Adaptive methods for PDE-based optimal control with pointwise inequality constraints [Elektronische Ressource] / vorgelegt von Winnifried Wollner
143 pages
English

Découvre YouScribe en t'inscrivant gratuitement

Je m'inscris

Adaptive methods for PDE-based optimal control with pointwise inequality constraints [Elektronische Ressource] / vorgelegt von Winnifried Wollner

Découvre YouScribe en t'inscrivant gratuitement

Je m'inscris
Obtenez un accès à la bibliothèque pour le consulter en ligne
En savoir plus
143 pages
English
Obtenez un accès à la bibliothèque pour le consulter en ligne
En savoir plus

Description

Inaugural-DissertationzurErlangung der DoktorwürdederNaturwissenschaftlich-Mathematischen GesamtfakultätderRuprecht-Karls-UniversitätHeidelbergvorgelegt vonDiplom-Mathematiker Winnifried Wollneraus Henstedt-UlzburgTag der mündlichen Prüfung: 22.07.2010Adaptive Methods for PDE-based OptimalControl with Pointwise Inequality Constraints19. April 2010Gutachter: Prof. Dr. Rolf RannacherGutachter: Prof. Dr. Hans Georg BockAbstractThis work is devoted to the development of efficient numerical methods for a certain class ofPDE-based optimization problems. The optimization is constraint by an elliptic PDE. Inaddition to prior work in this context pointwise inequality constraints on the control andstate variable are considered. These problems are infinite dimensional and their solution canin general not be obtained exactly. Instead the solution of such problems means to find anapproximate solution. This is done by (approximately) solving for some set of first ordernecessary optimality conditions. Hence an efficient algorithm has to find such an approximatesolution with as little effort as possible while still being accurate enough for whatever thegoal of the computation is.The work at hand contributes to this goal by deriving a posteriori error estimates with respectto a given functional. These estimates are required for two purposes, first, to generate efficientmeshes for the solution of the PDEs required in the process of solving the necessary conditions.

Sujets

Informations

Publié par
Publié le 01 janvier 2010
Nombre de lectures 9
Langue English
Poids de l'ouvrage 9 Mo

Extrait

Inaugural-Dissertation
zur
Erlangung der Doktorwürde
der
Naturwissenschaftlich-Mathematischen Gesamtfakultät
der
Ruprecht-Karls-Universität
Heidelberg
vorgelegt von
Diplom-Mathematiker Winnifried Wollner
aus Henstedt-Ulzburg
Tag der mündlichen Prüfung: 22.07.2010Adaptive Methods for PDE-based Optimal
Control with Pointwise Inequality Constraints
19. April 2010
Gutachter: Prof. Dr. Rolf Rannacher
Gutachter: Prof. Dr. Hans Georg BockAbstract
This work is devoted to the development of efficient numerical methods for a certain class of
PDE-based optimization problems. The optimization is constraint by an elliptic PDE. In
addition to prior work in this context pointwise inequality constraints on the control and
state variable are considered. These problems are infinite dimensional and their solution can
in general not be obtained exactly. Instead the solution of such problems means to find an
approximate solution. This is done by (approximately) solving for some set of first order
necessary optimality conditions. Hence an efficient algorithm has to find such an approximate
solution with as little effort as possible while still being accurate enough for whatever the
goal of the computation is.
The work at hand contributes to this goal by deriving a posteriori error estimates with respect
to a given functional. These estimates are required for two purposes, first, to generate efficient
meshes for the solution of the PDEs required in the process of solving the necessary conditions.
Second, to choose several parameters that occur in order to regularize the problems at hand
in such a way that the regularization error is both small enough, to obtain a ‘good result’,
and yet large enough to have ‘easy to solve’ problems.
These a posteriori estimators are supplemented with a priori estimates in several cases where
non have been available in the literature for the problem class under consideration.
Finally, all theory and all heuristics will be substantiated with several numerical examples of
different complexity.
Zusammenfassung
Ziel der Arbeit ist es effiziente numerische Verfahren zur Lösung von PDE basierten Optimie-
rungsproblemen zu entwickeln. Hierbei betrachten wir als Nebenbedingung eine elliptische
PDE sowie im Unterschied zu früheren Arbeiten zusätzliche (Ungleichungs-) Beschränkungen
an die Kontroll- und Zustandsvariablen. Es handelt sich bei diesen Problemen um unendlich-
dimensionale Optimierungsprobleme, so dass die Lösung im Allgemeinen nicht exakt bestimmt
werden kann. Stattdessen wird eine Approximation bestimmt. Diese erhält man durch die
(approximative) Lösung geeigneter Systeme notwendiger Optimalitätsbedingungen. Ein effizi-
enter Algorithmus hat die Aufgabe, eine solche approximative Lösung mit so wenig Aufwand
wie möglich und dennoch hinreichend genau zu bestimmen.
Die vorliegende Arbeit leistet hierzu einen Beitrag indem a posteriori Fehlerschätzer, bezüglich
eines gegebenen Funktionals, hergeleitet werden. Diese werden aus zwei Gründen benötigt.
Zum Einen, um sparsame Gitter für die Lösung der auftretenden PDEs zu erzeugen. Zum
Anderen, um diverse Parameter zur Regularisierung des Problems derart zu steuern, dass
einerseits der Regularisierungsfehler „klein genug” ist und andererseits die Probleme noch
immer „einfach zu lösen” sind.
Ferner werden die a posteriori Schätzer durch a priori Fehleranalysen ergänzt sofern solche
für die betrachtete Problemklasse noch nicht in der Literatur verfügbar waren.
Schließlich werden die theoretischen Resultate und die verwendeten Heuristiken durch mehrere
Beispiele unterschiedlicher Komplexität untermauert.Contents
1 Introduction 1
2 Foundations 7
2.1 Basic Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Abstract Optimization Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Discretization of the State Constraint . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 Existence and Regularity 17
3.1 Results on Non-Smooth Domains . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.1.1 Existence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.1.2 Necessary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2
3.3 Existence with L -regularization . . . . . . . . . . . . . . . . . . . . . . . . . 26
4 A Priori Error Estimates 27
4.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.2 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.3 A Priori Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3.1 State Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3.2 Control . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5 Algorithms for State Constraints 41
5.1 Barrier Methods for First Order State Constraints . . . . . . . . . . . . . . . 42
5.1.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.1.2 Barrier Functional and its Subdifferentiability . . . . . . . . . . . . . . 46
5.1.3 Minimizers of Barrier Problems . . . . . . . . . . . . . . . . . . . . . . 49
5.1.4 Properties of the Central Path . . . . . . . . . . . . . . . . . . . . . . 54
5.1.5 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.2 Formal KKT-Conditions for Solution to Regularized Problems . . . . . . . . 61
5.3 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6 A Posteriori Error Estimates 65
6.1 Control Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.1.1 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.1.2 A Posteriori Error Estimation . . . . . . . . . . . . . . . . . . . . . . . 70
6.1.2.1 Error in the Cost Functional . . . . . . . . . . . . . . . . . . 71
iContents
6.1.2.2 Error in the Cost Functional Reviewed . . . . . . . . . . . . 74
6.1.2.3 Error in the Quantity of Interest . . . . . . . . . . . . . . . . 76
6.1.3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.2 Regularization Error for State Constraints . . . . . . . . . . . . . . . . . . . . 87
6.2.1 Estimates for the Cost Functional . . . . . . . . . . . . . . . . . . . . 87
6.2.1.1 Barrier Regularization without Control Constraints . . . . . 87
6.2.1.2 Illustration of the Results for Two Specific Types of Constraints 92
6.2.1.3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . 94
6.2.1.4 The Influence of the Approximations to the Estimate . . . . 101
6.2.1.5 Barrier Regularization with Control Constraints . . . . . . . 104
6.2.1.6 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . 106
6.2.1.7 Penalty . . . . . . . . . . . . . . . . . . . . . 109
7 Algorithmic Aspects 117
7.1 Control Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.2 Statets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
8 Conclusions and Outlook 121
Acknowledgments 123
Bibliography 125
ii1 Introduction
Thisworkisdevotedtothedevelopmentofefficientnumericalmethodsforsolvingoptimization
problems subject to an elliptic PDE constraint and additional (inequality) constraints on
the control variable and zero or first-order constraints on the state variable. Since these
problems exhibit very rough adjoint variables their solution usually requires some kind of
regularization.
To state things precise, we will consider optimization problems that are constrained by an
elliptic partial differential equation
A(q,u) =f.
If the control is given by external forces the operator A is typically given in the form
¯A(q,u) =A(u)−B(q)
¯
with a (nonlinear) elliptic operator A and a (usually linear) control operator B. On the other
hand, for ‘control by the coefficients‘ this simple splitting will not suffice. For computational
purposes it is more convenient to rewrite this equation in a variational form, with a suitable
trial space V to be specified later on. It reads as follows
a(q,u)(ϕ) = (f,ϕ) ∀ϕ∈V.
The target of the optimization is to minimize a given cost functional J(q,u), e.g., a tracking
type functional
1 αd 2 r
J(q,u) = ku−uk + kqk rL
2 r
for some r≥ 2. The emphasis of this thesis is the consideration of additional constraints on
ad
the control and state variable. Therefore let Q be a closed convex set and g a functional.
Then, we require the control and state variable to fulfill
ad
q∈Q , g(u,∇u)≤ 0.
ad
ˆ
A typical choice of Q are ‘box-constraints’, e.g., let a<b∈R =R∪{±∞} be given. We
set
adQ ={q|a≤q(x)≤b a.e.}.
As state constraints we will consider zero-order state constraints, e.g.,
g(u,∇u) =g(u) =u−ψ
or first-order state constraints, e.g.,
2g(u,∇u) =g(∇u) =|∇u| −ψ.
11 Introduction
To summarize: The general problem under consideration is given as

  • Univers Univers
  • Ebooks Ebooks
  • Livres audio Livres audio
  • Presse Presse
  • Podcasts Podcasts
  • BD BD
  • Documents Documents