4.2 The Rosenbrok Problem
4.2.2 Linear Transformation
Inthis example we usethefollowing transformationof thene model:
u(x) = Cx + d =
1.1 − 0.2 0.2 0.9
x 1 x 2
+ − 0.3
0.3
(4.2.1)
The orresponding ne model optimizer rounded to
4
deimals isx ∗ = [1.2718 , 0.4951]
T.−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2
−1.5
−1
−0.5 0 0.5 1 1.5 2
Coarse model
z 1 z 2
z * = [ 1.0000 , 1.0000 ] T
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2
−1.5
−1
−0.5 0 0.5 1 1.5 2
Fine model
x 1 x 2
x * = [ 1.2718 , 0.4951 ] T
Figure4.2.1: Contourplots of theRosenbrokfuntion: oarse model(left)
andne model (right)
Weshowthe levelurvesof theoarseand nemodelsingure4.2.1,where
the objetive funtion is
F = k·k 2 2
. The ne model is very similar to theoarse model (the original Rosenbrok funtion) and the harateristi
ba-nana shapeis stillpresent.
In all the test results with the Rosenbrok funtion we use the options
ε 1 = 10 − 14
,ε 2 = 10 − 14
and opts=
[1e-8 1e-14 1e-14 200 1e-12℄. Theinitial guessforthe oarsemodeloptimizeris
x (0) = [ − 1.2 , 1.0]
T. Weshowthe iteration sequenes also after the algorithm has onverged to the
opti-mizer.
Eet of the Regularization
The onvergene is very fast for both the ase with regularization and the
asewithout. The performanes areshown ingures4.2.2 and 4.2.3.
We notie the absene of the points of iteriation
6
in gure 4.2.2, beausethevalue
0
is not visibleinthesemilogarithmi plot.0 2 4 6 8 10 10 −20
10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * ||
2 F(x (k) ) − F(x * )
Figure 4.2.2: Withregularization
0 2 4 6 8 10
10 −20 10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * ||
2 F(x (k) ) − F(x * )
Figure 4.2.3: Without
regulariza-tion
runs: Theonvergeneisalittlefasterwithouttheregularizationtermadded
to the residual vetor. When we solve the unregularized problem, we have
n p = 7
and thereisapossibilityofanoverdeterminedParameterExtration probleminiteration6
and onwards. For this problemtheunderdetermined Parameter Extrationproblems donot have anegative eet ontheonver-generate, sine we ndthene modeloptimizer before iteration
6
.Theiteration pointsorrespondingto gure4.2.3 areshowninthe
(x 1 , x 2 )
-plane with the objetive funtion
F = k f(x) k 2 2
. It is noted that the rstiteration point plotted is the initial guess for the oarse model optimizer.
Therst evaluation ofthene modelis madeinthepoint
x (1)
whih istheseondpoint plotted ingure4.2.3.
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2
−1.5
−1
−0.5 0 0.5 1 1.5 2
Iteration sequence
x 1 x 2
Iteration points
Figure 4.2.4: Sequene ofiteration points
Withthe SpaeMappingalgorithm we avoid theiteration sequenemoving
Eet of the Normalization Fators
Thenext test runsaremadewithall normalization fatorsequal to
1
.Theiterationsequenesingures4.2.5-4.2.6arealmostidentialwithgures
4.2.2-4.2.3,andwe onlude,thatthenormalizationfatorshavepratially
noeet inthe Rosenbrok problem.
0 2 4 6 8 10
10 −20 10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * ||
2 F(x (k) ) − F(x * )
Figure 4.2.5: With regularization
andwithout normalization
0 2 4 6 8 10
10 −20 10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * ||
2 F(x (k) ) − F(x * )
Figure 4.2.6: Without
regulariza-tionand withoutnormalization
Eet of the Weighting Fators
We an not use this problem for testing the eet of theweighting fators.
Theweighting fatorsfromthe Gaussdistributed weight funtion approah
areonlydierent fromzerofromiterationnumber
6
. Atthis point thesolu-tionis alreadyfound.
Eet of the Number of Mapping Parameters
Inthisasetheresultsareverydierent,whenweusethereduedparameter
vetorinsteadoftheomplete. We testtheperformane ofthealgorithm in
thefollowing threeases:
•
Withregularization•
Withoutregularization•
Withoutregularization and withweighting fatorsAll threeases are withnormalization of the residual elements. The results
0 5 10 15 20 25 30 10 −20
10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * || 2 F(x (k) ) − F(x * )
Figure4.2.7: With
regu-larization
0 5 10 15 20 25 30
10 −20 10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * || 2 F(x (k) ) − F(x * )
Figure 4.2.8: Without
regularization
0 5 10 15 20 25 30
10 −20 10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * || 2 F(x (k) ) − F(x * )
Figure 4.2.9: Without
regularization and with
weighting
Theonvergene is muh slower inall three ases ompared to gures 4.2.2
and 4.2.3. There is only a little dierene in the iteration sequenes in
g-ures 4.2.7, 4.2.8 and 4.2.9 in the last iterations, when we are lose to the
optimizer. The onvergene rateisthe same for all threeases.
The table below shows the values of
k x (k+1) − x ∗ k 2 / k x (k) − x ∗ k 2
fork = 1, . . . , 24
orrespondingto the results of gure4.2.7.k
k x (k+1) − x ∗ k 2
k x (k) − x ∗ k 2
k
k x (k+1) − x ∗ k 2
k x (k) − x ∗ k 2
k
k x (k+1) − x ∗ k 2
k x (k) − x ∗ k 2
1 6.4062e-01 9 2.0771e-01 17 2.0771e-01
2 3.4325e-01 10 2.0770e-01 18 2.0768e-01
3 1.6551e-01 11 2.0770e-01 19 2.0758e-01
4 2.1301e-01 12 2.0769e-01 20 2.0725e-01
5 2.0869e-01 13 2.0769e-01 21 2.0764e-01
6 2.0793e-01 14 2.0769e-01 22 1.0000e+00
7 2.0778e-01 15 2.0769e-01 23 2.0577e-01
8 2.0774e-01 16 2.0770e-01 24 1.9426e-01
We notethat theasymptoti error onstant is approximately
0.2
fromiter-ation
4
to21
. Theresults indiatelinear onvergene.In the ase of the redued parameter vetor we have
n p = 5
unknownpa-rametersineveryParameterExtrationproblem. Thetransformationofthe
nemodelparametersisdenedbyanon-diagonalmatrix
C
,andapparentlythisreatesproblems,whenaligningthesurrogatemodelwiththenemodel.
Ingure4.2.10theiteration pointsfromgure4.2.8areseenintheontour
plot of
F = k f (x) k 2 2
. It is similar to gure4.2.4, exept for the fat thatalotof pointsare lusterednearthe optimizer
x ∗ = [1.2718 , 0.4951]
T.−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2
−1.5
−1
−0.5 0 0.5 1 1.5 2
Iteration sequence
x 1 x 2
Iteration points
Figure 4.2.10: Sequene of iteration points
Optimal Mapping Parameters
TheRosenbrokfuntionisspeialinthe sense,thatthetwo response
fun-tions are qualitatively dierent. The rst response is a quadrati funtion
and depends on both
z 1
andz 2
, whereas the seond is linear and onlyde-pends on
z 1
. TheJaobian matrix is:− 20z 1 10z 2
− 1 0
Sine
∂c 2 /∂z 2 = 0
and with referene to setion 3.4 equation (3.4.7) thismeans that:
∂s 2 (x, p)
∂p = α 2 c ′ 2,z (z) · H
= α 2
− 1 0
x 1 x 2 0 0 1 0 0 0 x 1 x 2 0 1
= − α 2
x 1 x 2 0 0 1 0
We have no information of the mapping parameters
A 21
,A 22
andb 2
on-erning
z 2
, sine all the partial derivatives wrt. these parameters are zero.This inuenes the Parameter Extration for response funtion number
2
.The variables are never hanged during the residual optimization, and the
nal values are
A 21 = 0
,A 22 = 1
andb 2 = 0
orresponding to the initial values.This theory is onrmed when looking at the results from the Spae
Map-pingalgorithm. Theoptimalparametersets arefortheregularized aseand
A 1 =
0.9567 − 0.1489 0.0370 0.9930
b 1 =
− 0.0678 0
α 1 = 0.9899 A 2 =
1.0329 − 0.1878
0 1
b 2 = 0
0
α 2 = 1.0649
For the ase when
A
is redued to a diagonal matrix we get the optimalmapping parameters withuseofthe regularization term(gure 4.2.7):
A 1 =
0.7811 0 0 1.1721
b 1 =
0.1609 0
α 1 = 1.1091 A 2 =
1 0 0 1
b 2 = 0
0
α 2 = 1
Theresults withno regularization (gure 4.2.8)arenot qualitatively
dier-ent:
A 1 =
0.7923 0 0 1.3003
b 1 =
0.2548 0
α 1 = 0.9997 A 2 =
1 0 0 1
b 2 = 0
0
α 2 = 1
It is noted, that the mapping parameters for response funtion
2
areiden-tialwith the initial mapping parameters in both ases. Thisprobably has
something to do with the fat, that the seond response funtion is linear
andonly dependsontherst variable.
Diret Optimization
We nallypresent theresults from diretoptimization of thene modelby
the two algorithms diret and diretd from the SMIS framework
imple-mentedby FrankPedersen.
Bothiteration sequenes in gure4.2.11 onverge very slowly,whih isalso
seenfromtheplots ingure4.2.12.
Fromthe given initial guess
x (0) = [ − 1.2 , 1]
theiterates move through thevalley with a large number of small steps towards the optimizer. This
be-haviouroftheiterationsequeneisavoidedfortheSpaeMappingalgorithm.
Itisobvious, thattheSpaeMappingalgorithm ismuhmoreeient than
0 10 20 30 40 50 10 −20
10 −15 10 −10 10 −5 10 0 10 5
Performance
Iteration
||x (k) −x * ||
2 F(x (k) ) − F(x * )
0 5 10 15 20 25 30
10 −20 10 −15 10 −10 10 −5 10 0 10 5 10 10
Performance
Iteration
||x (k) −x * ||
2 F(x (k) ) − F(x * )
Figure4.2.11: Performaneofdiretoptimization ofthene model('diret'
left,'diretd'right)
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2
−1.5
−1
−0.5 0 0.5 1 1.5 2
Iteration sequence
x 1 x 2
Iteration points
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2
−1.5
−1
−0.5 0 0.5 1 1.5 2
Iteration sequence
x 1 x 2
Iteration points
Figure 4.2.12: Iteration sequenes for diret optimization of the ne model
('diret'left,'diretd' right)
Summary of the Results
•
The regularization seems to have slightly negative eet on the on-vergene speed.•
Thenormalizationfatorshavepratiallynoeetontheonvergene speed.•
Theeetoftheweightingfatorsisnotpossibletoinvestigateforthis problem, beause the optimizer isfound, before thehosen weightingstrategy hasanyinuene.
•
The redution of the mapping parameters results in a muh sloweronvergene rate.
•
Theoptimalmappingparameters arepratially notinuenedbytheregularization term.