crank-nicolson for optimal control problems with evolution equations · pdf filecrank-nicolson...
TRANSCRIPT
![Page 1: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/1.jpg)
Crank-Nicolson for optimal control problems withevolution equations
Optimize then Discretize == Discretize then Optimize
Thomas G. Flaig1
1Institut für Mathematik und BauinformatikFakultät für Bauingenieur- und Vermessungswesen
März 2009
Thomas Flaig Crank-Nicolson for optimal control 1
![Page 2: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/2.jpg)
1 Introduction
2 Derivation of the schemeDiscretize then OptimizeOptimize then Discretize
3 Numerical examplesSolution algorithmPDE Case
Only time disretization errorAlso spacial error
Thomas Flaig Crank-Nicolson for optimal control 2
![Page 3: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/3.jpg)
Introduction
1 Introduction
2 Derivation of the schemeDiscretize then OptimizeOptimize then Discretize
3 Numerical examplesSolution algorithmPDE Case
Only time disretization errorAlso spacial error
Thomas Flaig Crank-Nicolson for optimal control 3
![Page 4: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/4.jpg)
Introduction
Problem
min12
T∫0
‖y − yd‖2 + ν‖u‖2 d t
y,t + Ay = uy(0) = 0
A linear, self-adjoint operatorA ∈ Rn×n
A = ∆ + Boundary conditions
Thomas Flaig Crank-Nicolson for optimal control 4
![Page 5: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/5.jpg)
Introduction
Discretization
Optimize then Discretize == Discretize then OptimizeContinuous Optimal Control Problem
Discrete Optimization Problem
Continuous Optimality System
Discrete Optimality System
Discretize Discretize
Optimize
Optimize
initially based on a time stepping scheme
D. Meidner and B. Vexler and others: Space Time FEMinitially based on CG and DG schemes.second order convergence.
Thomas Flaig Crank-Nicolson for optimal control 5
![Page 6: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/6.jpg)
Derivation of the scheme
1 Introduction
2 Derivation of the schemeDiscretize then OptimizeOptimize then Discretize
3 Numerical examplesSolution algorithmPDE Case
Only time disretization errorAlso spacial error
Thomas Flaig Crank-Nicolson for optimal control 6
![Page 7: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/7.jpg)
Derivation of the scheme Discretize then Optimize
Discretization
min12
TZ0
‖y − yd‖2 + ν‖u‖2 d t
y,t + Ay = u
y(0) = 0
Problem
minτ
2
N−1∑k=0
(yk+1 − yk
2− yd k+1 − yd k
2
)2
+ντ
2
∑u2k+ 1
2
yk+1 − yk
τ+ A
yk+1 + yk
2= uk+ 1
2
y(0) = 0
Thomas Flaig Crank-Nicolson for optimal control 7
![Page 8: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/8.jpg)
Derivation of the scheme Discretize then Optimize
Lagrange functional
Lagrange functional
L(y , u, p) =τ
2
N−1∑k=0
(yk+1 − yk
2− yd k+1 − yd k
2
)2
+ντ
2
∑u2k+ 1
2︸ ︷︷ ︸Cost Functional
+
+ τ
N−1∑k=0
(yk+1 − yk
τ+ A
yk+1 + yk
2− uk+ 1
2
)︸ ︷︷ ︸
State Equation
· pk+ 12︸ ︷︷ ︸
Lagrange Multiplier
Thomas Flaig Crank-Nicolson for optimal control 8
![Page 9: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/9.jpg)
Derivation of the scheme Discretize then Optimize
State equation and variational equation
Lagrange functional
L(y , u, p) =τ
2
N−1Xk=0
„yk+1 − yk
2− yd k+1 − yd k
2
«2+ντ
2
Xu2k+ 1
2+
+ τ
N−1Xk=0
„yk+1 − yk
τ+ A
yk+1 + yk
2− uk+ 1
2
«· pk+ 1
2
∂L(y , u, p)
∂pk+ 12
= 0∂L(y , u, p)
∂uk+ 12
= 0
yk+1 − yk
τ+ A
yk+1 + yk
2= uk+ 1
2νuk+ 1
2= pk+ 1
2
Thomas Flaig Crank-Nicolson for optimal control 9
![Page 10: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/10.jpg)
Derivation of the scheme Discretize then Optimize
Adjoint equation (inner nodes)Lagrange functional
L(y , u, p) =τ
2
N−1Xk=0
„yk+1 − yk
2− yd k+1 − yd k
2
«2+ντ
2
Xu2k+ 1
2+
+ τ
N−1Xk=0
„yk+1 − yk
τ+ A
yk+1 + yk
2− uk+ 1
2
«· pk+ 1
2
∂L(y , u, p)
∂yk= 0
−pk+ 12
+ pk− 12
τ+ A
pk+ 12
+ pk− 12
2+
12
(yk+1 − yk
2− yd k+1 − yd k
2
)+
12
(yk − yk−1
2− yd k − yd k−1
2
)= 0
Thomas Flaig Crank-Nicolson for optimal control 10
![Page 11: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/11.jpg)
Derivation of the scheme Discretize then Optimize
Interpretation of RHSyi
yi−1
yi+1yi−1+yi
2+
yi+yi+12
2
ti−1 ti ti+1ti+ti+1
2ti−1+ti
2
yiyi−1
yi+1
yi−1+yi2
+yi+yi+1
2
2
ti−1 ti ti+1ti+ti+1
2ti−1+ti
2
Idea: Slightly different scheme(yk+1−yk
2 − yd k+1−yd k2
)+(
yk−yk−12 − yd k−yd k−1
2
)2
≈ yk − yd k
Thomas Flaig Crank-Nicolson for optimal control 11
![Page 12: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/12.jpg)
Derivation of the scheme Discretize then Optimize
Adjoint equation (last node)
Lagrange functional
L(y , u, p) =τ
2
N−1Xk=0
„yk+1 − yk
2− yd k+1 − yd k
2
«2+ντ
2
Xu2k+ 1
2+
+ τ
N−1Xk=0
„yk+1 − yk
τ+ A
yk+1 + yk
2− uk+ 1
2
«· pk+ 1
2
∂L(y , u, p)
∂yN= 0
pN− 12
τ+ A
pN− 12
2+
12
(yN − yN−1
2− yd N − yd N−1
2
)= 0
No need for ∂L(y ,u,p)∂y0
as y0 is given data.
Thomas Flaig Crank-Nicolson for optimal control 12
![Page 13: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/13.jpg)
Derivation of the scheme Discretize then Optimize
Basic facts about the discretization scheme
yk+1 − yk
τ+ A
yk+1 + yk
2= uk+ 1
2
pk+ 12− pk− 1
2
τ− A
pk+ 12
+ pk− 12
2= yk − yd k
−pN− 1
2
τ− A
pN− 12
2=
12
„yN − yN−1
2− yd N − yd N−1
2
«νuk+ 1
2= pk+ 1
2y0 = 0
Based on Crank-NicolsonBased on a time-stepping formulationDifferent discretization points for state and adjoint state
Thomas Flaig Crank-Nicolson for optimal control 13
![Page 14: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/14.jpg)
Derivation of the scheme Discretize then Optimize
Hamilton Systems and Störmer-Verlet-Scheme
Different discretization points for state and adjoint statealso used in Geometric Numerical Integration:Störmer-Verlet method (other names for this Method: see Hairer Lubich Wanner)
Hamilton System: Hamiltonian H(p, q) with
q,t = −H,p(p, q) p,t = H,q(p, q)
Optimal control problems are Hamilton Systems
H(y , p) =12〈y , y〉 − 〈yd , y〉+ 〈Ay , p〉 − 1
2ν〈p, p〉+ C
y,t = −H,p = −Ay +1ν
p
p,t = H,y = Ap + y − yd
Thomas Flaig Crank-Nicolson for optimal control 14
![Page 15: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/15.jpg)
Derivation of the scheme Optimize then Discretize
1 Introduction
2 Derivation of the schemeDiscretize then OptimizeOptimize then Discretize
3 Numerical examplesSolution algorithmPDE Case
Only time disretization errorAlso spacial error
Thomas Flaig Crank-Nicolson for optimal control 15
![Page 16: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/16.jpg)
Derivation of the scheme Optimize then Discretize
Optimize then Discretize
min12
TZ0
‖y − yd‖2 + ν‖u‖2 d t
y,t + Ay = u y(0) = 0
Lagrange functional
L(y , u, p) =12
T∫0
‖y − yd‖2 + ν‖u‖2 d t
︸ ︷︷ ︸Cost Functional
+
+
∫ T
0(y,t + Ay − u)︸ ︷︷ ︸
State Equation
· p︸︷︷︸Lagrange Multiplier
d t
Thomas Flaig Crank-Nicolson for optimal control 16
![Page 17: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/17.jpg)
Derivation of the scheme Optimize then Discretize
Optimality system
Lagrange functional
L(y , u, p) =12
TZ0
‖y − yd‖2 + ν‖u‖2 d t+
+
Z T
0(y,t + Ay − u) · p d t
Optimality system
y,t + Ay = u p,t − Ap = y − yd
y(0) = 0 p(T ) = 0νu = p
Thomas Flaig Crank-Nicolson for optimal control 17
![Page 18: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/18.jpg)
Derivation of the scheme Optimize then Discretize
Discretization State and variational equation
Optimality system
y,t + Ay = u p,t − Ap = y − yd
y(0) = 0 p(T ) = 0
νu = p
yk+1 − yk
τ+ A
yk+1 + yk
2= uk+ 1
2
y0 = 0
uk+ 12
=1ν
pk+ 12
Thomas Flaig Crank-Nicolson for optimal control 18
![Page 19: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/19.jpg)
Derivation of the scheme Optimize then Discretize
Discretization adjoint state
Optimality system
y,t + Ay = u p,t − Ap = y − yd
y(0) = 0 p(T ) = 0
νu = p
Discretization of p needed in pk+ 12. Therfore:
p,t − Ap = 0 in (T ; T +τ
2]
p,t − Ap = y − yd in (0; T ]
p(T +τ
2) = 0
easy to see p ≡ p in (0; T ]
Thomas Flaig Crank-Nicolson for optimal control 19
![Page 20: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/20.jpg)
Derivation of the scheme Optimize then Discretize
Discretization adjoint state (inner nodes)
Crank-Nicolson for
p,t − Ap = 0 in (T ; T +τ
2]
p,t − Ap = y − yd in (0; T ]
p(T +τ
2) = 0
Inner nodes
midpoint rule:pk+ 1
2− pk− 1
2
τ− A
pk+ 12
+ pk− 12
2= yk − yd ,k
trapezoidal rule:pk+ 1
2− pk− 1
2
τ− A
pk+ 12
+ pk− 12
2=
=12
(yk+1 − yk
2− yd k+1 − yd k
2
)+
12
(yk − yk−1
2− yd k − yd k−1
2
)
Thomas Flaig Crank-Nicolson for optimal control 20
![Page 21: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/21.jpg)
Derivation of the scheme Optimize then Discretize
Discretization adjoint state (last node)
p,t − Ap = 0 in (T ; T +τ
2]
p,t − Ap = y − yd in (0; T ]
p(T +τ
2) = 0
trapezoidal rule:pN+ 1
2− pN− 1
2
τ− A
pN+ 12
+ pN− 12
2=
=12
(yN+1 − yN
2− yd N+1 − yd N
2
)+
12
(yN − yN−1
2− yd N − yd N−1
2
)
Thomas Flaig Crank-Nicolson for optimal control 21
![Page 22: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/22.jpg)
Derivation of the scheme Optimize then Discretize
Discretization adjoint state (last node)
p,t − Ap = 0 in (T ; T +τ
2]
p,t − Ap = y − yd in (0; T ]
p(T +τ
2) = 0
trapezoidal rule: −pN− 1
2
τ− A
pk− 12
2=
=12
(yN − yN−1
2− yd N − yd N−1
2
)
Thomas Flaig Crank-Nicolson for optimal control 21
![Page 23: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/23.jpg)
Derivation of the scheme Optimize then Discretize
Altogether
Again
yk+1 − yk
τ+ A
yk+1 − yk
2= uk+ 1
2
pk+ 12− pk− 1
2
τ− A
pk+ 12
+ pk− 12
2= yk − yd k
−pN− 1
2
τ− A
pN− 12
2=
12
(yN − yN−1
2− yd N − yd N−1
2
)νuk+ 1
2= pk+ 1
2y0 = 0
Thomas Flaig Crank-Nicolson for optimal control 22
![Page 24: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/24.jpg)
Numerical examples
1 Introduction
2 Derivation of the schemeDiscretize then OptimizeOptimize then Discretize
3 Numerical examplesSolution algorithmPDE Case
Only time disretization errorAlso spacial error
Thomas Flaig Crank-Nicolson for optimal control 23
![Page 25: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/25.jpg)
Numerical examples Solution algorithm
System to solveLinear System for the midpoint rule
0BBBBBBBBBBBBBB@
K −Mν
L K. . .
. . .. . .
. . .L K −M
ν
−M −K −L. . .
. . .. . .
−M −K −L−M
4 −M4 −K
1CCCCCCCCCCCCCCA
0BBBBBBBBBBBBBBBB@
y1......
yN
pi+ 12
...
...pN− 1
2
1CCCCCCCCCCCCCCCCA=
0BBBBBBBBBBBBB@
−Ly0
0...0
−Myd,1...
−Myd,N−1
−Myd ,N−1+yd ,N
4
1CCCCCCCCCCCCCA.
K =Mτ
+A2, L = −M
τ+
A2
Solution algorithm: Matlab-\Improvement: Parallelization of Matrix-Vector-Multiplication
or Multigrid.
Thomas Flaig Crank-Nicolson for optimal control 24
![Page 26: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/26.jpg)
Numerical examples Solution algorithm
Error measurement
maxti∈[0:τ :1]
((yh,i − y(ti , xj)
)T M(yh,i − y(ti , xj)
))1/2
maxti∈[0:τ :1]
((ph,i+ 1
2− p(ti+ 1
2, xj)
)TM(ph,i+ 1
2− p(ti+ 1
2, xj)
))1/2
Thomas Flaig Crank-Nicolson for optimal control 25
![Page 27: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/27.jpg)
Numerical examples PDE Case
Only time disretization error
Problem
min
1∫0
12‖y − yd‖2 +
ν
2‖u‖2 d t
y,t −∆y = u in (0; 1]× (0; 1)2
∂y∂n
= 0 on (0; 1]× ∂(0; 1)2
y(0) = 0 in 0 × (0; 1)2
yd = t3 − 1.5t − 6νt + 3ν in (0; 1]× (0; 1)2
Thomas Flaig Crank-Nicolson for optimal control 26
![Page 28: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/28.jpg)
Numerical examples PDE Case
ν = 10−5
10−4
10−3
10−2
10−1
100
10−14
10−12
10−10
10−8
10−6
10−4
10−2
100
green: τ, τ2, τ3
τ y y p p
0.25 1.00 0.02 2.00 1.990.2 1.00 0.08 2.00 1.960.1 1.01 0.52 1.99 1.890.05 1.02 1.77 1.97 1.850.02 1.10 2.00 1.86 1.790.01 1.37 2.00 1.64 1.740.005 1.76 2.00 1.59 1.750.002 2.00 2.00 1.75 1.810.001 2.02 2.00 1.88 1.890.0005 2.01 2.00 1.93 1.930.0002 2.00 2.00 1.97 1.970.0001 2.00 2.00 1.99 1.99
red p, blue y for the midpoint rulemagenta p and cyan y for the trapezoidal rule
Thomas Flaig Crank-Nicolson for optimal control 27
![Page 29: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/29.jpg)
Numerical examples PDE Case
Also spacial error
Problem
min
1∫0
12‖y − yd‖2 +
ν
2‖u‖2 d t
y,t −∆y = u∂y∂n
= 0, y(0) = 0
wa = e−13π
2t cos(πx1) cos(πx2)
wb = e13π
2t cos(πx1) cos(πx2)
yd = c7wa(t, x) + c8wb(t, x)
+ c9wa(0, x) + c10wb(0, x)
+ c11wa(1, x) + c12wb(1, x)
y = y(ν), p = p(ν), u = u(ν) (0; T ]× Ω = (0; 1]× (0; 1)2
Thomas Flaig Crank-Nicolson for optimal control 28
![Page 30: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/30.jpg)
Numerical examples PDE Case
10−2
10−1
100
10−6
10−5
10−4
10−3
10−2
10−1
100
4 space-refinements10
−210
−110
010
−6
10−5
10−4
10−3
10−2
10−1
100
5 space-refinementsgreen: τ, τ2, τ3
red p, blue y for the midpoint rulemagenta p and cyan y for the trapezoidal rule
Thomas Flaig Crank-Nicolson for optimal control 29
![Page 31: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/31.jpg)
Conclusions
time stepping scheme withOptimize then Discretize == Discretize then Optimizebased on Crank-Nicolson-schemerelationship to Störmer-Verlet-scheme / Hamilton systemsnumerical examplesFurther results
I interpretation as continuous Galerkin methodI variable time step sizeI proof of second order convergence
Further researchI constraintsI more efficient solver (multigrid)
Thomas Flaig Crank-Nicolson for optimal control 30
![Page 32: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/32.jpg)
Thank youfor your attention!
Thomas Flaig Crank-Nicolson for optimal control 31
![Page 33: Crank-Nicolson for optimal control problems with evolution equations · PDF fileCrank-Nicolson for optimal control problems with evolution equations OptimizethenDiscretize==DiscretizethenOptimize](https://reader031.vdocuments.mx/reader031/viewer/2022022504/5ab854547f8b9ab62f8c86e4/html5/thumbnails/33.jpg)
1 Introduction
2 Derivation of the schemeDiscretize then OptimizeOptimize then Discretize
3 Numerical examplesSolution algorithmPDE Case
Only time disretization errorAlso spacial error
Thomas Flaig Crank-Nicolson for optimal control 32