Beruflich Dokumente
Kultur Dokumente
ACOE
402 Neural Networks and Fuzzy Logic
ArtificialNeuralNetworks
EfthyvoulosC.Kyriacou(PhD)Assoc.Prof.
ComputerScienceandEngineeringDepartment
Resources:
Dr.Neocleous Costas,CyprusUniv.ofTechnology
NeuralNetworks,S.Haykin,Prenticehall1999
FundamentalsofNeuralNetworks,L.Fausett,Prentice hall1994
ComputationalIntelligence(CI)isaspecialcaseof
A ifi i l I lli
ArtificialIntelligence(AI)
(AI)
What is intelligence?
Whatisintelligence?
It is difficult to be clearly defined
Itisdifficulttobeclearlydefined.
It may not be needed to be defined to be understood
Itmaynotbeneededtobedefinedtobeunderstood.
Terman*suggeststhatitis
gg
thecapacityforabstractthought.
p y
g
Mostpsychologiststhough,agreethatitisthecapacityfor
effectiveadaptation toanenvironment, whichisdonethrough
changeintheorganism,orchangeintheenvironment,oreven
creationofanewenvironment.
ti
f
i
t
__________________________________________________________________________________________________
*TermanL.(1916).Theusesofintelligencetests.nThemeasurementofintelligence.Boston,HoughtonMifflin. 2
Hereisamorepracticalsuggestion:
Intelligenceisthegeneralmentalability involvedin
processessuchascalculation,reasoning,
h
l l ti
i
classification,learning,usinglanguage,
understandingtheenvironment,selfcorrecting,
d
d
h
lf
inventingandadjustingtonewsituations.
ArtificialIntelligencethenishumancreated,non
g
,
biologicalintelligence.
ComputationalIntelligenceisintelligencethatemerges
through some form of computation
throughsomeformofcomputation.
3
AIandCImayleadtosolutionsofdifficult
y
problemsinscienceandengineering.
Theyhavealsoinstigatedarevitalizationof
fundamental philosophical questions
fundamentalphilosophicalquestions.
Theyalsogeneratednewfrontiersofexplorations
They
also generated new frontiers of explorations
andintroducednewquestions.
What is consciousness?
Whatisconsciousness?
Can universal consciousness exist?
Canuniversalconsciousnessexist?
How does creativity appear?
Howdoescreativityappear?
How are illusions love hate generated?
Howareillusions,love,hate,generated?
Can matter generate cognition ? And how?
Canmattergeneratecognition?Andhow?
Wh t i
Whatismemory?
?
Can there be olle ti e memor (or national) memor ?
Cantherebecollectivememory(ornational)memory?
How are numbers words meaning codified?
Howarenumbers,words,meaningcodified?
How does the brain process logic?
Howdoesthebrainprocesslogic?
Howcanthebrainhandlewitheasebothcrispandfuzzy
How
can the brain handle with ease both crisp and fuzzy
logic?
IbelievethattheArtificialNeuralNetworks are
presentlythemostpromisingtechnologythatmay
lead to nonprogrammed,
leadtonon
programmed,intelligentartificial
intelligent artificial
systems.
Artificial Intelligence
ArtificialIntelligence
Weak
approach
Humanknowledgeand
specificexpressionsofitcan
besimulatedthrough
computationalsystems
Strong
approach
Humanknowledgeand
specificexpressionsofit,
emergenaturallyin
computationalsystems
ClassicAI (symbolicAI)
ConnectionistAI
Itisexpressedthroughsymbolicentities
thatmaybeproperlycoded
Expressedthroughnonsymbolicmethods such
astheANN.Thevariouscognitiveprocesses
emergenaturallyfromthedynamically
connectedcells,astheyevolvewhen
influencedfromanenvironment.
Someimportantproponentsare:
Dennet,Newell and Simon,Chomski,Minsky,
Fodor and Pylyshyn
Fodorand
Someimportantproponentsare:
Smolenskyand Hameroff.
10
MaincharacteristicsofExpertSystems:
p
y
Theprogrammingandtheconclusionsarebasedonrules.
Theyarenoteasilymadetoleranttomistakes.
Iftheknowledgebasechanges,reprogrammingmaybe
If
the knowledge base changes reprogramming may be
needed.
Explanationofresultsiseasiertoimplementwiththese
systems.
12
Othernames fortheArtificialNeuralNetworks
Artificial
A
tifi i l Neural
N
l Systems
S t
Parallel Distribution Processing Systems
Connectionist Systems
Neurocomputing Systems
Adaptive Networks
Associative Networks
Collective Computation Systems
NeuromorphicSystems
14
TheANNsarebasedontheefforttomimicthe
operationalpropertiesofthenatural(biological)neural
(
)
networks.
Theyarecomposedofmanyartificialneurons
connectedinanorganizedsysteminwhichthereis
directorindirectcommunicationandinteraction
amongallitsmembers.
Thereisusuallyprovisionforinformationinput andfor
the desired output
thedesiredoutput.
Groups of neurons may organized into layers orslabs.
Groupsofneuronsmayorganizedintolayers
or slabs
15
AnArtificialNeuralNetwork
An emerging
A
i global
l b l behaviormayappearthroughtheuseof
b h i
th
h th
f
simple,andlocal learningandevolutionrules.
Artificial neural network
Artificialneuralnetwork
Processingelement
SYSTEM
SYSTEM
INPUT
SYSTEM
SYSTEM
OUTPUT
N1
N4
N6
N3
N2
N5
N7
N8
Signalflow
16
Togainabetterunderstandingoftheirfunctioning,letushaveabrief
encyclopaediclookonthebiologicalneuralnetworks.
Synapse
Typical
biological
neuron
Dendriteofreceivingneuron
xonofsendingneuron
Nucleus
Cellbodysoma
Internalvoltage 60to 80mV
Membrane
Thickness 5to10nm=0.000005to0.00001mm
Capacitance 0.944F/cm2
Fieldintensity 12000000V/m=120kV/cm
AXONHILLOCK
INITIALSEGMENT
TRIGGERZONE
NODEOFRANVIER
yelinsheath
xonornervefiber
Diameter:0.5 22minvertebrate
500 1000minthesquidgiantaxon
DENDRITEOFRECEIVINGNEURON
17
INFORMATIONTRANSMISSION
(Generation of the Action Potential)
(GenerationoftheActionPotential)
Excitationfrom
theenvironment
Habituatio
n
Stimulus
Sensory
receptor
Time
Synapse
Sy
apse
mV
Thegradedpotential is
GradedPotential graduallyweakened asitis
propagated towards the soma
propagatedtowardsthesoma.
Time
Neuron
mV
ActionPotentiall
Time
Anotherneuron
Thereceptivesensors convert
changesintheenvironment
(light pressure chemical
(light,pressure,chemical
constituency,...) inGraded
Potential.
Manygradedpotentialsfrom
differentdendritesare
accumulatedtocreateabigger
orsmallerone,which,onceit
reachesathreshold(40mV),
it becomes an Action Potential
itbecomesanActionPotential
(atrainofpulsedvoltages,
nerveimpulses orspikes)
18
Synapse
Diameter 1m=0.001mm
Gap20to40nm
=0.00002to0.00004mm
Diameter/Gapratio 100
Delayinthetransmissionofthe
presynapticpotentialtoapost
synapticpotential 0.3to1.0
ms
Velocity 0.2cm/minute
19
Themediumforthesignaltransmissionisbasicallyof
electrochemicalnature.
Thereexistsagreatvarietyofsynapses.
Eventhepositionofasynapsemakesadifferenceinsignal
transmission.
transmission
y p
Thetransmissionofcodedinformationatthesynapsesis
primarilydonewithchemicalsubstancesknownas
neurotransmitters [e.g. acetylcholine,norepinephrine,
d
dopamine,5hydroxytryptamine
h d
serotonin,aminobutyricacid
b
d
GABA].
Thesephenomenaaremainlyoccurringinmammalnervous
systems.
y
20
Comparisons
Thereareabout1010 1012 neuronsinthebrain
andabout1013 1016 synapses.
Organism
leech
Worm
Fly
Cockroach
Bee
Man
Numberof
Synapses
>104
>105
~109
<1011
>1011
~1014
21
Therealsoexistelectricalsynapses,buttheseoccurmainlyinlower
There
also exist electrical synapses, but these occur mainly in lower
animals.
Thesystematicuseofasynapseisbelievedto
improve its efficacy
improveitsefficacy.
learning,memory
g,
y
HebbsRule
22
23
= 60mV
= 50mV
= 77mV
= 54.402mV
54 402 V
= 120mmho/cm2
= 36mmho/cm2
= 0.3mmho/cm
=
0 3 mmho/cm2
E =V Vrest
25 E/10 1 ,
m =0.125 E/e
/ 25E/10
E/18
m =4eE/18
n =0.0110 E/e10E/10 1,
n =0.125eE/80
h =0.07eE/20,
h =1/1+e30E /10
24
duj/dt = uj uj 3/3 zj + x
d j/dt = k1j zj + k2j uj + k3j
dz
25
BRAIN NERVOUSSYSTEM
26
BRAIN NERVOUSSYSTEM
9
Thebiologicalneuralnetworksarenonlinear
computational systems thatarecharacterizedbyahigh
computationalsystems
that are characterized by a high
degreeofparallelism,robustness andfaulttolerance.
Theyshowacapacityforlearning,generalization,and
handling vagueness.
Itisestimatedthatthetotallengthofthebrain
It
i ti t d th t th t t l l th f th b i
connectionsareabout 109 meters,whichisabout25
times the earth perimeter!
timestheearthperimeter!
27
Eventhoughthebrainisonlyabout2%ofthetotal
weightofaperson,itneedsabout 15%ofthetotal
bloodandabout 25%oftheinhaledoxygen.
Livingorganismsthatdonothavenervoussystems
h d
h
(bacteria,protozoa,invertebrate)canexhibitsomeform
of behavior memory and even learning
ofbehavior,memoryandevenlearning.
Forinstance,the
For
instance, the paramecium (asingle
(a singlecell
cellorganism)
organism)
candemonstratelearningbehavior,eventhoughitdoes
nothavebrainandsynapses.
28
Everyneuroncommunicates
locally anddirectly withmany
other neurons and indirectly
otherneuronsandindirectly
withallothers.
LAYER
LAYER
Groupsofneuronsform
subsystems,layers,slabs,
y
, y ,
,
LAYER
Theinformationtransmissioninsynapsesisdoneinparallel.Becauseofthis,the
frequencyofchangesisabout1016 persecond.Takingintoconsiderationthefact
thatateachsynapsemanyneurotransmittersarealsotransferred,theinformation
transmissionisimpressive.
29
Comparisonofbiological
and artificial neurons and networks
andartificialneuronsandnetworks
BIOLOGICAL
NEURONS
ANDNETWORKS
ARTIFICIAL
NEURONS
ANDNETWORKS
Denseconnections
~ (1012 neurons)(10
~(10
)(104 synapses)=
)
=1016 connections
Fewconnections
Singleneuronsaredifferenttooneanother
Single
neurons are different to one another
Modularstructures
Autonomouslocalinteraction
Mostlysimilartooneanother
Mostly
similar to one another
Partlymodular
Nonautonomous
Usually supervision is needed
Usuallysupervisionisneeded
Mostlyserialprocessing
Muchenergyconsumption
Mostlymathematicaloralgorithmicdescription
y
g
p
Parallelprocessing
Verylittleenergyconsumption
Nonmathematical
Non
mathematicaloralgorithmicoperation
or algorithmic operation
30
ArtificialNeuralNetworks
Theyareflexible,nonlinearsystems
They
are flexible non linear systems composedofmany
composed of many
cooperatingprocessors,that,amongothers,helpin:
Studyofbiologicalneuralsystems.
Dataprocessingandknowledgemining.
Mappingandrelationextraction.
M
i
d l ti
t ti
Classification,Patternrecognition.
Forecasting.
Forecasting.
Studyofdynamicalsystems.
Adaptivesignalprocessing.
Automaticcontrolofsystems.
31
Itishopedthatthroughanautonomousevolutionof
It
is hoped that through an autonomous evolution of
suchconnectionistsystems,someformofintelligence
may emerge
mayemerge.
Will it lead to artificial brain?
Willitleadtoartificialbrain?
Artificial mind?
Artificialmind?
Artificial
Artificialconsciousness?
consciousness?
32
Artificialneurons
TheyarethebasicbuildingblocksofANNs.
They
are the basic building blocks of ANNs
Theyarealsoknownas:
Units or ProcessingElements
Theirmaincharacteristicsare:
Theyhavemultipleinputs oneoutput (MISO).
Theyarenonlinear.
They exhibit adaptivity
Theyexhibitadaptivity.
33
Theycanbeimplementedthroughappropriate
They
can be implemented through appropriate
software orhardware.
Thehardwareneuronscanbe:
Electronic,chemical,optical,micromechanical,
nanomechanical,...
nanomechanical
FFunctionally,theycanbesimulatedascellular
ti
ll th
b i l t d
ll l
automata.
Hereisageneralizedformofanartificialneuron.
34
Generalformofasingleneuronmodel
Inputinformationfromtheenvironment
orfromotherneurons
Outputtotheenvironmentortoother
neurons
PREACCUMULATOR
PROCESSING
(Subsystemoffunctional
anddynamicalprocessors)
(including cross
(includingcross
correlations),p1()
Main
Accumulator
POSTACCUMULATOR
PROCESSING
(Subsystemoffunctional
anddynamicalprocessors)
,p2()
DISTTRIBUTOR
Learningsystemthatadaptsthe
variousparameters
p
FFeedforward/Feedback
df
d/F db k
subsystem
h()
35
Someusefulpropertiesoftheartificialneuralnetworks
Theycangeneralize inthesensethattheyconcludeaboutanarea
of interest in which they hadnttbeenpreviouslytaught.
ofinterestinwhichtheyhadn
been previously taught
Theyshowrobustness. Theycontinuetobehaveacceptablywell
evenifsomeneuronsaredeleted,orsomeparametersmodified.
if
d l d
difi d
y
Theyarefaulttolerant.
Therulesareexpressedinadistributedprocessingand
representation manner.
manner
Theyusuallylearnthroughappropriatetraining,whichcanbe
guided(supervised)orunguided(unsupervised).
id d (
i d)
id d (
i d)
36
Structuring an ANN
StructuringanANN
Artificialneuralnetwork
Processingelement
SYSTEM
INPUT
SYSTEM
OUTPUT
TN1
TN 4
TN 6
TN 3
TN 2
TN 5
TN 7
TN 8
Signalflow
37
ANNTopologies,Architectures
NETWORKARCHITECTURES
STATICARCHITECTURES
DYNAMICARCHITECTURES
RANDOMCONNECTIONS
AUTONOMOUSLYADAPTED
ORGANIEDCONNECTIONS
EXTERNALLYGUIDED
ORGANIEDFEEDFORWARD
CONNECTIONS
(Withoutfeedback)
SINGLELAYER
TWOLAYER
THREELAYER
MULTI LAYER
ORGANIEDRECURRENT
CONNECTIONS
(With feedback)
(Withfeedback)
SINGLELAYER
TWOLAYER
THREELAYER
MULTI LAYER
CELLULARNETWORKS
38
SomeexamplesofANNarchitectures
Multilayerperceptron(MLP)
x1
y1 =sgn(u
= sgn(u1)
x2
xN
Weight,w
y2 =sgn(u2)
Forclarity,therestoftheweightsarenotshown
39
Multilayer,multislabnetwork
(hidden)
INPUT
OUTPUT
SLAB2
(hidden)
SLAB4
SLAB 5
SLAB5
SLAB1
(hidden)
SLAB3
40
Input
put
Vector
Weight
p(1)
Matrix
p1(1)
w(1a)
Receptive
p
Field
x1
.
.
.
xk
.
.
.
xK
f1
.
.
.
fk
.
.
.
fK
pI((1))
1(1)
.
.
.
i(1)
.
.
.
I((1))
p1(2)
.
.
.
pj(2)
.
.
.
pJ(2)
1(2)
.
.
.
j(2)
.
.
.
J(2)
.
.
.
IxK
p (1)
i
.
.
.
JxK
Weight
Matrix
w(2a)
fk =ake
(1)
p(2)
p(L) =w(La)f
i(1) =
j(2) =[pj(2)]
(2)
Weight
A(1)
Matrix
(1)
W(1b) A1
OxI
.
.
.
Ao(1)
.
.
.
AO((1))
OxJ
Weight
Matrix
W(2b)
A1(2)
.
.
.
Ao(2)
.
.
.
AO(2)
+
+
A1
.
.
.
Ao
.
.
.
AO
y1
.
.
.
yo
.
.
.
yO
A(2)
A(L) =W(Lb)(L)
41
Kohonenarchitectures
42
Recurrentarchitecture
43
Cellularneuralnetwork(CNN)architecture
Theartificialneurons
The
artificial neurons
communicateonlywiththeir
peripheral
peripheral.
44
Generalcommentsonneuralarchitectures
Itisnotnecessarytocodetheinformationprocessinginthe
y
p
g
formalismshowninthepreviousfigures/structures.
Wecoulduseanalgebraic/differentialformalism,orotherkind
We
could use an algebraic/differential formalism or other kind
ofgraphssuchas:
INPUT
HIDDENUNITS
OUTPUT
etc
45
LearninginANNs
Weightadaptation
x1()
Desiredoutput
w1()
d()
x2()
w2()
Error
u( ) = f xiwi
i =1
e()=d() u()
xN()
Comparison
wN(())
Thelearningalgorithmgovernsthewaythattheweightsadaptsothattheerrorgradually
The
learning algorithm governs the way that the weights adapt so that the error gradually
LEARNINGALGORITHM
becomesacceptablysmall.
46
LearninginANNs
Haykin (1994):
Learningisaprocessbywhichthefreeparameters(weights)ofa
neural network are adapted through a continuing process of
neuralnetworkareadaptedthroughacontinuingprocessof
stimulationbytheenvironmentinwhichthenetworkis
embedded.
embedded.
Thisdefinitionleadstothepossibilityofachieving
learning through:
learningthrough:
Synapticweightmodification.
Networkstructure(topology)modifications.
Bycreatingordeletingneuronsorconnections.
By creating or deleting neurons or connections
Throughtheuseofsuitableattractorsorother
suitablestablepoints.
it bl t bl
i t
Learningthroughforgetting.
Throughappropriatechoiceofactivationfunctions.
Throughmodifyingcontrollableparametersina
Through modifying controllable parameters in a
lookuptabledefininganactivation
transformation.
transformation
Combinationofabove.
48
Learningisusuallyaprocessofminimizinganerror
functionormaximizingabenefitfunction.
Inthisrespect,learningresemblestheoptimization,
whereweseektoextremizeadesiredcriterion
function(Error,Loss,Penalty,Computational
energy Lyapunov Hamiltonfunction)
energy,Lyapunov,
Hamilton function)
49
Learningrules
Some of the learning rules are significantly different than others, while
there are many that are of a minor variation, or are special cases of
more general ones,
ones usually identified by a different name.
name
Some important
p
rules are:
Hebb
Delta and its variations such as:
Backpropagation (the most common)
Adaline
Perceptron
Competitive learning
Hopfield
p
Kohonen
ART
50
Applicationsofartificialneuralnetworks
Iftheobjectiveistofindamappingsuchthatthesystem
capturesthedynamicsrelatinganinputdatasettoanoutput
dataset,theproblemiscalledfunctionapproximationor
systemidentificationincontrolengineeringapplications.
Iftheobjectiveistofindamappingsuchthesystemloosely
i
interpolateswithinaninputoutputdataset,theproblemis
l
i hi
i
d
h
bl
i
calledgeneralization.
Iftheobjectiveistofindamappingsuchthatthesystem
generalizestodiscreteoutputclasses,theproblemiscalled
classification.Insomefieldsthisproblemisgivendifferent
namessuchasassociativememories,patternassociation or
patternrecognition.
51
Applicationsofartificialneuralnetworks
In Engineering
Modeling
manufacturing
processes.
processes
Monitoringand
diagnosisofmachinery
failures.
= x1
Roboticsand
autonomousvehicles.
Automaticcontrol
systems.
Qualitycontrol.
& = x2
l
u
mc
x=x3
x& = x 4
52
InMedicineandHealthSciences
DiagnosticapplicationsinEEG,ECG,EMGandother
medical signatures
medicalsignatures.
Artificialvision.
Artificialhearing.
Artificialnose.
Artificial nose
Braincontrolledtelekinesis.
IdentifyingRNAandDNAinproteins.
53
Neuromusculardiseasediagnosis(UCY,ING)
54
QuantitativeAssessmentofthePrognosticFactorsinBreastCancer(UCY,ING)
Negative
025
25%
%
Weak
14
1+
2650% 2
Moderate
58
2+
5175% 3
Strong
3 912
3+
Very
Strong
4+
76%
>13
55
InSignalProcessing
Classification patternrecognition.
Si
Signalidentification,
l id ifi i
speechrecognition,
signaturerecognition,
i t
iti
imagerecognition,
speakeridentification.
k id tifi ti
Seismicimageanalysis.
Identificationof
explosives and military
explosivesandmilitary
targets.
Opticalcharacter
recognition.
56
In Meteorology
Modeling meteorological processes.
Filli i missing
Fillingin
i i rainfall
i f ll data
d (HTI,
(HTI UCY).
UCY)
y
Creation of isohyets.
In Cognitive studies
InCognitivestudies
AoccdrnigtoarscheearchatCmabrigdeUinervtisy,itdeosn
Aoccdrnig
to a rscheearch at Cmabrigde Uinervtisy it deosn'tt
mttaerinwahtoredrtheltteersinawrodare,theolny
iprmoetnt tihngistahtthefrist
tihng is taht the frist andlsatltteerbeattherghit
and lsat ltteer be at the rghit
pclae.Thersetcanbeatotalmsesandyoucansitllraedit
wouthitporbelm.Tihsisbcuseaethehuamnmniddeosnot
p
raederveylteterbyistlef,butthewrodasawlohe.
57
In Financial Engineering
Indecisionmaking.
Forecastingofmarketindices.(UCY,HTI)
F
i
f
k i di
(UCY HTI)
Stockexchangetrends.(HTI,UCY)
g
(
)
Bankrupsyprediction.(UCY)
Optimization.(HTI,UCY)
58
InDataclusteringandcompression
InTelecommunications,
Telemedicine,
Telemedicine
Geology,Meteorology
60
PossibleresearchareasinANNs
BasicResearch
Newlearningmethods
l
h d
Newarchitectures
Genetic algorithm learning
Geneticalgorithmlearning
Newsingleneuronmodels
Medicine
Diseasediagnosis
Engineering
Powerloadforecasting
l df
Meteorologicalapplications
Robotic and other complicated system control
Roboticandothercomplicatedsystemcontrol
Shipresistancecalculations
61
Dynamicalsystems
Conflictresolution,conflictavoidance(FuzzyCognitive
Maps?)
Dynamicanalysis (FuzzyCognitiveMaps?)
D ii
DecisionmakingdynamicsintheCyprusFederalState
ki d
i i th C
F d l St t
(FuzzyCognitiveMaps?)
Miscellaneous
Earthquakeprognosis
Athletic capability prognosis
Athleticcapabilityprognosis
62
ANNandStatisticalmethods
Thereexistssomeconflictamongneuroscientists
and statisticians on who invented first a particular
andstatisticiansonwhoinventedfirstaparticular
methodology!
Manyofthenon
Many
of the nondynamical
dynamicalmodelsusedinANNshave
models used in ANNs have
theirrootsinknownstatisticalmethodssuchas:
Generalizedlinearmodels
P l
Polynomialregression
i l
i
Nonparametricregression
Discriminantanalysis
Principalcomponents
Clusteranalysis
63
Comparisonofterminology
STATISTICS
ANN
Variables
Features
Independentvariables
Inputs
Predictedvalues
Outputs
Dependentvariables
d
i bl
Targetsortrainingvalues
i i
l
Residuals
Errors
Estimation
Training,learning,adaptation,orselforganization.
Estimationcriterion
Errorfunction,Costfunction,orLyapunov function
Observations
Patternsortrainingpairs
Parameterestimates
(Synaptic)weights
Interactions
Higherorderneurons
Transformations
Functionallinks
Regressionanddiscriminant analysis
Supervisedlearningorheteroassociation
Datareduction
Unsupervisedlearning,encoding,or
autoassociation
Clusteranalysis
Competitivelearningoradaptivevector
q anti ation
quantization
Interpolationandextrapolation
Generalization
64
Evolutionary Systems ES
EvolutionarySystems
Lately has been given great momentum
Latelyhasbeengivengreatmomentum.
TheyarusuallyknownasGeneticAlgorithms(GA)
Th
Theyarealgorithms
l ith foroptimizationandsearch.
f
ti i ti
d
h
Initiallywerestudiedby JohnHolland ofthe
University of Michigan back in the 70s.
UniversityofMichiganbackinthe70
s.
They resemble the natural evolution
Theyresemblethenaturalevolution.
65
Theymaybeusedin:
Function optimization.
Operationsresearch.
Learninginartificialneuralnetworks.
f l
l
k
Evolutionofnewartificialneuralnetworkarchitectures.
Evolutionoffuzzyrules.
Theyarerelativelysimple
y
y
p toapply,especiallyincomplicated
pp y, p
y
p
optimizationproblems.
They can find global optimization solutions
Theycanfindglobaloptimizationsolutions.
66
Somemoremoderntermsthatarebeingusedare:
EvolutionaryComputationor Evolutionary Algorithms.
Theseinclude:
GeneticAlgorithms GA
Genetic Programming GP
GeneticProgramming
EvolutionaryProgramming EP
Evolutionaryhardware EHW
Evolutionary Strategies ES
EvolutionaryStrategies
LearningClassifierSystems LCS
67
Theycanshowanemergingbehaviour.
Canbeusedin:
Optimization problems
Optimizationproblems.
Neural network learning
Neuralnetworklearning.
...
ModernParticleSwarmSystems isanextensionofALS.
69
70
Someparadoxes:
Iamaliar!
(Epimenides)
AmIsayingthetruth?
Thebarberwhoshavesallthosewhodontshave
themselves
(Bertrand Russel)
(BertrandRussel)
Wh h
Whoshavesthebarber?
h b b ?
71
Fuzzy Systems
FuzzySystems
Example:Whenissomeonetall?
If P isthesetofallhumansandh theirheight,afuzzysubset(h) that
couldanswertheabovequestionis:
q
(h) =
0
(h 1,5)
if
h < 1,5 m
if
1,5 h 2,2 m
if
h > 2,2 m
0,7
1
72
Differencebetweenfuzzyandconventional:
Inbinarylogic,theresultcanbetrueorfalse:
e g 1+1=2
e.g.1+1=2
Theworldismostlyfuzzy.
e.g.
Itscold,
It is green
Itisgreen,...
73
FuzzySystems
Membershipfunctions
(h)
1
0.5
0
0.5
1.85
2.2
Height,h [m]
SSuchfunctionsmaybeconstructedforotherparameters,
hf
i
b
df
h
andgenerateafuzzylogicsystemwithrules:
74
FuzzySystems
SomemembershipfunctionsfromtheMATLABlibrary:
trapmf
gbellmf
trimf
gaussmf
gauss2mf
smf
1
0.8
0.6
0.4
02
0.2
0
zmf
psigmf
dsigmf
pimf
sigmf
0.8
0.6
0.4
0.2
0
75
FuzzySystems
y y
ANDfuzzyrule:
AND
76
MostapplicationsareinthefieldsofAutomatic
,
y
andinsupporting
pp
g
Control,inInformationSystems
Decisionmaking.
77
FuzzyCognitiveMaps
y g
p FCM
Itisakindof
combinationbetween
ANNsandfuzzylogic.
Theyaremadeofasetof
junctionsthatrepresent
conceptsandfromarrows
thatconnectthese
junctions.
Canbeusedincomplicatedproblemsinordertosimulatethe
dynamical interrelationships asinDecisionMaking
dynamicalinterrelationships
as in Decision Making andin
and in
StrategicPlanning.
Essentiallyarerecurrentdiagrams thatusefuzzyvalues.
78
AnattempttostudyinterrelationsintheCyprusProblem
Conceptsthathavebeenconsidered
Concepts
that have been considered
30 influencingparameters(concepts)havebeenexplored.
These are:
Theseare:
INFLUENCING PARAMETERS (Concepts)
C1
C2
C3
yp
State
Welfare of the Turkish Cypriot
C4
C5
Christian religiousness
C6
C7
C8
C9
C10
Islamic religiousness
79
Conceptsthathavebeenconsidered
C11
C12
C13
C14
C15
C16
C1
C17
Political
li i l interests
i
off UK
C18
C19
C20
80
Conceptsthathavebeenconsidered
C21
C22
C23
C24
Military interests of UK
C25
Mili
Military
iinterests off USA
C26
C27
C28
Oil fields
C29
Quality of environment
C30
81
300%
Effects
EffectsofchangesinIslamicreligiousness,EuropeanandUSAinterests
of changes in Islamic religiousness European and USA interests
250%
Percentage
g change
g in parameters
p
due to major
j increase in Islamic Religiousness
g
es
rc
ou
te
es n t
sta
l r me
ed
ra
at
tu ron
er
na vi
ed
er f en
th
eF
O ty o
t h rs
li
in ett le
ua
Q elds ri sm an s
u li
fi
o
SA
t
to
il
O l of ana f U
o
ve of ts K
Le sts eres f U ey
o k
e
t
ter in ts ur
In ary eres f T ce
o
t
t
ili in ts ree
M ary res f G
l
e o
t
ili int ts rae
M ry res f Is y
e o ke
ta
r
ili int s
t
M ry res Tu e
f
e
ta
c
ili int s o ee
M ry rest Gr
f
l
ta e
ili nt s o ae
M al i rest f Isr
ic e o
lit nt s K
t
i
Po cal eres f U ia
i
o
lit int ts uss
Po cal eres f R
i
lit int ts o SA
Po cal res f U e
p
e
i
lit int ts o ro
TC
Po cal res f Eu C the C
e
i
lit int ts o e T by e T
Po cal res f th ry y t h
o
e
i
o
lit int el ist e b
Po cal lev ek h uag
i
l re ng
lit na
a
Po atio of G ek l
uc ge re
Ed led of G ss
C
w e ne
s
eG
no g
K led iou
C y th GC
w ig
e G b t he
no el
K i c r ism f th tory by
o is e
m al
la ion vel h h uag
Is
t
le i s g
a
N al urk lan
n
TC atio of T ki sh
uc ge ur
Ed led of T ness
w e us
no g
K ed i o
g
l
w eli
s
r
e
no
ru
K tian li sm Stat
yp
ris na C e
fC
t
Ch atio he T St a te o
a
t C
N
St
C of
G re he G eral
fa t ed
el of
W re he F
fa t
el of
W re
fa
el
W
-50%
P
Percentage
change
h
iin parameters due
d to major
j increase
i
in
i USA interest
i
150%
100%
50%
0%
82
CaseofincreaseintheIslamicreligiousness
ThisresultsinsignificantincreaseintheGreekCypriot
nationalism(+36%)andmostimportantlyaveryhighincrease
inChristianreligiousness(+283%).
Also,itwillresultinanincreaseinthemilitaryinterestsof
Al
it ill
lt i
i
i th
ilit
i t
t f
Turkey,GreeceandIsrael.
Ofsignificanceistheobservationthatthewelfareofthe
FederalStateofCypruswillreduceby7%andthewelfareof
yp
y
theTurkishCypriotcommunityby4%.
Somereductionintheleveloftourismandthequalityofthe
environmentisalsoobserved.
83
CaseofincreaseintheEuropean politicalinterestsinCyprus
Themodelpredictsasignificantincreaseoftheknowledgeof
GreeklanguageandhistorybytheTurkishCypriotsandvice
versa.
Thisismorepronouncedinthecaseofknowledgeofthe
Thi
i
d i th
fk
l d
f th
TurkishlanguageandhistorybytheGreekCypriots.
Itishoweververyinterestingthatthemodelsuggestsa
significantreductioninChristianreligiousness(28%)and
g
g
(
)
somewhatless(5%)intheIslamicreligiousness.
Mostinterestingthough,isthesuggestionthattheinterestsof
theAnatoliansettlerswillreduceby20%.
84
CaseofincreaseinthepoliticalinterestsofUSA inCyprus
Themodelpredictsa12%increaseinthepoliticalinterestsof
Israelanda12%increaseinthewelfareoftheTurkishCypriots.
Whatismostinterestingthough,andsomewhatunusual,isthe
reductionoftheChristianreligiousnessby16%,oftheTurkish
d ti
f th Ch i ti
li i
b 16% f th T ki h
Cypriotnationalismby18%,oftheIslamicreligiousnessby8%,
of the political interests of Greece and Turkey by 11% and 16%
ofthepoliticalinterestsofGreeceandTurkeyby11%and16%
respectively,andofthemilitaryinterestsofUKby18%.
85