4.4 From blak-box state-spae models to gray-box ones
4.4.1 An H ∞ -norm-based approah
As explained in Setion 4.1, instead of resortingto the standard
L 2
-norm, it issuggested herein using the
H ∞
-norm in order to estimate the unknown parametervetor
θ
aurately. More preisely, by assuming that, from the availabledata-set
{ u(k), y(k) } N−1 k=0
a reliable blak-box ontinuous-time representation( A , B , C )
has been estimated beforehand, i.e., a blak-box model able to apture aurately
the frequeny response of the real system is available, we aim at minimizing the
H ∞
-norm of the dierene given by Eq. (4.9), i.e.,J(θ) = k G(θ, ω) − G (ω) k 2 ∞ = k F(θ, ω) H F(θ, ω) k ∞ ,
(4.16)with
F(θ, ω) = G(θ, ω) − G (ω).
(4.17)Astandardapproahforthiskindof
H ∞
-norm-basedoptimizationproblemonsists in using the Kalman-Yakubovith-Popov lemma whih would herein lead tobilin-ear matrix inequalities [20℄. In this Chapter, we have hosen to minimize
J(θ)
byusing the proximity ontrol algorithm. As shown in [5, 6℄, the applied tehnique
in this hapter avoids the use of Lyapunov funtions, feature whih is beneial
beause these variables are often a soure of numerial troubleand inreased
om-putational load. The prie to pay for avoiding the previously mentioned numerial
troubles is the introdution of a semi-innite and non-smooth program. Indeed,
the optimization problem (4.16) is non-smooth and non-onvex with two soures
of non-smoothness, the innite max operator and the maximum eigenvalue
fun-tion, respetively [5℄. The use of this spei ost funtion dened by Eq. (4.16),
more preisely the
H ∞
-norm, an be motivated by the availability of quite new and eient optimization algorithmsdediated to the minimizationof non-smoothand non-onvex funtions with additionalstruture [4, 5, 6, 53, 8℄. The use of the
H ∞
-norm an also be justied by notiingthat, by denition, theH ∞
-norm is themaximal singular value of the omplex gain matrix over all the frequenies. Thus,
thanks tothenormproperty,if theost funtion
J (θ)
inEq.(4.16)issmallenough(i.e.,alotsmallerthan1),thenthedistanebetweentheatualsystemandthe
iden-tied one issmallaswell. Notie thatin [14,15℄asimilar ost funtionisproposed
to estimate the unknown parameter vetor of an identiable gray-box struture by
employing the hinfstrut Matlab funtion. Notie also that during the
devel-opments proposed in this thesis, the proximity ontrol algorithm[6℄ is used whih
an be seen as a novel version of the algorithm realized by the hinfstrut. This
hoie an mainly be explained by the fat that the hinfstrut funtion annot
dealwithunstable modelsduringitsoperation. Thisispartiularlyimportantwhen
J (θ)
is initialized, before the optimization, by uniform random numbers resulting sometimes in unstable models.Before desribing into detail the
H ∞
-norm-based method, it is important to point out that the availability of an aurate blak-box form( A , B , C )
is not astrong assumption. Indeed, this estimation step an be performed eiently by
using one of the many identiationtoolboxesavailablein the literature(the
Mat-lab System Identiation Toolbox 12
, the CONTSID Toolbox 13
or the LTI
Sys-tem Identiation Toolbox 14
to name few). In this thesis, more preisely in
Se-tion 6.2, a spei attention is paid to the Multivariable OutputError State-sPae
(MOESP) algorithms[144, 93℄ aswell as the SimpliedRened Instrumental
Vari-able method for Continuous-time systems (SRIVC) [46, 47℄ (see Setions 9.2-9.3 in
the Appendixforabriefoverviewofthe identiationmethods). Thishoieanbe
justied asfollows. First, theMOESP lass of subspae-basedalgorithms, available
through the LTI System Identiation Toolbox, is indeed full of tehniques able to
deliverreliable(suboptimal)blak-boxstate-spaerepresentationsfromthedata-set
{ u(k), y(k) } N−1 k=0
undermanydierentdisturbaneassumptions(whiteoutputnoise, olored output noise, white state noise, error in variables) [144℄. Notie also thatMOESP-like algorithmsyielding diretly ontinuous-time 15
blak-box models have
beendeveloped duringthelastdeade[98,13℄. Ontopofthat,theappliationofthe
subspae-based algorithmto get this initialstate-spae representation an be
justi-ed by two otherreasons. First,this step is sharedby the standard
L 2
-norm-based tehniques, at least in the blak-box model framework [78℄. Thus, this estimationstep of
( A , B , C )
, whih is ompulsory for future developments, does not inrease the omputational load of this novelH ∞
-based tehnique in omparison with theaforementioned
L 2
-norm-basedtehnique. Seond, the subspae-based methodsare numerially robust and are one-shot estimation proedures whih do not require,apart from an upper bound of the system order, any prior knowledge about the
system toidentify. Thus, the state-spae form
( A , B , C )
an beestimated withoutresortingtoany iterativealgorithm. Seond, theSRIVCmethod,whihan be
on-sideredasone ofthemosteientidentiationtehniquesforthediretestimation
of ontinuous-time models, is known to yield onsistent estimates when the model
belongs to the system lass [46℄. This tehnique is, more preisely, available in the
CONTSID ToolboxforidentifyingMISOmodels,aonstraintsatisedbythe
simu-12. http://www.mathworks.fr/fr/produts/sysid/.
13. http://www.iris.ran.uhp-nany.fr/ontsid/.
14. http://www.ds.tudelft.nl/~datadriven/lti/ltitoolbox_produt_page.html.
15. TheuseranfurthermoreombineaMOESPalgorithmyieldingadisrete-timestate-spae
latedsystemstudiedinSetion6.2. Notienallythattheimplementationavailable
inthe CONTSIDToolboxhas theadvantageofnot requiringany designparameters
tobespeied. Ofourse,users familiarwith othertehniquesan usetheirfavorite
algorithms to estimate the initial state-spae representations
( A , B , C )
as long asthese blak-box models mimi the system behavior aurately.
Thetransformationoftheinformationontainedintheinput-output(I/O)
data-set into a blak-box model an be interesting for three main reasons. First, the
(potentiallyhuge) availabledata-setof size
N (n u + n y )
is summedup into amodelontaining
n 2 x + n x (n y + n u ) + n u n y
parameters. Thus, the dimension of the in-formation used by the iterative algorithm required for theH ∞
-norm optimization, i.e, the seond step of the identiation proedure developed in this Chapter, is alot smaller. The omputational load of the proximity ontrolalgorithm used
here-after should then be redued. Seond, by having aess to blak-box models quite
easily from the available data-sets, it is straightforward to hek if the available
I/Oinformationisgoodenough by usingstandard toolsfor modelvalidation[78℄.
Third, the noise treatment, i.e., the attenuation of the noise eet for the
onsis-teny ofthe estimated parameters,is performedduringthis initialstep by hoosing
eientlythe tehniqueyieldingaonsistentform
( A , B , C )
fromthe(noisy)data-set
{ u(k), y(k) } N−1 k=0
. This problem is quite standard in LTI system identiation and, as said previously, an be solved eiently and in one shot by using speisubspae-based algorithms(see, e.g., [144℄,when the MOESP lass of tehniques is
handled). Of ourse, this rst step is not omputationally ost free. For instane,
the omputationalost of the dierent steps omposing the PO-MOESPalgorithm
have been evaluated and gathered in Table 4.2. These values show that this step
anbequitenumeriallydemanding 16
when,e.g.,MOESP algorithmsareusedbut,
fortunately, should be performedonly one.
Step Multipliations
RQfatorization
2(α 2 + β 2 )(N − (α + β)/3)
Singularvalue deomposition
αn 2 y f 2
Extration of
A
andC n y f n 2 x
Estimationof
B n 3 u (n x + n y ) 3
Table 4.2: Computation load of the PO-MOESP algorithm [144℄.
f
andp
arethe user-dened future and past window hyper-parameters hosen greaterthan
n x
.α = (n u + n y )p
andβ = (n u + n y )f
. See [144, Paragraph 9.6.1℄ for a summary ofthe main steps omposing the PO-MOESPalgorithm.
As mentioned in the previous Chapter, the proximity ontrol algorithm is
de-signed to minimizesuh ost-funtions as the one dened by Eq. (4.16). Here, the
followingspeiset-upisintroduedbeausethe proximityontrolmethodisused
tominimizethefrequeny-domaindistane"betweenblakandgray-boxmodelsby
tuningtheparametersgathered inthevetor
θ
. As aonsequene,the Hamiltonian matrix dened by Eq. (3.13) and used during the proximity ontrol algorithm has16. Ifneessary,thePO-MOESPalgorithmanbereplaedbyareursivesubspae-based
algo-the following inner matries, e.g.,
E A = A(θ) − A , E B = B(θ) − B , E C = C(θ) − C .
(4.18)In the next Subsetion, the onvergene properties of the proximity ontrol
algo-rithm is desribed by onsidering the spei ase when
F ( θ , ω)
has a state-spaerepresentation.