Upload
vu-thao
View
218
Download
0
Embed Size (px)
Citation preview
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 1/22
Marakas: Decision Support Systems, 2nd Edition © 2003, Prentice-Hall Capter ! - "
Chapter 9:
Machines That Can Learn
Decision Support Systems in the
21st
Century , 2nd
Edition by George M. Marakas
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 2/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 2
9-1: Fuzzy Logic and
Linguistic Ambiguity(ur 'anguage is rep'ete )ith *ague and
iprecise concepts, and a''o)s +or
con*eyance o+ eaning through seanticapproiations.
These approiations are use+u' to huans,
but do not readi'y 'end these'*es to the ru'e%
based reasoning done on coputers.
-se o+ +uy 'ogic is ho) coputers hand'e
this abiguity.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 3/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % #
The Basics of Fuzzy Logic
/n a 0pure1 'ogica' coparison, the resu't is
either +a'se "3 or true 43 and can be stored
in a binary +ashion.The resu'ts o+ a +uy 'ogic operation range
+ro " abso'ute'y +a'se3 to 4 abso'ute'y true3,
)ith stops in bet)een.
These operations uti'ie +unctions that assign
a degree o+ 0ebership1 in a set.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 4/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 5
A Simple Membeship Function !"ample
The #Tallness$ function ta%es a peson&s height and
con'ets it to a numeical scale fom ( to 1)*ee the statement #*e is Tall$ is absolutely false
fo heights belo+ , feet and absolutely tue foheights abo'e feet
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 5/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 6
Fuzziness .esus /obability
There are soe subt'e di++erences:
$robabi'ity dea's )ith the 'ike'ihood thatsoething has a particu'ar property.
7uy 'ogic dea's )ith the degree to )hichthe property is present. 7or eap'e, aperson 8 +eet in height has a .6 degree o+ta''ness.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 6/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 8
Ad'antages and Limitations
of Fuzzy Logic Advantages: +uy 'ogic a''o)s +or the ode'ingand inc'usion o+ contradiction in a kno)'edgebase. /t a'so increases the syste autonoy
the ru'es in the kno)'edge base +unctionindependent o+ each other3.
Disadvantages: /n a high'y cop'e syste,use o+ +uy 'ogic ay becoe an obstac'e to
the *eri+ication o+ syste re'iabi'ity. 'so, +uyreasoning echaniss cannot 'earn +ro theiristakes.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 7/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 %
9-0: Atificial eual et+o%s
7irst proposed in 495"s as an attept to
siu'ate the huan brain;s cogniti*e 'earning
processes.
They ha*e abi'ity to ode' cop'e, yetpoor'y understood prob'es.
<<s are sip'e coputer%based progras
)hose +unction is to ode' a prob'e spacebased on tria' and error.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 8/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % =
Leaning Fom !"peience
The process is:
4. piece o+ data is presented to a neura'net. The << 0guesses1 an output.
2. The prediction is copared )ith the actua'or correct *a'ue. /+ the guess )as correct,no action is taken.
#. n incorrect guess causes the << to
eaine itse'+ to deterine )hichparaeters to ad>ust.
5. nother piece o+ data is presented and theprocess is repeated.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 9/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 9
Fundamentals of eual 2omputing
The basic processing e'eent in the huanner*ous syste is the neuron. <et)orks o+these interconnected ce''s recei*e in+oration
+ro sensors in the eye, ear, etc./n+oration recei*ed by a neuron )i'' eitherecite it and it )i'' pass a essage a'ong thenet)ork3 or )i'' inhibit it suppressing
in+oration +'o)3.Sensiti*ity can change )ith passing o+ tie orgaining o+ eperience.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 10/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 4"
/utting a Bain in a Bo"
n << is coposed o+ three basic 'ayers:
4. The input 'ayer recei*es the data
2. The interna' or hidden 'ayer processes the data.
#. The output 'ayer re'ays the +ina' resu't o+ the net.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 11/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 44
3nside the euode
The neurode usua''y has u'tip'e inputs, each
input )ith its o)n )eight or iportance.
bias input can be used to ap'i+y the
output.The state +unction conso'idates the )eights o+
the *arious inputs into a sing'e *a'ue.
The trans+er +unction processes this state*a'ue and akes the output.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 12/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 42
Taining the Atificial eual et+o%
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 13/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 4#
Sending the et to School:
Leaning /aadigms/n unsuper*ised 'earning paradigs, the <<
recei*es input data but not any +eedback about
desired resu'ts. /t de*e'ops c'usters o+ thetraining records based on data sii'arities.
/n a super*ised 'earning paradig, the <<
gets to copare its guess to +eedback
containing the desired resu'ts. The ostcoon o+ these is back propagation, )hich
does the coparison )ith s?uared errors.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 14/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 45
Benefits Associated +ith eual
2omputing *oidance o+ ep'icit prograing
@educed need +or eperts
<<s are adaptab'e to changed inputs
<o need +or re+ined kno)'edge base <<s are dynaic and ipro*e )ith use
b'e to process erroneous or incop'ete data
''o)s +or genera'iation +ro speci+ic in+o ''o)s inc'usion o+ coon sense into theprob'e%so'*ing doain
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 15/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 46
Limitations Associated +ith
eual 2omputing <<s cannot 0ep'ain1 their in+erence
The 0b'ack bo1 nature akes accountabi'ity
and re'iabi'ity issues di++icu't
@epetiti*e training process is tie consuing
&igh'y ski''ed achine 'earning ana'ysts and
designers are sti'' a scare resource
<< techno'ogy pushes the 'iits o+ currenthard)are
<< re?uire 0+aith1 be iparted to the output
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 16/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 48
9-4: 5enetic Algoithms and
5enetically !'ol'ed et+o%s/+ a prob'e has any so'ution, it suggests that
there is an optia' so'ution soe)here.
The +ie'd o+ anageent science has been
ab'e to tack'e increasing'y cop'e prob'esand +ind optia' so'utions.
This success 'eads us to tack'e prob'es
e*en ore cop'icated, creating a need +or
ore inno*ati*e so'ution ethods.
(ne such ethod is the genetic a'gorith.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 17/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 4
3ntoduction to 5enetic Algoithms
Like neura' nets, genetic a'goriths G3 are
based on bio'ogica' theory.
&ere, ho)e*er, Gs +ind their roots in the
e*o'utionary theories o+ natura' se'ection andadaptation.
The po)er o+ a G resu'ts +ro the ating o+
t)o popu'ation ebers to produce o++spring
that are soeties better than the parents.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 18/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 4=
Basic 2omponents of a
5enetic Algoithm
The sa''est units o+ in+oration are dubbed
genes, )hich cobine into chroosoes.
+ter a G is initia'ied, it uses a 0+itness
+unction1 to e*a'uate each chroosoe.The G then eperients by cobining the ost
+it chroosoes.
<et, the crosso*er phase sees these t)o 0good1chroosoes echange gene in+oration.
The utated chroosoes then >oin the poo'.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 19/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 49
Basic /ocess Flo+ of a
5enetic Algoithm
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 20/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 2"
Benefits and Limitations Associated 6ith 5As
$opu'ation sie is a critica' +actor in the speed o++inding a so'ution, but at 'east it is re'ati*e'y easy topredict this speed.
Crosso*er and utation are interesting ideas, but
they shou'd not be used too +re?uent'y or toosparing'y, either3.
(ne ad*antage is that you are a')ays guaranteed tocoe up )ith at 'east a 0reasonab'e1 so'ution.
Ae can a'so app'y the to prob'es +or )hich )erea''y ha*e no c'ue on ho) to so'*e.
7ina''y, their po)er coes +ro sip'e concepts, not+ro a cop'icated a'gorithic procedure.
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 21/22
Marakas: Decision Support Systes, 2nd Edition ! 2""#, $rentice%&a''
Chapter 9 % 24
9-7: Applications of Machines That Lean
<ippon Stee': b'ast +urnace contro' systethat uses <<s
Dai)a Securities and <EC: stock price chart
pattern recognitionMitsubishi E'ectric: neura' net and optica'scanning to recognie tet
<ippon (i': neura' net used +or diagnosis o+
pup *ibrationCredit scoring on 'oan app'ications, both toindi*idua's and corporations
7/24/2019 Marakas-Ch09
http://slidepdf.com/reader/full/marakas-ch09 22/22
Marakas: Decision Support Systes, 2nd Edition !2""# $ ti & ''
Chapter 9 % 22
The Futue of Machine Leaning
'ready, arti+icia' neura' nets eceed huan
capacity +or iso'ated instances.
Theoretica''y, a coputer can process datai''ion ties +aster than a huan.
7ortunate'y +or us, huans are so uch
better at ac?uiring data. Coputers >ust don;t
ha*e anything 'ike the +i*e senses.