An iterative Bregman regularization method for optimal control

An iterative Bregman regularization method for
optimal control problems with inequality
constraints
Frank Pörner
Joint work with Daniel Wachsmuth
Universität Würzburg
Joint Annual Meeting of DMV and GAMM 2016
Outline
1 Introduction
2 Iterative Bregman regularization method
Bregman distance
Iterative Bregman regularization
Regularity Assumption
Convergence results
3 Outlook/open problems
Introduction
Problem setting:
Minimize
such that
1
kSu − zk2Y
2
ua ≤ u ≤ ub
a.e. in Ω
Assumptions
S : L2 (Ω) → Y linear and continuous, Y Hilbert space, z ∈ Y
Example: Y = L2 (Ω), S = (−∆)−1 : L2 (Ω) → H01 (Ω) ⊆ Y
Control constraints
ua , ub ∈ L∞ (Ω) with ua ≤ ub a.e. in Ω, Uad is the set of admissible
functions: Uad := {u ∈ L2 (Ω) : ua ≤ u ≤ ub }
Introduction
Uad weakly compact → solution u † exists
optimal state: y † = Su † unique
Problems
solutions may be unstable w.r.t. to pertubations
convergence of numerical schemes
Solution: Regularize the problem
Tikhonov regularization
Let uk be the solution of
min
u∈Uad
1
αk
kSu − zk2Y +
kuk2L2
2
2
(αk )k positive, monotonically decreasing with αk → 0
Let ū be solution with minimal L2 -norm, solving
min kukL2
s.t.
u ∈ Uad , Su = y †
Properties
(kuk kL2 )k monotonically increasing, kuk kL2 ≤ kūkL2
unconditional strong convergence uk → ū for αk → 0
Problems
For αk → 0 (T) becomes ill-conditioned
(T)
Iterated Tikhonov / Proximal Point Method
Let uk+1 be the solution of
min
u∈Uad
1
αk+1
kSu − zk2Y +
ku − uk k2L2
2
2
(αk )k positive, bounded: 0 < αk ≤ ᾱ
Properties
nice monotonicity properties
uk * u ∗ , u ∗ not necessarily minimum norm solution
[Rockafellar ’76], [Tichatschke]
Problems
No strong convergence: Counterexample by Güler
Source condition not applicable if y † 6= z
[Güler ’91]
Bregman distance
Solution: regularize with Bregman distance
Bregman distance associated with regularization function
J : L2 (Ω) → R
D λ (u, v ) := J(u) − J(v ) − (λ, u − v )
and subdifferential λ ∈ ∂J(v ).
Properties
D λ (u, v ) non-negative, convex w.r.t to u
D λ (u, v ) = 0 if v = u (if and only if for strictly convex J)
[Bregman ’67]
For J(u) = 21 kuk2L2 one has D λ (u, v ) = 12 ku − v k2L2
Bregman distance
We use
1
J(u) := kuk2L2 + IUad (u)
2
hence
∂J(u) 3 λ = u + w
with w ∈ ∂IUad (u)
For v ∈ Uad we obtain:
1
D λ (u, v ) = ku − v k2L2 + IUad (u)
2 Z
+
w (ua − u) dx +
{v =ua }
Z
w (ub − u) dx
{v =ub }
Bregman distance adds two parts that measures u on sets where
the control constraints are active for v .
Iterative Bregman regularization
Prototypical iterative method:
Algorithm A0
Let u0 ∈ Uad , λ0 ∈ ∂J(u0 ) and k = 1.
1. Solve for uk :
Minimize
1
kSu − zk2Y + αk D λk−1 (u, uk−1 ).
2
2. Choose λk ∈ ∂J(uk ).
3. Set k := k + 1, go back to 1.
Problems
How to choose u0 and λ0 ?
How to choose λk ?
Iterative Bregman regularization
Problems
How to choose u0 and λ0 ?
How to choose λk ?
Set
u0 := arg min
u∈L2 (Ω)
1
2
kukL2 + IUad (u) = PUad (0)
2
and
λ0 := 0 ∈ ∂J(u0 )
Iterative Bregman regularization
Problems
How to choose u0 and λ0 ?
How to choose λk ?
First order conditions for uk : ∃wk ∈ ∂IUad (uk ) such that
S ∗ (Suk − z) + αk (uk − λk−1 + wk ) = 0
Rearranging ( J(u) = 12 kuk2L2 + IUad (u) ):
λk := uk + wk =
1 ∗
S (z − Suk ) + λk−1 ∈ ∂J(uk )
αk
By induction:
k
X
1 ∗
λk :=
S (z − Sui ) ∈ ∂J(uk )
αi
i=1
Iterative Bregman regularization
Algorithm A
Let u0 = PUad (0) ∈ Uad , λ0 = 0 ∈ ∂J(u0 ) and k = 1.
1. Solve for uk :
Minimize
2. Set λk :=
k
P
i=1
1
kSu − zk2Y + αk D λk−1 (u, uk−1 ).
2
1 ∗
αi S (z
− Sui ).
3. Set k := k + 1, go back to 1.
Here: 0 < αk ≤ ᾱ.
Algorithm A is well-defined (see also [Osher, Burger, Goldfarb, Xu,
Yin 2005])
Iterative Bregman regularization
Monotonicity
(kSuk − zkY )k is monotonically decreasing
Convergence rate
kSuk −
zk2Y
†
− kSu −
zk2Y
≤c
k
X
1
αi
!−1
i=1
Compare to [Osher, Burger, Goldfarb, Xu, Yin 2005]
Weak convergence
Weak limit points of (uk )k are solutions of the original problem
and Suk → Su † .
Regularity Assumption
Regularity Assumption
Let u † be a solution of the original problem and assume that there
exists a set I ⊆ Ω, a function w ∈ Y , and positive constants κ, c
such that the following holds
(source condition) I ⊃ {x ∈ Ω : p † (x) = 0} and
χI u † = χI PUad (S ∗ w ),
(structure of active set) A := Ω \ I and for all ε > 0
|{x ∈ A : 0 < |p † (x)| < ε}| ≤ cεκ ,
(regularity of solution) S ∗ w ∈ L∞ (Ω).
[D. Wachsmuth & G. Wachsmuth 2011]
Regularity Assumption
Let u † satisfy regularity assumption.
Improved optimality condition
Improved optimality condition for u † :
1+ 1
κ
(S ∗ (Su † − z), v − u † ) ≥ ckv − u † kL1 (A)
Similar to [Seydenschwanz 2015]
∀v ∈ Uad
Regularity Assumption
Special Case 1: I = Ω
Obtain Source Condition: u † = PUad (S ∗ w ) - Equivalent to
existence of Lagrange Multipliers for the minimum norm problem:
min
u∈Uad
1
kuk2L2
2
s.t.
Su = y †
[Neubauer]
Special Case 2: I = ∅
Obtain purely structural assumption:
|{x ∈ Ω : 0 < |p † (x)| < ε}| ≤ cεκ ,
suited for bang-bang problems.
[Wachsmuth & Wachsmuth]
Convergence results
Convergence results under the regularity assumption
Special Case: I = Ω
Convergence:
lim
k→∞
1
kuk − u † k2L2 = 0
αk
1
kS(uk − u † )k2Y = 0
2
k→∞ αk
lim
Convergence rates:
kuk − u † k2L2
1
+ min
kS(ui − u † )k2Y ≤ c
i=1,...,k αi
k
X
1
αi
i=1
!−1
Convergence results
Convergence results under the regularity assumption
General case: I 6= Ω
Convergence:
lim kuk − u † kL2 = 0
k→∞
Convergence rates:
1
kuk − u † k2L2 + min
kS(ui − u † )k2Y
i=1,...,k αi


k
X
≤ c γk−1 + γk−1
αj−1 γj−κ 
j=1
with the abbreviation γk =
k
P
i=1
αi−1
Convergence results
Special case: αk := cα k −s with s ≥ 0, cα > 0:
General case: I 6= Ω

−κ(s+1)

k
kuk − u † k2L2 ≤ c k −(s+1) log k

 −(s+1)
k

−κ(s+1)−s

k
† 2
min kS(uj − u )kY ≤ c k −(2s+1) log k

j=1,...,k
 −(2s+1)
k
if κ < 1,
if κ = 1,
if κ > 1,
if κ < 1,
if κ = 1,
if κ > 1,
Outlook/open problems
Open Questions:
Stopping criteria
Discrepancy principles for disturbed data (a-priori,
a-posteriori)
Comparison of Bregman iteration with other methods
(Tikhonov, projected gradient,...)