ISSPIT07 Sabbir Pres

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 25

DSP DSP DSP DSP DSP DSP DSP DSP

Group Group Group Group Group Group Group Group


A Multiobjective Genetic Algorithm for the
A Multiobjective Genetic Algorithm for the
Design of Asymmetric FIR Filters
Design of Asymmetric FIR Filters
Sabbir U. Ahmad and Andreas Antoniou Sabbir U. Ahmad and Andreas Antoniou
Dept. of Electrical and Computer Engineering Dept. of Electrical and Computer Engineering
University of Victoria, Canada University of Victoria, Canada
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
Introduction
Classicalvs.geneticoptimization
Multiobjectiveoptimization
Theelitistnondominatedsortinggenetic
algorithmapproach
Designexample
Conclusions
OVERVIEW
OVERVIEW
2
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
INTRODUCTION
INTRODUCTION
3
FIRfiltersareusuallydesigned
withsymmetriccoefficientsto
achievelinearphaseresponse
withrespecttothebaseband,i.e.,
b
0
=b
N-1
,b
1
=b
N-2
,etc.
Efficientdesignmethodsare
available,e.g.,windowmethod,
Remezalgorithm
Largegroupdelay
b
N-2
b
0
b
1
b
N-1
T
T
T
y(nT)
x(nT)
FIR Filter
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
INTRODUCTION (Cont
INTRODUCTION (Cont

d)
d)
4
FilterswithAsymmetricCoefficients FilterswithAsymmetricCoefficients FilterswithAsymmetricCoefficients FilterswithAsymmetricCoefficients
Approximately linearphaseresponseinpassband
Relativelysmallgroupdelay
Canbedesignedbyusingclassicaloptimization
methodswithamultiobjectiveformulation.
Canalsobedesignedbyusingamultiobjective
geneticalgorithm(GA)knownastheelitist
nondominatedsortedGA (ENSGA).
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
CLASSICAL OPTIMIZATION ALGORITHMS
CLASSICAL OPTIMIZATION ALGORITHMS
Fastandefficient
Verygoodinobtaininglocalsolutions
Unbeatableforthesolutionofconvex(concave)
problems
Inmultimodalproblems,theytendtozoomtoa
solutioninthelocaleoftheinitializationpoint.
Notequippedtodiscardinferiorlocalsolutionsin
favourofbettersolutions.
5
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
6
GENETIC ALGORITHMS
GENETIC ALGORITHMS
Areveryflexible,nonproblemspecific,androbust.
Canexploremultipleregionsoftheparameter
spaceforsolutionssimultaneously.
Candiscardpoorlocalsolutionsinfavourofmore
promisingsubsequentlocalsolutions.
Theyaremorelikelytoobtainbettersolutionsfor
multimodalproblemsthanclassifiedmethods.
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
7
GENETIC ALGORITHMS (Cont
GENETIC ALGORITHMS (Cont

d)
d)
OwingtotheheuristicnatureofGAs,arbitrary
constraintscanbeimposedontheobjectivefunction
withoutincreasingthemathematicalcomplexityof
theproblem.
Multipleobjectivefunctionscanbeoptimized
simultaneously.
Theyrequireaverylargeamountofcomputation.
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
MULTIOBJECTIVE OPTIMIZATION
MULTIOBJECTIVE OPTIMIZATION
8
Inmanyapplications,severalobjectivefunctions
needtobeoptimizedsimultaneously.
Inclassicaloptimization,multipleobjectivefunctions
areusedtoconstructamorecomplexunified
objectivefunctionwithorwithoutconstraints.
Ontheotherhand,withGAs multipleobjective
functionscanbeoptimizeddirectlytoobtainasetof
compromisesolutionsoftheproblemathand.
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
MULTIOBJECTIVE OPTIMIZATION (Cont
MULTIOBJECTIVE OPTIMIZATION (Cont

d)
d)
9
Amultiobjectiveoptimizationproblemwithkobjective
functionscanberepresentedas:
Minimize
)] ( ..., ), ( ), ( [ ) (
2 1
x x x x f
k
f f f =
subject to
X x
Asetofcompromisesolutionsobtainedby
multiobjectiveapproachesisknownasaParetooptimal
solutionset.
Theusercanselectthebestcompromisesolutionfrom
aPareto-optimalset.
where X is the solution space
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
PARETO OPTIMALITY
PARETO OPTIMALITY
10
Asolutionx
1
issaidtodominateasolution x
2
ifthe
followingconditionshold:
Solutionx
1
isnoworsethanx
2
forallobjectives
Solutionx
1
isstrictlybetterthanx
2
inatleastone
objective
Asolutionthatisnotdominatedbyanyotherissaidto
beanondominatedsolution.
Asetofnondominatedsolutionsinthesolutionspace
isaPareto-optimalsolutionset.
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
PARETO OPTIMALITY (Cont
PARETO OPTIMALITY (Cont

d)
d)
11
Thenondominatedsetoftheentirefeasiblesolution
spaceisthegloballyPareto-optimal set.
ThesolutionspacecorrespondingtothePareto
optimalsolutionsetiscalledtheParetofront.
f
3
(x)
Pareto Pareto
front front
Feasible Feasible
solution solution
space space
) (
1
x f
) (
2
x f
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
12
ELITIST NONDOMINATED SORTING GA
ELITIST NONDOMINATED SORTING GA
- ENSGAintroducesdiversityinsolutionsbysorting
thepopulation accordingtothenondominationprinciple.
- classifiesthepopulationintoanumberofmutually
exclusiveclasses
- assignshighestfitnesstothemembersoftheclass
thatareclosesttothe Pareto-optimalfront
- usestheelitismprincipletoincreasethenumberof
Paretosolutions.
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
13
ENSGA STEPS
ENSGA STEPS
Initializepopulation
Evaluatefitnessofindividuals
Selectnondominatedsolutions
byusing nondominatedsorting
byassigningfitnessvaluesaccordingtocrowdingdistance
Performcrossovertoobtainnewoffsprings
Mutateindividualsintheoffspringpopulation
Evaluateoffspringsolutions
Performelitistreplacement
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
14
NONDOMINATED SORTING
NONDOMINATED SORTING
Identifythebestnondominatedset.
Discardthenondominatedsolutionsfromthe
populationtemporarily.
Identifythenextbestnondominatedset.
Continuetillallsolutionsareclassified.
F2: Rank =2
F1: Rank =1
F1: Rank =1
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
15
CROWDING DISTANCE
CROWDING DISTANCE
Crowdingdistanceisadiversitymetric.
Crowdingdistanceisdefinedasthefrontdensityina
specificlocale.
Eachsolutionisassignedacrowdingdistance.
Solutionslocatedinalesscrowdedspacearepreferred.
SolutionA islocatedinaless
crowdedlocalethanB
SolutionA islocatedinaless
crowdedlocalethanB
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
16
ELITIST REPLACEMENT
ELITIST REPLACEMENT
Combineparentandoffspringpopulations.
Selectbetterrankingindividualsandusecrowding
distancetobreakanyties.
Parent
population
Offspring
population
Nondominated
sorting
Crowdingdistance
sorting
Population
for next
generation
Population
discarded
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
17
DESIGN OF ASYMMETRIC FIR FILTERS
DESIGN OF ASYMMETRIC FIR FILTERS
Apopulationofpotentialsolutionsiscreatedfroman
initialleast-squaressolution.
Simulatedbinarycrossovers andpolynomial
mutations areappliedaccordingtopredefined
probabilitiesofoccurrence,P
x
andP
m
,respectively.
Theobjectivefunctionsusedtoevaluatethefitness
oftheindividualsolutionsarebasedonthe
amplituderesponseandgroup-delayerrors.
TheENSGAisterminatedafteraprespecified
numberofgenerations.
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
18
CHROMOSOME STRUCTURE
CHROMOSOME STRUCTURE
Chromosome Chromosome Chromosome Chromosome Chromosome Chromosome Chromosome Chromosome ( (candidatesolution candidatesolution): ):
ThecoefficientvectoroftheFIRfilter,b, isusedas
thecandidatesolution.
Toavoidverylongbinarystrings,afloating-point
representationisusedinencodingthe
chromosomes.
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
19
OBJECTIVE FUNCTION
OBJECTIVE FUNCTION
Passband for ) ( 1 max =
j
p
e H F

L
Stopband for ) (
1
2
=

=
k
K
k
j
a
a
k
e H F

Threeobjectivefunctionshavebeenused:
- normofthepassbandamplitude-responseerror
- normofthestopsbandamplitude-response
error withaconstrainedimposedonthepeakerror
- AparameterQ whichmeasuresthedegreeof
flatnessofthepassbandgroup-delaycharacteristic


+

=
) (
) ( 100


Q
2
L
{ } d | ) ( | log 20 min subject to
10
B e H
a
j
k


DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
20
Specifications: Specifications: HighpassFIRfilter,

a
=0.4,
p
=0.55,
s
=1rad/s,

a
=52dB, N =35
Initialization: Initialization:AnFIRfilterwithnonsymmetric
coefficientsdesignedusingaweightedleast-
squaresmethodwasusedastheinitialdesign.
Solutions: Solutions:Thenexttwoslidesgivetheresultsfor Thenexttwoslidesgivetheresultsfor
twosolutionsfromthePareto twosolutionsfromthePareto- -optimalsolutionset. optimalsolutionset.
DESIGN EXAMPLE
DESIGN EXAMPLE
, 5 . 16
2

+
=

DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
21
Amplituderesponse Delaycharacteristic
Solution1: Solution1:
DESIGN EXAMPLE (Cont
DESIGN EXAMPLE (Cont

d)
d)
DesignobtainedwithENSGAInitialdesign
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
22
DESIGN EXAMPLE (Cont
DESIGN EXAMPLE (Cont

d)
d)
Amplituderesponse Delaycharacteristic
Solution2: Solution2:
DesignobtainedwithENSGAInitialdesign
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
23
DESIGN EXAMPLE (Cont
DESIGN EXAMPLE (Cont

d)
d)
3-DscatterplotofthePareto-optimalsolutionsobtained
byusingtheENSGA
Stopband for | ) ( | log 20
10
=
j
a
e H A
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
24
CONCLUSIONS
CONCLUSIONS
TheENSGAcanbeusedtodesignnonsymmetric
FIRfiltersthatwouldsatisfymultiplerequirements
imposedontheamplituderesponseandthedelay
characteristic.
Theapproachyieldsanimproveddesignwith
respecttotheinitialweightedleast-squaresdesign.
Thedesignthatisbest-suitedtoaspecific
applicationcanbechosenfromthesetofPareto-
optimalsolutionsobtained.
IncommonwithotherGAs,theENSGArequiresa
largeamountofcomputation.However,thisisnota
seriousproblemunlessthefilterdesignhastobe
carriedoutinrealorquasi-realtime.
DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group DSP Group
25
Thankyou Thankyou Thankyou Thankyou
Thankyou Thankyou Thankyou Thankyou

You might also like