4.4 The TL T7 Problem
4.4.2 The Results of the T est Runs
Eet of the Regularization
Thealgorithm does not onvergeto theoptimizerinthe asewhere we use
theregularized residual inthe Parameter Extration. When solvingthe
un-regularizedParameterExtrationproblemwegettheperformaneresultsof
the algorithm depited in gure 4.4.2. The result is with normalization of
the residualelements anda full sizeparameter vetor.
The optimizer is found in 10 iterations within the auray
10 − 16
of theobjetive funtion. Thenumberofunknown mappingparameters ineahof
the Parameter Extration problems is
n p = 57
. In the ase of noregular-ization this means, that until iteration
51
we will have lessequations thanunknowns. Thisis obviouslynot an obstale, sine we ndthe optimizerin
12
iterations.It is not known, why the test senario with theregularized Parameter
Ex-0 5 10 15 10 −20
10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * ||
2 F(x (k) ) − F(x * )
Figure4.4.2: Withoutregularization
Eet of the Normalization Fators
Ingure4.4.3we seethe iterationsequenefor the aseofnonormalization
of the residual vetor, orresponding to
a 1 , . . . , a k = 1
andd = 1
. Thetestisomputed withfull parametervetor.
Theonvergene is alittle faster than theresultfrom gure4.4.2.
0 5 10 15
10 −20 10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * ||
2 F(x (k) ) − F(x * )
Figure4.4.3: Withoutnormalization
Eet of the Weighting Fators
With no redution of the number of unknown mapping parameters in the
Parameter Extration we have
n p = 57
, and hereby the weighting fatorsmay inuene the results from iteration number
51
. Beausethe optimizerisalreadyfoundatthispoint,weannotusethistestsenariotoinvestigate
Instead we look at the senario withredution of the parameter vetor. In
thisasewehave
n p = 15
,andsomeoftheweightingfatorswillbedierentfrom
1
from iteration9
and onwards, if we use the strategy from setion3.6. Figure 4.4.5 on the right shows the iteration sequene with the useof
weightingstrategy. Foromparison theresults withoutweighting areshown
ingure4.4.4 onthe left.
0 5 10 15
10 −20 10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * ||
2 F(x (k) ) − F(x * )
Figure 4.4.4: Withoutweighting
0 5 10 15
10 −20 10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * ||
2 F(x (k) ) − F(x * )
Figure4.4.5: Withweighting
The gures are almost idential, and the weighting fators are pratially
withoutinuene inthis problem.
Eet of the Number of MappingParameters
Wenowshowtheresultsfromusingareduedparametervetor,
orrespond-ingtoadiagonal matrixforthe inputmappingmatrix
A
. Figure4.4.6isforthe regularized ase, andgure 4.4.7for theunregularized ase.
Thealgorithmdoesnot onvergeintherstasewithregularization. Inthe
seondase theoptimizeris foundin
12
iteration steps withan aurayof10 − 15
on the ne model objetive funtion. It is not known, why the testrunwith regularizationdoesnot onverge to theoptimizer.
Thereispratiallynodierenebetweentheiterationswiththefull
parame-tervetor(gure4.4.2)andthereduedparametervetor(gure4.4.7). The
algorithm works well when solving the unregularized Parameter Extration
problem,and ithasno eet whetherthe problemsareunderdetermined or
0 10 20 30 40 50 60 70 80 10 −20
10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * || 2 F(x (k) ) − F(x * )
Figure4.4.6: Withregularization
0 5 10 15
10 −20 10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * ||
2 F(x (k) ) − F(x * )
Figure 4.4.7: Without
regulariza-tion
Optimal Mapping Parameters
This problem is of larger dimension than the other test problems, and we
havenot investigated the optimal mapping parameters.
Diret optimization
Belowwe seetheperformaneof thediretd-algorithm.
0 20 40 60 80 100 120
10 −20 10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * ||
2 F(x (k) ) − F(x * )
Figure4.4.8: Performaneof diretoptimization ofthene model
The lassial optimization method onverges in
127
iterations, and we on-ludethatthe SpaeMappingtehnique isveryuseful inthisproblem,sineitreduesthe numberofnemodelevaluationswithapproximately afator
10
.Summaryof the Results
•
When the Parameter Extration problem is regularized we have no onvergene to theoptimizer, both inthe aseof afull and a reduedparameter vetor.
•
The normalization fatorshave littleeet on theonvergene speed.•
The weighting fators have pratially no inuene on the results inthe aseofa redued parametervetor.
•
The redutionof themapping parameters hasno eet on theperfor-mane inthis test problem.
•
It isnotpossibletogetasgoodonvergene resultsastheresultsfrom[1 ℄ and [2 ℄, probably beause the problem is not idential to the one
solvedinthis report.
Chapter 5
Future Work
5.1 Improvements of the SMIS Implementation
Anumberofhangesoftheimplementation ouldbemadeinorderto make
theSMISframeworkmoreexibleanduser-friendly. Alsotheaurayofthe
omputationsouldbeimprovedforsomeproblems. Theurrentframework
onsistsof many diretories and les linked together ina ertain struture.
The suggested hanges will therefore inuene many of the inluded les,
whihwill eet theduration of thework involved.
Intheurrentimplementation thegradientsofthesurrogatemodelwrt.the
parameter vetor are alulated by forward dierene approximations. In
ordertominimizethetrunationerrors,itwouldbeanadvantageto exploit
theexat gradients, ifthey areavailable.
TheMatlab-lediffjaobianusedtoalulatetheforward dierene
ap-proximation. In the urrent implementation the parameter
η
dening therelative step length is xed at a ertain value, and an only be altered by
modifyingthe Matlab-le diretly.
Theuser friendliness ouldbeimproved byinluding more optional
param-eters to the problem setup-le. The setup-le should dene the following
optionalparameters:
∆
Initialtrust region radius for Algorithm1.max f1
Maximal numberof funtionevaluations inAlgorithm1.max f2
Maximal numberof funtionevaluations inAlgorithm2.max f3
Maximal numberof funtionevaluations inAlgorithm3.η x
Steplength indiffjaobianwhenalulatingJaobian wrt.x
.η p
Steplength indiffjaobianwhenalulatingJaobian wrt.p
.ε F
Usedinthe stopping riterion for thene modelobjetive funtion.ε K
Usedinthe stopping riterion for thegradient mathing.ε hx
Usedinthe stopping riterion for thestep lengthforx
-iterates.ε hp
Usedinthe stopping riterion for thestep lengthforp
-iterates.opts Options formarquardt ([1e-8 1e-4 1e-4 200 1e-12℄).
diagA Parameter dening thenumberof mapping parameters
(0: fullmatrix
A
,1: diagonal matrixA
)Someof the above options arealready used in theSMIS framework, ie.
∆
,max f1
,max f2
,anddiagA,andalsotwo toleraneparameters orresponding toε F
,ε K
,ε hx
andε hp
.We also list some possible hanges inorder to inrease the auray of the
omputations:
•
Introduingthestep lengthparameterη
asaninput parameterto thefuntion diffjaobian.