pieces threatened by Lack,
Thos our hearning fregpram will ye present VW) aa Linearty @ tte
— Boum
[Pereos weet 18 + Asks 48 ety Beko tdg |
where We te Op ae numerical CoeKiclents or weights. ‘Aissten. by
de Aesaming ag. which determine the velakive Importance
dhe Verious board fecbures +| @) roo sing « Funckion Agproximation Alserithm
ot ty Ven he Aenggt 0 wee_reqpitt set ge Lvaingcam pled
in Ake form lee b, ers
Eo Tee flaming Arnining ex Aeseriber a board ste & , in which
black hes won Ae at o
LL Hd, HHO, KHEV, MH he =O, 2% :
Mele? 7,25 Mea no ved pieces remaining .
tele) og Estimating trosning values
bee ALjuating he weit
+ Eshimatirg Lrotning oluen ha Iwler mediate Goard States as follow
I je 9 (socomatt)) seer (VD
where Succemor(b) dewter die next ae sea clebe ollouing, b
+ Adtysting Ae weights is 4 specify Ate hesrwing oy fr
choosing Abe weight 4 to bert fit Ake set sp training Champley |
| § de, awit > 4 | . |
Batt means definig Me Lest hyoduests oy set weights
as Me A Riess Me squared exnr E bekween -
and fe Vober _redictest by je ye ai vl
Ake red. Valo
ae _ Lo = So)
E=
4 1 Mena) re. E TreinExanglen
re
As| Several alg ane Known fa Finding weights oP a hiner fin flat
winimize defines? 4 Hay vey & LOS (Least Meon Squeer)
‘ Los weight update Rule -
€or eae Aveiving ene db, Vaan) -
A
"Use AL. cued eights fe caladate Vb)
"FO cack weight
lowes Wr (Ya lar Ve) x ‘\
-
Here Yc Swell
Update ab as
Canton (Ex: 0.1) tut moderates Ma size 4 4
Goeight update «
* when 4
en Abe ener 1g 27 Ho coeigits ow Changed
. % 4
then (Mea. ve) is 4ve (i Ub) Is tee how ) 7
VS increavect in
Alen -eacl eiefat-
Prprtion bake value ote Corverpading fecdure
This will poche ener.
a a a a % 's zew
1 den ts veidte 4S nok
akered regedlers Perey:
© Tee Final Design
The final design any learning System can be roturelly
described by 4 Preqram wodules +
Lrg v J
"WF rmource System
*! “Gite
Generalizer
<\
Experiment Generator - : oy=Tervtamance System! ts Ma module Wat wut Sele tte givens
RErfamarce desk vsing hewmed Compt functions -
~ NF Aeker aun instance Poo new protlem ao input aed
~ grees a hace og Ws Sluhion (gem history ) as ofp
~ Trowr ca, Mor Sealey wed by Mis | whedale te
As select nex\ Gest move at cack Step Using evaluation tal):
Qritic +
= a tarea ar Yp Ate kistry al produces a set F training
exams zt tbe Acyrget function
~ Ta cur care, this module corresponds te Ma Lrainwg role (rey
Generalizer 5
- ge tae as ‘ve Mu Krainin exs
hypottesis tat ts Hy erkumale op Als Lacget fe
and produce an ofp
— Te Wyfferse sizing x general fy Mat covers give Lrevining €xcimtles,
awl other Car alio
= Jr our caw, Was module currespards to Ae Lins aa
Experiment Geweretor |
AR hares od Sp que current hy portaesis ( curently Jeamned fn Jaa
aes a new problem FMS permone System te explore ,
Note Many ML Systems can be vsefully chancterized Th terms
op Mee A gerteric modules
ome Sequence ap design choices mace fy He checkers preqram Is
Summorized jw [ Figure ve]
be
- te]@ Perspectives avd Iscues WW ML *
Perspectives
~ PML Anvelves oy 4 & Very hare space p possitie ly potheses
fe determine asp est Fits Me observed data
~The LMS aly fo Filing wserghts | tkeredively Auning Ame
warts , adding Copyections to cach weigdt each Lwne fhe
hypothesized eval. fy gredicts « Volue Atak dilfers Pram he
Kroawing data:
“t
Merent hygodresis regresenkotions ane agpropriake fr heanning
AR Kinds og ctonagt functions:
Assuey (Humber $F gpneric question qhowk tw)
= whet algorithnn exists fd Learning oynerch Aorget Fn's From specific
Ayoaning exam les
Sit alas qertormn beat, fy dich ty 964 + poblenn freprenattiont
~ How much Arearing Aska 1S suicienk 9
~ Caw Prior knowledge be helpful even when ik is only agproximately
Corveck 4
— whet |
Wo Ake best way de reduce Abe hewrning et
An agpreximstion pratlems J
Hd Can dee Learner auttometically alter ts regreseadation to
Twngrove Ws ability to represent and Learn dhe target fn
i Course 1g concerned with Amsuering age above question,| Concept Learning Task
|
+ Muda heernivg, ‘volves aqwiring general Concepts from
sqecttic Lretining Catarmgles.
Fora given Ayeining examples labeled as members oy
nonmembers of Ake concept 6 referred o> “cence gt leaving!
©) agtrorimading a Lrclean-velued fn Prom examples:
[Table oy
alle 2) dercribey a sd ap examfle due — each re prevented
by a sk ge nthriluter
© the abtvitute “EnjoySpert" indicokers Whedler or not 9 person |
Enirys kis favorite Welter spat on this daw. \
4
"The tose ts to Ream to prediet Ae Ualue g alvdecte
“Exjoysyert” PA an ay bikrony day bared ot the values
% ks cher odfributes- \
wyfotbesis + Let each hypothesis bea vector of Six conhainh,
; : a
Specehying the valor sp tte sin edbviluken wm AK Tablery +
For Each abtritule, He ryfowesis iW either ga B , repuredul
L~ twcltester, drat omy value %¢ acceptable fy stu oblvitute »
¢- “ "no value is acceptable “ «
rw \wqpesis Abas a fersen enjoys his Favorite sport only
on MA dows eomk th Kigh Ian Omcagendek op te
ole addi bube a) regrrnted by Aa enpreniion
fea]
7 |EEO TL modk gener hypoikesis — Ast every doy Sa te keane
1S veprembed by
rapa agg
LUA |
* The definition =e the “Ewioy sport” concert henoning hack |
im Abe General form 1% given [Tate vr)
Twtences; The sek op thems over wich Re concept te AeFinad
Ws called a Sek @ imsbemces -densted by KX
The — Concegt 07 function 40 be Learnect 1 called Abe torept Concept
denled by c-
Th agnert ‘ce! can be any boclean urduest fy Refned over X.
i’ cr X> fon
coat 1 Exiysqaty = yes (fositwe Exsmp ie)
cay =o if Eriyyszat = No (Megane Semple )
COHN a abso Collect Member op the Lerget Concept «
Oe 8
thes is
Note — she Goat # the heonney is te fined o hy pothesis h Such
Het WOW =e) Fa a a in X:
‘eatectve Aevwing hyteeris: tony hy pohesis fuind appmsnse
He Lavet 21 well overa Shi ciently derge set op Araining exlangle
Wi abe Gpprtimste well over the, Unobserved exempler «
Ths 8 th firdamental ti
incluctive Leasing, — oeC2 comce tt Lesrning 04 Search
"The oak of iy Sear is te find He hypothesis tt best
Fis te brauning rermpler - y
~Ceider Gor nncty X cand hypotefis Hin she "Bajey sent"
earning teak +
‘Sky’ nthille her 3 potitle valu ay
Airleme Hewitt, Wind, wader od fore cost each have 2 pot
tae a
> Whence Space X — conteins erally B+2+2-2-22 =96 distinct
Veluers
= Sinclely SKYUth AY GAG = S120 Syntacticetl, Aislinet
hy pie ses ain
MS ee sry sintle fecrnivg AER wiih Swall
Fiwike bypotteio space
Mose Proctical Le maning Aerka invelue much, Langer ,
Some biomer infinite nplesis space
Fw aed 15 find cloprittens capable 4 Mcveth,
ee
hypettesis det Let At tL. TD (rainy Data) «
General b Specific ordering of Hippies
Conder two hypothesis (yp)
he S sm, 1,4, Swog, 4, 2>
Fie ce age ee
Vere Ve ptten, fewer comshoiah on the instances , Tk claaBes
Wak iwhewces as tye.
adPy tartemee Clam fied te by Ly will abo be clanitees
ave by he
Hr 1S mae qeneret Ly
* Gwen Wy ant Ie hy i “ade general Abe or 2g tual 46 hw,
1
VEE aany whence Rut Sabis tien hy alto cake fies hy
| Gro ou =1) —(h; @e= ) |
[ Fe a] :
The me velation 6 imp becawe (+ Provider a wcbd
Shruckwe over He hyp» space H fy any
*
concepr
Aa mrning Problem
eek eee LA maximally Specific hy potlesis :
Final S Algpritlen : a a
T, oe Wy & Me est Spedfic hype in e i
heen Pestve raiweg nttance 1,
7 ORR eaeh albu Cantal a wh
TE He coment A, is oats felt by x
An do welling
de refine ow by Ae West mde gener
1S setisfiey by x,
byt h
“Th fist She # Fads IER Wwittabze oh dy sant
h< £86, 0,0,0 05
Seunider fink ave iusemue (mw) ian Teele 21 rot Cabs Bing
Combine
3+ outgue
syeceh,
7 cMSo each iS peglaced by the
Net Wd Generel Cambria Host
fits Hu ex
| h ae Sunny , Werm , Normal, Strorg. Warm, Same Ss
‘
tmuoh we seh Nery cgecific
> 4
Dy Keni ex Rexeey Me ay te fur her Fewerelize h,
| hed Sunny wean , 7 ) SITE Wer Same >
PR FA kate :
| Bi Reiiiex, & (goes ve creampie.
|
| A
| ~? FP AM teniniey ex, ead & Beetles ener lization h
e.| a se
|
We Gsm tom, L shee 1 TS
No: pk enh Ghaye, Abe hw ‘is Me oct Specific hye. Cons tet
with AL rect eX clterved ug AMIS Pritt = CLence
| mame Fords) +
| (Fa, rr}
| Nee: Find-S lc quemanteed to cutgut Ale imesh ¢ pecific hyp
e| toubin HL ALES cenaistent eal Mle AVE traaing eracmple.
Ts find hyp. will abe be coustet with He Ve
example
1
| ID ID) Version sgaces ond de Cardidete- Elimination imabion “Alger
+ Comdidale Elimination (CE) aly aatrenies Several Luvttakion of Find-<
joe
| The RY Mea in Me co alq ig te outgut 4 dercripliow -f Ay
“set # all lypedaeses” contistent with «the Lreunieg exermples
6 coy ad Find-s dg ome himted ey bce perform Poorly
when Wen noisy tunis dota -
x wisy 9) dake, ie+ Applicekions “ cE-aly aan
+ hemrning regueution tw” chemi col mes spect swpy" avd
~ heanivg conto] ruler H "heyristic Search"
“B Consistent — defjnation
A bypothesis W 1S consistent wilh A sek sf kei: Examgler D
wee Wy = CO) Fr each ex Lx, GOD MD
| consistert(h, D ) (44x oy ED) ke =e |
sokisty: pn ee ete said solisfy lnypoanesis by whew hoy =)
wegerdles 4 gheter 1m ne ~ ctamtle: § @
Consistent; An 2% 1 ie conisterk with Wh depends on Ae Leryet Concept
je hG) =c@)).
Version sgace: subset ad hypetesis 1s cae the version space
tort hypotlesis space +.
Tae = [he [emsnalio)) |
subset ot hypoteesds from H consistent wir Ae Aeaining Cxarutly in D>
+ we Com represent dhe Version Space in terma g tks wos specific ¢
aad most apnerel memberlG) pluc hae hypesi: UE Lbtucen
Gad S wm Me qoattally ordered hyp desis space
[V0 ek | (4s €5)(34€6)( 34 hays) j
[a3 )The CE -ahy computer de VS (erin spate) Containing all
byphests Fem H Hub om comstet uith au observed
Seqwence ge keainey etarmpler
* Tk begin, by Wwikiebzing AL WS te dln set gp all hyp. in H,
by iwi kcabizang Ate G bumdey see te
GO §
and
(wikis
Me Leandiony Set te Me mmesy specie. hyp
So <— § £6,¢.0,0,6,0> 4 ad
+ even other bypebess wm Wis in between G, ond Se
| (Tables) [Rg 24)
| Figure pt shuws ce-aly applied te de fink tue Keesiosey eonmple
from Feuble a
| 7 Pestve training Ok. may Ane da S burday op fe VG bk
lecome wereaurgly qpneral:
eo. Negative trainiey exs Woy Frrcing Me G bey fe bewme
| Wworentmgty specific
Commer dhe Aud Arininy Ge, Sheen in (Fy rs |
© The yp. in Ae G ww be specialized Until ip corseetly clemities
War new — neqalive €amgle
oe ote ts excluded becom wh is
Imcewistent sith He previolily enteuntercd prstne mmples
The a teeing Ge, as thee tn (re a6 |
| «Overall “VS 1S sham in LFig 24 |
9Ehense!
the WS earned by he CEraly dat correctly dercri ber Ahe
Layer Concept — providect (when)
Vi here a Ro errors fn Ahe training Excmplea
Be Tew 15 Some yp. Tn Ho dla Correctly dearriber dha tavggl Conceg
What training Be should dhe _Leamer Request Next 9
~ Uf bis point Se have attumedt te Loniniry
"But if incervectly covers da 3 nega trees ex»
DPT gable 1s aur be he bined dhe hearer de conider
oly Conjunciwe ly petheses
+ Leb vc reformulate flu Exyy Stat earning fark in an Unboiasect Way
by defining nee hypothesis space yy) te
allow arbitrary
| Aisjunckiens Conjunction, aud Regalions ap cur easler hyp...
SO Samy LAD VC chk (1,1,3, 21> )
S$. £(WVUVG }} Aisyonction
Gq: § 4 (x4 V ty) \ neglion
Aurdludive | Lessniny Algeritina (1) =
deductive bias gL te be Me Set 4 assumption B Such det
Ar alk new instances 1%,
(BA Der x) ELA 1) [
Notation Yb 2 indiceter dt 2 Allens deductively tom ¥-
[¥% 28 |CD) vecision ee Leaning — traduction
Decision Tree (DT) tering ts cme fp dhe westly, widely wed
ard prackicd metiod -fr incluctive infe rence
+ Dr Ts yebust to netsy dakar acl Capable oe Lemming
Aigjunctive ex gresions
Tormily 9p 7-L algorithms one 103, ASSISTANT ad Ch-S.
+ Lewmed trees com abo Le vepeprted 2, set
| # Wren
| whe bt imprve hwman readability
|
» Aoficettons + Applied te a breed Youxe ¢ traks fry
deannivg, & diagnose medical cass tu @
deeming fo aren credit risk lean epplicelrends
Decision Tree hepresentation ;
Toukleck
| Sunny / ove e
yes Lwin]
nig” Shey _
No yes No Ye ®
[e037
[TRAST 6 feomed pp te clanity Sodurcley martina, gccordingy
whelker | they awe Saiki ble fy play
ny Kennis .
Jo general OT eget a disgunction op Conjuncdion x
| compainks on He athitute values ff imbenes -
6 Gack goth fom He eee root be a Leak corerparch © 4
conjunction cp atbyiluke teats ad
| Me keee thelf fg dis inction gf soe Conjunction.
| EY (ouklor = sunn :
1 A Ruwidity = Hormel) ¥ (cuter suerest Vv
(Cowllor= Rain A inde eak) «* Each node wm the bree Speci ties a Ley F Some albyibute
te witance ard
+ 20. branch der corcting fom Aet node eovrerpands & one et.
fesible values fr his odtystake
GD Aerrepecte pubtem fo pT-Leawing (071)
' DAL 1s Feverclly bese Suited te Problemy will Whe fellesing
chavact evi stics
V Daskances on Yepreiertecd by tlriluke value pairs. (Purnia = bah
2. The Kewaet function bar dis cvete cak put values -
SIA clothes Tiscele: Ase g data ® sad fo be discrete
abe values belonging be MRE ae Aistind v sep.
anu
2 disjunctive description may be required ,
GoM. fresning data
s
Th
wey contain
ervers
Sin ‘y tal va
MSSING Athalde vals Cece, a
‘J
:
CD eave bre (Pere Len
LAkyrbon )~ 193
3
Selecting, whidh cthibde
v The cervel clove Wa pz aly 1s
Re CO AE each vetle 4g) Me tae
© WS wes Ty pte Gain measure te select
ook 1S mosh wet Pa chow Fy;
whale ory He tree -
(Fave 31] = OTL alain
“ Comments Wer Ae" Ear,
Characteri2es He Cm) purity Marlys hey Callechion -y ekamples =
Ie dhibate
we eames ok cack sep
Ta fanetion Gai TY) mesure Abek
te| Given a Collection Ge Comeswindy Positive anf pee
"GW A
| F Some eseh Gncept ;
SN Crm les
Sen) 6. fe bay ie (okg.fe) )|
te "S$ tee Dreger tion F Positive ewamples in cs .
fo ‘ it newline i
| ' To understand Consider [ taue or]
wher Ss
'
~ WS Collection
"4 enamples — inebudks
F posite ad negative excmpley + e
~ (Yu) 14, Vy - (uy ) de
Eres (44 5 ) =
]
‘be (s/ )
~ e940
eee cee eee
NAC Grhepy is few) 1p
Enbo py iS Tne) wshes dha Conta aq excl no. £ PSitive
1
all members FS Lelong be A Game clay
Od necahire
7 TE He S cont Sere OS os Pest ad Agjalee ex’, jae
| Emery IS behead \. -
~ More generally, A doret athitate ean fore on ¢ dip Wali,
Ahem He enhepy
11 $ Sb Ry Cute Shem ety is
Exbora(s) = 7 “Py bog, r
there Pi iS he Propatin gs Ce te claw 4
Informer Gain
ee
(Bain) «
iS fw eXpecdkedt reduction in coheyy
comnd by Pevtitionng Abe
Srtampley According 4, Abn obit. -Sv
Gain (5,8) BE Entropy (5) — F_ vs Entvopy(Sy)
ve VeluesCay
Ee Asm Ss Collection @ Me Crcempler (a+s-]
SOROS atbinte tind have 2 vols je Lean endl strong
a (4+, 5. J
Sweag = [es, 2-J
Pale
a Isv| ea
Gain(S, und )= Exoty(S) ~ > Te ota(S,)
eS teas shag
= Eniery (S) — i BrbeP9 (Suen) ~ € St Sony)
OHO (Yq) o21 (ig) 0
= C04g a
~~ Ci) 4.4%) -(%) la&%
~ Gan
(Fa 33] shows tbr Gan F Hmiding and Wind attitdes 4
determine wich & better athikube F clerifying. the Leeining CXerm plon
hie. 403 determines He mEr opin fy each otbinte then
one wh Kees opin value
Gains, aikleck) = 01246 W~ (ifort vale )
rans, Huity ) = OS]
Sin 5, wind ) = bo4g
Sain (3 Fearpersbn) = 0025
owtlcer is selected a AL decision atbibute fa Re rect node
and branches an creakect fy each of its possible clues ¢
Sommy , overemst , vein) 4) Shoon in [Fa a4 |
* — ontloee = Sumy ] ok ae
outlook, = ore Sill have won-zew enbopy. | heme Ae” DT
wil be further elaboraked ay shun in (Fa % |
selects yy
CY Aly pothesis. SPS WM DTL (decision Treelearni )
LE [Decision Tree Lenaning
The bye Syre Semrchet & 103 ‘s
* ee seh PANE Recisin trees.
+ 13 (ert a Siangle - te complen bill climbing search
Horcusle Mis hyqothe sis space fing cSt, fhe
crnths tree hon emwadering Prgrersively mac eloberake
Wyethess tm Search ot a py ay
coveectly lena Pies ee te ty
Using, inf mek ot BM Measure,
Capabilities y Limittion op ID3,
BH 9s's Irypedhe sis Syace Fall DT is a Complete Space ¢ finite
Atscrete-valid tas A aveilatte otbittey
~ Trails He map 94
© 103 mointosing onl, « Single ey
Ae space ¢ DT.
Yow
SKF Search Weanplete hyp Spaces.
mre” hyp. as it seraches Aorensy
Ty dees't have Ae ability Ae determing
mong allernchve DT an Gusset with de awilate Yay ales© 103 ts Pure foo Performs he back va ching: iw tts Search.
This focably optimal Solution may be Lem desirable Ahoy bees
Hoe csould bese been encoutenesl hag diff branch “f be Searls. |
(D3 uses all Keating exompler at each Steg Mm de Search bo
make gets Siabistically bared Becisions «
103 Cam be ently erlendid ft handle netsy erasing ee
+ Muck Lets sensitive t2 evrers sbule Seancling
Gy) Selective Bios OTL
+ Seductive pian is the Set of assumptions cht fo gti Wit
| Areawing, date , deducbively Jwbity tle clamiticolions eavigneel
&y te kearner te Lubin inskeamees.
We aly cheeses te fir actegtalle tree ik encounters imide
Siew fle fe compen | bill climbing Search Awouge the space
posible brees-
Loa Search Shrabeay is
() Selects dn tawr of shorter bees over longer ones
() Selecks trees tuk place dhe atlricates with higheat inPamta,
Spin closest 4© the yoot -
Preterome bias Vs Restriction bins +
she inductive bias 103 is Ha a preference. fr certain hyp.
oer ders ( €g: 42 shoter We) calls a preference bias -
Preferene bias Is we desirable dhan a yestrtation Las , beraue
1k allows the Learner f wort widhin A Comyle te lryp- Spat °
“Ge continst, restriction ies Alot Limits strictly Ake Set frtential hyp
is qenerally lets dlesiva ble cE
prefer de srptat byp- Abst fits 4 data) Focuen in Decision Tee Learning (5)
1s OFFI be dato
Le Guarding aginst
Hemndling Continuous valued ablrintes
& Welling missing cdbviute values
S Herdbiog cablributes wth differing, costs:
bak abritute chorees
Ue over fitting tha data
Definitions Givens hyp-gpace Ho hyp HEN is Said te overfit
dhe Areinivg doka If dhere exists Some alternative hyp- hie}
Sach Abst W has sweller ever jhon hover tle foinry temps
buk h* hes swuller ervey evra than h over Jte entire
Aishribukion Iwutemees ( fe tmclading slammer Leyord le traning set )
Hes Con me prevent overfitling 7
> dott dey fe fi all eremples, Stog befre Ae Kraining Bet i exhaust
: et
FE alk examples then Prune Re vesltont tree.
2 Eruanding egeinvh bad cbtvibute choices
~The nf gain measured les abi Mat Fouors aityitutes
With any values over Ahebe wilh ondy 4 few
+ GRr Date allrble sth Unie Valuer fn each traning €X
+ gbviowly no cher edttitute can cl better ,
+ This ot peylt in a very Sroad bee depth |
eats: Gusad
GaimBodio 6A) =
inst ALIS, we Can WE GainRekio (5A) ited ¢
Gain 6A) manent Gain( < Ng 2) >
FIGURE 11
Final design of the checkers leaming program
Determine Type
of Training Experience
Games against
expen Table of eotec
value
Determine Representation
‘of Leamed Function
Potynomist
(Lincar function Arifcial neural
‘fem fearures network
Determine
Learning Algorithon
Linear
“Grasiont
sat \\ programming
Completed Design
FIGURE 12
Summary of choices in designing the checkers learning program.Example Sky ArTemp Humidity Wind Water Forecast’ EnjoySport
1 Sumy Warm Normal Suong Warm = Same Yes
2 Sumy Warm High —Srong Warm Same Yes
3 Rainy Cold. “High Strong Warm Change -——_(No
4 Sumy Warm High Strong Cool Change Yes.
TABLE 21
Positive and negative training examples for the target concept EnjoySport.
+ Given: -
‘@ Instances : Possible days, each described by the attributes
‘Sky (with possible Values Sunny, Cloudy, and Rainy),
AirTemp (with values Warm and Cold),
Humidity (with values Norma! and High),
Wind (with values Strong and Weak),
Water (with values Warm and Coo!) and
Forecast (with values Same ard Change).
1 Hypotheses #7: Each hypothesis is described by a conjunction of constraints on the at-
tibutes Sky, AirTemp, Humidity, Wind, Water, and Forecast. The constraints may be “?”
ny value is sccepible), “O" (no value is acceptable) or a specific value.
4 Target concept e: EnjoySport : X ~ (0,1)
‘© Training examples : Positive and negative examples of the target function (ee Table 2.1)
© Determine:
‘© A hypothesis fin 2 such that h(x) = e(x) forall x in X.
seoee
TABLE 22
‘The Enjonsport concept learing tsk.
1, Initialize 4 to the most specific hypothesis in AT
2. For each positive training instance x
© For each attribute constrain: a in h
Ifthe constraint a; is satisfied by
‘Then do nothing
Ble replace a in h by the nett more general constraint that is satisfied by x
3. Output hypothesis
TABLE 2.3
Finv-S Algorithm.
The List-THex-ELIMIvare Algorithm
L. Version Space « a list containing every hypothesis in H
2. For each training example, (x,c(x))
remove from VersionSpace any hypethesis for which h(x) x e(x)
3. Output the list of hypotheses in VersionSpace
TABLE 2.4
‘The List-THEN-ELIMuNATE algorithm,Instances X Hypotheses Ht
2 ow lt
ah a
- AON
j= Sin: Warm, High ght Warm Same ye Sunn 222
FIGURE 2.1
Instances X
Sree
y= <2.2,2,0,2,0>
x © «Say Warm NomalSrong Worm Same>, «hy =
Suey arm Hick Srons Warm Sme>, hy
Rain Cold High Strong Warm Change>,- hy =
-,+ hy =
[}
_— |
2, Strong, 2, ?> — <2, Warm, ?, Siromg, 2, ?>
NAN
(eSuamy, 2,7, 2.2, 2>,<2Warm, 7,22)
FIGURE 2359: |{<0,0,0,2,0,071
{ }
I
Sq: [ ()
Gy G:[ (RnR RAID
‘Training examples:
1. , Enjoy Sport
2. , Enjoy Sport
FIGURE 24
Sy 32 | }
Gx | {, EnjoySport=No
FIGURE 25
CaNDIDATE-ELIMINATION Trace 2. ‘Training example 3 is a negative example that forces the G2
boundary to be specialized to G3. Note several alternative maximally general hypotheses ace included
in Ga,5: [[)]
Ph
G4, [[)
I
G3: [[eSunny, 2, 2, 2,.2.2> <2, Warm, 2,22, 2> <2.2% 2%? Same>}
Training Example:
4., EnjoySpori = Yes
FIGURE 26
< , Warm, 2, Strong, ?. ?>
NON
FIGURE 27
{)
‘The final version space for the EnjoySport concept learning problem and training examples described
earlier.
Instance Sty AirTemp — Humidity — Wind
A Sumy — Warm Normal Strong
B Rainy Cold. ~—Normal—_Light
C Sumy = Warm = Normal Light
D Sumy Cold. Normal Suong
TABLE 26
New instances to be classified.
Water
Warm
Warm
Ferecast EnjoxSport
Change
‘Same
Same
‘SameTndactive system
‘Classification of
Taig enampicr | Etanson “fonttknon?
Algorithm ee
‘New imtance__,| Using Hypothesis
Space
Equivaeat deductive system
| Ciasatteaon of
Toning examples
‘Theorem Prover
New instance
Assertion “Heontatns
the target concept
i
/
Inductive bias
‘made explicit
FIGURE 28
3CExamptes, Targer-atiribute, Attributes)
‘Examples are the training examples. Target.attibute is the atribute whose value is 10 be
predicted by the tree. Attributes isa list of osher attributes that may be tested by the learned
decision tree, Returns a decision tre that correctly classifies the given Examples.
1 Create a Root node for the tree
‘ Ifall Ezamples are postive, Return the single-node tee Root, with label = +
‘« If all Examples are negative, Return the single-node tree Koot, with label = ~
1+ If Artrioutes is empty, Return the single-node trve Root, with label = most common value of
Target.attribute in Examples
«© Otherwise Begin
(2.4 = the atribute from Attributes that bost"classifics Examples
The decision attribute for Root < A
«For ezch possible valu, vj, of A,
12 Adi a new tre branch below Root, corresponding to the test A = vs
«© Let Examplesy, be the subset of Examples that have value vy for A
eo 1f Examples, is empty
‘© Then below this new branch add a leaf node with label
value of Target.attribure in Examples
+ Else below this new brasch add the subtree
D3 Examples, Target-attribute, Attributes —(A}))
‘most common
© End
‘© Return Root
*"The best attribute is the one with highest information gain, 2s defined in Equation (3.4).
TABLE 3.1Outlook Temperature Humidity Wind —PlayTennis
Di Sunny Hot High Weak No
D2 Sumy Hot Strong No.
D3 Overcast. = Hot Weak Yes,
D4 Rein Mild Weak Yes
DS Rein Coot Weak Yes
De Rain Coot Strong No.
D7 Overcast. Cool Strong Yes
Ds “Sunny Mild Weak No
D9 Sunny Coot Weak Yes
DIO Rain Mild Weak Yes
DIL Sunny Mild Strong Yes
DI2- Overcast Mild Strong Yes
DI3. Overcast Hot Weak Yes
Did Rein Mild Strong \
TABLE 32
Training examples for the target concept PlayTennis.
Sted ses
E0940 50910
amity wn
oka (6124 13034
098s B08 E100
ame (S. Hamidty ) Gain 8 Wind)
‘49 -criansas -c.e = s9-caneatt igi
= 949 -cay9n3 7
Whick attrBoute should be tested here?
Sunny = (BLD2.DB.DSDL
Gain Sumy Hain)
70 - GI5)00 -2/5)00 = 970
Gain Sunny. Temperature) = 97) - @1S) 09 - (IS) 1.0 - (U5) 00 = 500
Gain Sumy: Wind) = 910 ~ 28) 1.0 ~ (3) 918 = 019
FIGURE 34