NTRODuCT naN TO MACH1NE LERNIN4
DECIS QN TREE Lc9eNINE
ART-A INRODUCIION TO AcHLNE Lcn2NINe]
|lap1c- 1:
WeLpased leQniogproblems
Designi a learning Probiems learning
1$sUes trn Machne
P'etspectives and
ANSHCRS
platn about WELL-p 05ED leatning
Ex
problems.
|4) ACtNE LEARNING A computer program
E Wrtn
1S Sad to learo rom expeience
and
Tespect t0 some class of t asKsCT)
|perfotmance Cp), f itS pcformance at tas
0as neasuied (P), mpraYes wrLn Expei
ence E
ExnnplCS,
Acheckes leaofng problem
Vlask : playing checkers
performance P: PerceDt 9f 9ames wO
X aganist apponents
3 Irain1ng i Play'ng piact ice
expelence qames agansl I sef
AtLnNDNRIILNG RrcasNLTION. 1En2N1NH
PROBLCM
iaskC): Re caan ying ood classifyn3
w|thio
hand wtttco Words
Images
PerfQt mance Cp): percent or wods Corsecty
class ied
raroing Experience CE): 4 séguenceof
2 database 0f handuor LEen
wOr ds wth the given class1-fications.
YA RaBor DRIING, LneNING lROBIE
lask(T: Diirng on pUblic four -lane high
Ways USing vision SenSOS
t
performance Cp): Average distance trove
Ied before dn etror
3 ltaDiog Experi ence CE): ASequence
|O1 mages nd Steeing Cammands reCOH ded
while ohserving human crirer
3: PerspecHiYes g Issyes in MyINE
LEARNING Explatn
4) Perspective af machine lecu ong lovglves
Searc hng ver| large Space f poss1b)e
nypothesis to det emune one that best
rIES the gbserred data nd Qny pior
Knowledge held by learne r
MacNE LERNING
what alartbms shaulc.be
Used
Which alqorhms pcroms best for which
types of a poblem
tranng. and estingdatCL
lau much
methods should beused
. whgt Kind of
should be used to tedce
S What methods
leaoOg achieyements
data which methaS
. tar wbich Lupe oc
Shoutd be USed
s4 Hhat are the applicaLIon of MAcI1
MAHLNE
LEARN NG
|4)
earning to ecagnze Spaken woids
Learnng to dve an qutonomgu S
vechicle
3Learong to classrf new astionomral
SttucLures
LeaO0g to plGy wo ld chess backq -
loan applicatioos
fiognacial Liarn scqLIOns
gnedical iecoids
G Explatn the Concept ofa designing C
Learnng system
esigning
In Order ta Lloder SLond abaut
aLEARNLNG Sys IEM let S- Under SLand
T Lo leatO
eTample te A Prgg ram to
through an eIample e
enteri ngE
tO Play cheCKers,wwrth goal of
IoURNAMENT
Tn the WORLD Ct ECKERS.
foLLOWENG ARE STEPS
| Chogsing the trainLOg Experience
Chogsing the lARGEL EUNcTLON
RepreseniatIon for the Targee
Y hogS Inq a
functIOn
funcIOn 4 Pproximctrgn
| . ChggsLOg g
AlgrILhm,
|IstímatLog Irainiog ralues
he fnal design
Iet uS IsCUSS an e by One
CHagSt NG HE TRAINING ExptRIENCE
avalabe
Ihe type of ttarotng ecpetience
Can have SigoIfiCant Impact Gn SucceSS
faulure Gf the learoer. Heeloneof theattibU
of t a i n gepetrece iS
A/DrRECT FEEDB4CK
IE Chagsed de pendroq on performance
a System
DrRECT; or ccample a learning tOpiey
duect
checkers, the System ght learnfrom
|LIatntng eCamples consISt Of Indvd
ChecKers bgard LOtes nc Lhe CaneCt
fo, each, SomcONE Ls i#ERE | (Spo ntaneUSJ
JNOIRECI Indtrect insarmatign onsiStirg
of the move seqUences and ftnal gutcorn e3
|S Variousqomes plaued. tletc Out one
ay be wgn or L ast B u learner fore J
Problem is
Cte dit asSiGnement Can be aport tculry
diffcUIt problem beca use the qome o b
lost even when n O
e ab early nores Gie OpIInL
Lhese are follo uwed latter by poor mgre
CNO one isthere JLNOt SpontGneoUS3
KaNCLUSION, Henceieorning fron direcE
Lraining feedback it t4picoliy eoS1er tbar
learning fro m Indrect feed baCIe
SEcOND ÎMpGRTANT ArrRiBureS
A SecOnd important attribute of thetrainig
experience is degree 10 wtbch leaner CONr2
SEqUENCE Of traiono examples
A Lhrd mportant attrbute Of trarnn g
e perience 1s how )el|TE re presents the
d I n b l o gE eamples arer whrch tne
-final perrormanCe P muSt be measUTe d
Cre different poss1 ble combingtIon I
Io Praceed wth Cur desiqo, let us ecide
Lhat GUr 1cm) )|| Lraio by play i g9mnes
aqanst TLsetf Ihts has lhe clvant age
Lhat hos no
O be present an tt theeOrC
Q-2)
Suustem to generatc as much tranoq data as
permAS
CtIOSING H LaRGET fuNC TLON
1ne nert deSI qn chgice, is Lo determine
What tpe oc Kooole dae w! belearned
|and how this wl be used by the perfOrn
Cin epigamChe it ís userol or ngt)
ExAnpLE checker game
what
EGAL nov ES 1e let us Un derstand
` checker cames (§ X8Jboard
2
+o4 5
2 Peces
C
2 PIeces
lyraret only
in forward
diectIOn
R/Onty One move
per chance
3/only tndiag
Onql ectic)
tJump orer
4 Kled
legalmoves
Hhle mare diaqanally Set of all poss1ble
moreS iS Called
LEGOLHOYes
Ie one mave(Lor RI
ARGE Move
Let the Iagek funcL Ian vCb)
Board State :b fle oH al Iddie Orfina
Set of Legal moves B
"Here Tarqet fns nothing but what g
of Knowledge did machtnelearn, whether
that kngwledge is userulfor implemen
ation or not.
No let us duscUSs abiut b°.
ie how to select tarqe-function.
4ASE-1: f b is the fogl board Stqte
that.t's won,then veb) =I00
{ ASE-2 Tf b is fnal bogrd stqte th t
1s lasEtheo VCh)-I00
ASE 3 içb 15 the fiogl bOardstat e
that i's diawthed vcb) -
CASE 4 Tf
Jf bis not fingt sLQte thne
Stqte
isthe
best -fira
Stae
chaoSing a Re present ation for
Lhe iarget funcLiOn
(ie lagstng r wining or
baard Slale
norma), we calculale funcion c as inea r
Cambnation f follawtng bgard fe qAGr es
Ie cCb).
tEATURES
bdard
| No of black pieces On bgard
| Ng 0f red preces Go
Gn baard
3 Ng Of black Kings
boqrd.
X4 NO Of ed kings On
threqed byied
5: N o of black preces begteobyied )
Cblacks which can be
threqtecl by black
DG : No of red pieces
Uepresenigt1oO of Cb))
i f , 3:means^blacle
|e
4ddLYe COnS
CGnstant s y Omeans zero
Feed cOnS'
YChaasing
AlqorLhni tHere we aie choosng
learning
Alqorrthm: As we Knqw that ror
a LageifuncLIan Cf)
uWe need a se t of
ttauning exa ples Ie we need piedefined
data. This tiatong CTamples defines
Or gives ihfomat ion about baard State
Cund tIining value Ie b) 5 Vtab)
te Spective ly.
In geneIql tiarOng erample can be
defined as
Ordered
parr = (b, VttainCb) )
ExAmpLE:
Let black wan the qame e x=O,which
neanS no red te X4
red (e Xa =0,X50, s =0)Xj=3
X3 =3
ttaetE
VtraroCH)=+Io0 [Won it i's tiao J
<X-3,(,=©,3=,X4=9)NG-6X6=0)tioo7
He need 'to do two Steps in thephas
VESTIMATING IRA1NI NG VALUES
in eve4step, we conSider,suces sg r
Çdepending on Lhe next sLep of 0,PPO
nent):
nert boad StGLe
Hele Success ar tepreSents
tha his morewrll belp/
tepre5ens GPproI
destrOy oppane n ) e v
Ad Tusting the Hcights
-od we (ghts
alqor rthrn to
ihete are some Here we aie Using
near funtio ns
Ot SqUgre (-e Used o
LMs eCeast mean
etror.
Erro =(vCranb)-vcb)*
nced' to change the
etror=0 te NO
werqhES
weiqht ls IòCIease
If errgr IS pasire, each
1
proportiona lly wert is decte-
Negqt lve, each
Y f error is
aScd prOport ion ally
Ihi's can be done wr
L ns WELGHI UponTE RuLE
for each iatning erample b,anCb) 3
USe the cuIent we ighs to calcul ate
VCb) fdr each weight wj, Updaqte the
|w,wi+n(Vuainch) Vo)
n: Sma | constan t Ceg o)
: Error =0, 1e iaioCb)- ta)=oie n weigh
change, Same as above
5/ rNAL DESIGN
1Lngencral to inderstandfnal design
We have 4 Modu!es.
|PEREORNNCE SySTEM:
Critie,
3. Generahzer
Experimental Generattian
[tntjalg
b O
aor
m de
Expen meds
NepW
orblem
Genetato
)histL
o
performone
Sysiem
rqcê
Ga
y S
moelutian
(Gcoeraljzc
kh,Vtronch
<b3 ran(b
Crit'c
(y PERFORANCE SysTEM Here IE takes an
|ÒStOnce neo problem Cnew qome) aS input
cf
cnd poduce a tiace of its Sal vt IGngam
1StOr9) as oULpUE. In dur case Strate990e
by the performgnce Sutem to Select IL nea
iropoe
"ove qt cach step L` deerminecl te
There
by the leornec!V evairiIdO function.
1Ore ,we expect its perrormance o ImptoreS
becomes InCIea
S thS ealuation-untGO
Singly accurate
LRITIC i e crtc aKe game Gnd input
as-he ptoduces
af the.
the hstory Orttoace tratntng et QmpieS ,
aS GUtpyta Set of
Case Corresponds to Some game
in the
|State In the ttace algng wrth an eStimat
targe fonciion value forths
Mtrain Gf the
erample. 1e rule is
Vttaincb)< VSicessoi)
EN ERALIzER
ñe Generaljzer takes @s input -the trainin
leramples and produces an output hypothess
that is ILs estimate of target of hetiaget
funclon It generalizes from the SpeCf)C
tiaining erampies, hypothesiztng a genera
5unction that orets the Se eIamples 5 Gbr|
cases be yond the tratntng erampls i n
OUr eam ple he generalize Corresponds t0
the L s AlGORI TH1e Negh
Eoneps).
OR.
ExpERINEN T GENERAT
Generator takes as inpOE
ihe epeimentpothesis (CUurrent ly learned-f
the Curent hy problem (1e Intial
and OUL pUts a new Systcm
Performance
|boa rd SLQLe) for the newpiacuc¢
n role s t0 pclc
Ao ezplore
w maTmze the leqrOg
probiems t h q
OrerallSusLem Ion OUr eXImpl
|Tq te Of the Genergor folJ aws a
the Exper1mnent propases h4
SLiatcay. Tt aluuaL1s
very imple o beqin a n
lotIgl game baard
Same
ame
ENODT
fo p 1 C
