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.
•
Making itpossible to exploitexat gradients, ifthey areavailable.•
Providinga setof reommendabletolerane parameters.•
Testingdierent strategies for updating thepenaltyfator.•
Testingother weighting fator strategies.•
Further investigations of how the number of mapping parameters in-uenethe results.•
Further analysison theoptimalmapping parameters.•
Testingdierent strategiesonerningthenumberof mappingparam-eters,eg. releasing aparameter inevery main iteration.
•
Further analysisof thetest results,eg. looking at thetermination ri-teria for the sub-algorithms.•
Further analysisof theiteration sequenes.•
Deningdierentresidualformulations,eg.lettingtheformulation de-pend onthe numberofiterates available. Possiblyomittingthegradi-ent residual inthe rstiterations in orderto minimize thenumber of
ne modelevaluations.
Chapter 6
Conlusion
Inthis thesis aSpae Mappingalgorithm hasbeen presented and tested on
some test problems. We have made a number of investigations in order to
make a robust implementation and analyze the harater of the Parameter
Extration probleminregard to thesolutions,wean expet tond.
Thesteplengthusedinthenitediereneapproximationsisfoundtobeof
greatimportane,inasethe exatgradientsofthemodelarenot available.
The step length inuenes both the trunation errors and the rounding
er-rors,and basedonresults fromtheTLT2problem andthetheoretial error
funtions, we nd a suitable value for the step length to be approximately
10 − 5
.The number of equations in the Parameter Extration problems depends
on, how many main iterations we have made. A large part of the
Param-eter Extration problems are therefore underdetermined, and onsequently
thereareinnitelymanysolutions. Thereisalsothepossibilityofhaving an
overdeterminedproblem,inwhihasewe maybeabletoeetthesolution
bymeans of the residual formulation.
The formulation of the residual inluding the normalization fators, the
weighting fators and the penalty fator makes it possible to give eah of
the residual elements an individual priority. In this way we an more or
less hoose whih order of loal and global agreement the surrogate model
shouldsatisfy. Thesolutionwendisinuenedbythehoieoftheresidual
formulation.
Itmustbelaried,thatwearenotalwayssure,thatthevariousfatorseven
inuenethe results. Thisdependsentirely ontheharater oftheproblem,
WehavehosentodenetheParameterExtrationproblemsasleastsquares
problems, and solve them by the Marquardt algorithm. The Marquardt
methodiswell-suitedfortheproblem,sine thedampingparameterontrols
the step length and ensures, that the Marquardt equations have a unique
solution.
Ifthe rankof
J r
isnot full,the solution depends onthedamping parameterµ
. For smallµ
-valuestheMarquardtsolutionisnearlyidentialtothemini-mumnormsolution, whihisthesolutionlosesttothepreviousone. Inthis
asetheMarquardtequationsgivethesameeetasaformofregularization
wrt.the previous solution.
If we regularize the problems, we are guaranteed, that the Parameter
Ex-tration problems are overdetermined in all iterations. The regularization
termensures,thatwe ndasolutionlosetotheprevioussolution. We have
shown, that solving the Parameter Extration problem for the regularized
residual formulation orresponds to a speial ase of the Marquardt
equa-tionswiththedampingparameter
1 + µ
andahangedrighthandside. Theregularizationmeans,thatwe annot besureto satisfythegradient math,
ifthe solutionis farfrom theprevious.
Thealgorithmwiththevariousresidualformulationsistestedonthree
prob-lems: The Rosenbrok problem in its original form and in the augmented
version provides a theoretial example, and theTLT2 and TLT7 problems
areexamplesofengineering design problems ofmore pratial harater.
Thetest resultsshow, thatthe algorithm workswell inaseof both solving
theregularized and the unregularized problem. In several ases thelatteris
atually resulting inafaster onvergene to theoptimizer,and thesolution
is found, before there is a possibility of having overdetermined Parameter
Extration problems. Weonlude, thatfor theonsidered test problems it
isno obstale, iftheParameter Extration problems areunderdetermined.
TheRosenbrokproblemshowssomeinteresting featuresonerningthe
op-timalmappingparameters. TheRosenbrokfuntionisspeial,beausethe
seond response funtion is linear and only depends on one of the design
parameters. Thishasonsequenesfor themapping parametersthroughthe
iteration sequene, inthe sense thatseveral of themapping parameters are
not hanged form theinitial values.
For theaugmentedRosenbrokproblemthesamefeaturesarefoundfor the
optimalmapping parameters.
Inthe engineeringdesignproblemTLT2wegetgoodonvergeneresultsfor
But when we onsider the approximation errors from using the surrogate
model dened by the optimal mapping parameters, the two versions give
very dierent results.
The regularized problem provides good surrogate approximations and
gen-erally, the surrogate model is better than a lassial Taylor model
approx-imation - not only over a large region of the design parameter spae, but
alsoneartheexpansionpoint. Forthesenariowithoutregularizationofthe
residualwe getaverypoorsurrogatemodel, even though thealgorithm has
onvergedto the ne modeloptimizer.
It shows,that for this problemthe normalization fatorshave an inuene.
Weahieveaeptableresultsbothregardingonvergeneandthequalityof
thesurrogatemodelapproximationfor salingofonly thegradient residual.
It is not possible to onlude anythingin general on the normalization
fa-tors, but in this ase they must be omitted to ensure a suitable surrogate
model.
The importane of the quality of the surrogate model approximation of
ourse depends on the design problem: If we are only interested in the
optimal design parameters, the surrogate model is perhaps unimportant.
Butinmodellingproblemswe maybeinterested inreplaing thenemodel
withthesurrogatemodelandexaminingthisinsteadatalowomputational
expense. Insuhaaseanunauratesurrogateapproximationisnotuseful.
The last test problem TLT7 is of larger dimension. The results here show
fastonvergeneintheaseoftheunregularized ParameterExtration
prob-lems,andthe redutionofthe mappingparameters havenoinueneonthe
results. Butwhenthe regularizationterm isadded,there isnoonvergene.
It isnot possible to explainthis behaviour.
Itappliesforalltheonsideredtestproblems,thattheSpaeMapping
algo-rithm provides fast onvergene results ompared to lassial optimization
methods.
Appendix A
Short User's Guide for The
SMIS Framework
InthefollowinghapterabriefoverviewoftheSMISframeworkinMatlab
isgiven.
A.1 The Problem Setup-le
Thesetup-le for thegivenproblem isplaed intheorrespondingproblem
diretory. The setup-le ontains information on the ne and oarse model
optimizers and linear equality/inequality onstraints for the optimization
problem.
Thesearedened by:
A (1:leq,:) x + b (1:leq,:) = 0
andA (leq+1:l,:) x + b (leq+1:l,:) ≥ 0
Thesetup-le mustreturn astruture
S
ontaining thefollowing elds:n The numberofdesign parameters.
m The numberofn parameters.
fine The m-leimplementing thene model funtion.
oarse The m-leimplementing theoarse modelfuntion.
parf Vetor withthe sampling points forthene model.
par Vetor withthe sampling points fortheoarse model.
paro Struture withinformation usedfor plotting.
xast The nemodeloptimizer.
zast The oarse modeloptimizer.
A
l
-by-n
matrix deningtheonstraints (l =
size(
A, 1) − leq
)b
l
olumn vetor deningtheonstraints.leq Numberof linearequality onstraints.
x Initial guessfor oarse modeloptimizer.
delta Initial trustregion radius for surrogate optimizationproblem.
eps1 Used instopping riterion forthe mainproblem.
eps2 Used instopping riterion forthe Parameter Extration problems.
maxfun1Maximal numberof funtion evaluations inAlgorithm1.
maxfun2Maximal numberof funtion evaluations inAlgorithm2
(also usedinAlgorithm 3).
p1 Norm deningtheobjetivefuntion of themain problem.
p2 Norm deningtheobjetivefuntion of theParameter
Extration problems. When usingthealgorithm version
withmarquardt thisoption hasno eet.
iontr Parameter usedforontrolling the omputation:
1: starts algorithm,2: startsalgorithm andprints information.
diagA Parameter deningnumberof mapping parameters
(0: fullmatrix
A
,1: diagonal matrixA
)Theoptions formarquardt aredened inthele'lmm'aseg.opts
=
([1e-81e-14 1e-14 200 1e-12℄).