Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Optimal Solvers for Elliptic Optimal ControlProblems
F. Ibrahim, S. Turek and M. Köster
Institut for Applied Mathematics and Numerics (LSIII)Technische Universität Dortmund
ICAAMMP 2014Istanbul, Turkey, 20.08.2014
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Outline
Motivating Examples
Existence and Uniqueness
Optimize first, then discretize
Numerical results
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Outline
Motivating Examples
Existence and Uniqueness
Optimize first, then discretize
Numerical results
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Outline
Motivating Examples
Existence and Uniqueness
Optimize first, then discretize
Numerical results
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Outline
Motivating Examples
Existence and Uniqueness
Optimize first, then discretize
Numerical results
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
References
Optimal Control of System Governed by Partial DifferentialEquations: Lions (1971)Optimal Steuerung partial DifferentialgleichungenTheorie:Tröltzsch (2005)
Exact and Approximate Controllability for DistributedParameter Systems: A Numerical Approach:Lions, Glowinskiand He (2008)Optimization with PDE Constraints:Ulbrich and others (2009)Optimal Control of Partial Differential Equations:Theory,Methods and Applications Tröltzsch (2010)
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
References
Optimal Control of System Governed by Partial DifferentialEquations: Lions (1971)Optimal Steuerung partial DifferentialgleichungenTheorie:Tröltzsch (2005)Exact and Approximate Controllability for DistributedParameter Systems: A Numerical Approach:Lions, Glowinskiand He (2008)
Optimization with PDE Constraints:Ulbrich and others (2009)Optimal Control of Partial Differential Equations:Theory,Methods and Applications Tröltzsch (2010)
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
References
Optimal Control of System Governed by Partial DifferentialEquations: Lions (1971)Optimal Steuerung partial DifferentialgleichungenTheorie:Tröltzsch (2005)Exact and Approximate Controllability for DistributedParameter Systems: A Numerical Approach:Lions, Glowinskiand He (2008)Optimization with PDE Constraints:Ulbrich and others (2009)
Optimal Control of Partial Differential Equations:Theory,Methods and Applications Tröltzsch (2010)
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
References
Optimal Control of System Governed by Partial DifferentialEquations: Lions (1971)Optimal Steuerung partial DifferentialgleichungenTheorie:Tröltzsch (2005)Exact and Approximate Controllability for DistributedParameter Systems: A Numerical Approach:Lions, Glowinskiand He (2008)Optimization with PDE Constraints:Ulbrich and others (2009)Optimal Control of Partial Differential Equations:Theory,Methods and Applications Tröltzsch (2010)
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
References
Optimal Control of System Governed by Partial DifferentialEquations: Lions (1971)Optimal Steuerung partial DifferentialgleichungenTheorie:Tröltzsch (2005)Exact and Approximate Controllability for DistributedParameter Systems: A Numerical Approach:Lions, Glowinskiand He (2008)Optimization with PDE Constraints:Ulbrich and others (2009)Optimal Control of Partial Differential Equations:Theory,Methods and Applications Tröltzsch (2010)
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Part 1: Motivating Examples
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Optimal stationary boundary temperature: Heating of a body Ωbya controlled boundary temperature u to reach the desiredtemperature yd
infy,u
J(y, u) =1
2||y − yd||2L2(Ω) +
α
2||u||2L2(Γ) (P)
subject to state equation
−∆y = 0 in Ω
y = u on Γ,
and control constraints
a ≤ u ≤ b on Γ
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Optimal stationary heat source: The body Ω is heated e.g. bymicrowaves. The goal is to find u that minmizes the distance of yand yd
infy,u
J(y, u) =1
2||y − yd||2L2(Ω) +
α
2||u||2L2(Ω) (P1)
subject to state equation
−∆y = u in Ω
y = 0 on Γ,
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Notation and dataJ is the objective functional and yd is the desired stateα > 0 is regularization parameterOptimal control problems with linear state equation andquadratic objective functional called linear-quadratic(P) is a linear-quadratic elliptic boundary control Problem(P1) is a linear-quadratic elliptic distributed control Problem
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Part 2:Existence and Uniqueness of optimal controlproblem
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Problem P1: Defining the control- to- state map G : u → yfrom L2(Ω) to H1
0 (Ω) , a linear and bounded operator whichassigns to a control u ∈ L2(Ω) the unique solution y = y(u)of the state equation,writing the state y = Gu
introduce the reduced objective functional
infy,u
Jred(u) :=1
2||Gu− yd||2L2(Ω) +
α
2||u||2L2(Ω)
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Problem P1: Defining the control- to- state map G : u → yfrom L2(Ω) to H1
0 (Ω) , a linear and bounded operator whichassigns to a control u ∈ L2(Ω) the unique solution y = y(u)of the state equation,writing the state y = Gu
introduce the reduced objective functional
infy,u
Jred(u) :=1
2||Gu− yd||2L2(Ω) +
α
2||u||2L2(Ω)
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
TheoremThe reduced optimal control problem has a unique optimal solution(y, u) ∈ H1
0 (Ω)× L2(Ω) .
ProofDirect method of the calculus of variations(minimizing sequence) .
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Part 3: Optimize first, then discretize
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Theorem
Let (y, u) ∈ H10 (Ω)× L2(Ω) be the optimal solution of Problem
P1. Then, there exists an adjoint state p ∈ H10 (Ω) as a weak
solution of−∆p = y − yd in Ω
p = 0 on Ω
and the control equation
p+ αu = 0 in Ω.
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Finite Element Approximation of Distributed Control ProblemLet Th := Th(Ω) be a shape-regular, quadrilateral triangulation andlet
Vh := vh ∈ C(Ω) | vh|T ∈ Q1(T ) , T ∈ Th(Ω), | vh|Γ = 0
be the FE space of continuous piecewise linear finite elements.Then a possible FE approximation of (P1) is as follows:
inf J(yh, uh) :=1
2
∫Ω
|yh − yd|2 dx +α
2
∫Ω
|uh|2 dx
subject toa(yh, vh) = (uh, vh), ∀vh ∈ Vh
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Optimality Conditions for the FE Distributed Control Problemthere exists an adjoint state ph ∈ Vh such that the triple(yh, uh, ph) satisfies the following optimality conditions,
a(yh, vh) = (uh, vh), ∀vh ∈ Vh
a(ph, vh) = (yh − yd, vh), ∀vh ∈ Vh
(uh, vh) = −α−1(ph, vh) ∀vh ∈ Vh
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
The discrete KKT system
Ah 0 −Mh
−Mh Ah 00 Mh αMh
yhphuh
=
0−ydh0
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
A Reduced KKT systemIf we substitute uin the state equation by means of the controlequation according to u = −α−1 ph, the discrete KKT system canbe stated as (
Ah α−1Mh
Mh Ah
)(yhph
)=
(0
−ydh
).
There exist also solution methods, which rely on a reduction of theKKT system. In this method, the optimality system is reduced to asingle integral equation for the control u
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
A Reduced KKT systemIf we substitute uin the state equation by means of the controlequation according to u = −α−1 ph, the discrete KKT system canbe stated as (
Ah α−1Mh
Mh Ah
)(yhph
)=
(0
−ydh
).
There exist also solution methods, which rely on a reduction of theKKT system. In this method, the optimality system is reduced to asingle integral equation for the control u
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
The integral equation methodWe introduce the operator A : y 7→ (−∆y) and the adjointoperator A∗ : p 7→ (−∆p). The KKT system reads in the followingform:
A = u in Ω
A∗ = y − yd in Ω
u =− α−1p in Ω.
The last equation can be reformulated by the first two equationsto the fixed point equation
u = (−α−1A−∗A−1)u− (−α−1A−∗)yd
=Ku+ q
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
which we can rewrite it in the compact form as follows:
(I − K)u =q.
Based on this compact form, we formulate the following fixedpoint iteration:
un :=Kun−1 + q, n ∈ N ,
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Part 4:Numerical results
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Example
u =sin(3πx1) sin(3πx2)
y =sin(3πx1) sin(3πx2)
p =− α sin(3πx1) sin(3πx2)
We get the corresponding desired state as
yd =18π2α sin(3πx1) sin(3πx2) + sin(3πx1) sin(3πx2).
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Table: Comparison of iterations to solve Example 1. with different valuesof α and different mesh sizes for tolerance of 1E − 12 for Multigrid withdifferent smoother (A Reduced KKT system)
smoother level \α 1 1E-02 1E-04 1E-08 1E-126 7 7 9 div. div.
Jacobi(0.5) 7 6 6 7 10 div.8 5 5 5 7 div.9 4 4 4 5 div.6 5 5 6 div. div.
Gaus Seidel 7 5 5 6 div. div.8 5 5 5 27 div.9 6 6 6 9 div.6 7 7 8 div. div.
SSOR(0.5) 7 6 6 7 div. div.8 6 6 6 10 div.9 6 6 6 8 div.
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Table: Comparison of iterations to solve Example 1. with different valuesof α and different mesh sizes for tolerance of 1E − 12 for Multigrid withdifferent smoother and preconditioner (A Reduced KKT system)
smoother Preconditioner level \α 1 1E-02 1E-04 1E-08 1E-126 7 8 19 div. div.
SSOR(1.5) 7 8 10 10 163 div.8 8 8 9 28 div.9 8 8 9 26 div.6 4 5 7 39∗ div.
BICGSTAB SSOR(0.5) 7 5 5 6 16 div.8 5 5 6 10 419 5 5 5 8 296 4 4 7 9∗ 9∗
SSOR(0.5) 7 5 5 6 8 98 5 5 5 7 10
GMRES(30) 9 5 5 5 7 10
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Table: Comparison of iterations to solve Example 1. with different valuesof α and different mesh sizes for tolerance of 1E − 12 for Multigrid withdifferent smoother and preconditioner ( KKT system)
smoother Preconditioner level \α 1 1E-02 1E-04 1E-08 1E-126 4 5 9 div div
SSOR(0.5) 7 5 6 8 137 div.8 5 5 6 13 div.
BICGSTAB 9 5 6 7 9 div.6 4 4 7 98∗ 104∗
SSOR(0.5) 7 5 5 6 35 7678 5 5 5 11 1106
GMRES(30) 9 5 5 5 7 23
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Table: Comparison of iterations to solve Example 1. with different valuesof α at different grid level for tolerance of 1E − 12 with different solver(integral equation)
solver level \α 1 1E-02 1E-04 1E-08 1E-126 2 2 2 576∗ -
BICGSTAB 7 2 2 2 1281 -8 2 2 2 2318 -9 2 2 2 2429 -6 2 2 4 3∗ 3∗
GMRES 7 2 2 3 7 88 2 2 3 4 49 2 2 4 4 4
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
Figure: Example 1: Change in L2-error with level for different values of α
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems
Motivating Examples Existence and Uniqueness Optimize first, then discretize Numerical results
TU Dortmund University
Optimal Solvers for Elliptic Optimal Control Problems