Robust Optimization in Robust Control

Robust Optimization in Robust
Control
Carsten Scherer
Delft Center for Systems and Control (DCSC)
Delft University of Technology
The Netherlands
Supported by Dutch Technology Foundation
Outline
• Dissipativity and Linear Matrix Inequalities
• Linear parameter-varying system analysis and synthesis
• Robust LMI problems
• Matrix sum-of-squares relaxations
• Verification of relaxation exactness
• Examples
2/48
Carsten Scherer
Biking = Optimized Control
feedback
Riding a bicycle means
dynamical system
3/48
Carsten Scherer
Biking = Optimized Control
feedback
Riding a bicycle means
• Stabilization
keep bike upright
dynamical system
3/48
Carsten Scherer
Biking = Optimized Control
feedback
Riding a bicycle means
• Stabilization
keep bike upright
• Tracking
follow a given path
dynamical system
3/48
Carsten Scherer
Biking = Optimized Control
feedback
Riding a bicycle means
• Stabilization
keep bike upright
• Tracking
follow a given path
• Disturbance Suppression
counteract wind gusts
dynamical system
3/48
Carsten Scherer
Biking = Optimized Control
feedback
Riding a bicycle means
• Stabilization
keep bike upright
• Tracking
follow a given path
• Disturbance Suppression
counteract wind gusts
A biker is a feedback controller!
dynamical system
3/48
Carsten Scherer
Simple Model of Bike Dynamics
Forces
steering
angle u
velocity
v
Newton’s law:
"
#
" #
0 1
1
ẋ =
x+
u
γ 0
0
h
i
2
y =
αv βv x.
tilt angle y
Dynamical system u → y.
4/48
Carsten Scherer
Simple Model of Bike Dynamics
Forces
steering
angle u
velocity
Newton’s law:
"
#
" #
0 1
1
ẋ =
x+
u
γ 0
0
h
i
2
y =
αv βv x.
v
tilt angle y
Dynamical system u → y.
Model neglects many aspects:
non-linearities, resonances in frame, ...
Mismatch between real system and model: Uncertainty.
4/48
Carsten Scherer
Controlling a Bike
Keep Small
Wind Gust
Bike Model
Tilt angle
Steering angle
Controller
Automatic Control
• Given a model design controller which performs multiple tasks
• Develop algorithms for optimal design of controllers
• Include effects of uncertainties: Robust optimal design
5/48
Carsten Scherer
How to Handle Neglected Effects?
Consider static nonlinearity:
∆
ẋ = Ax + B∆(Cx)
w
z
Pull out nonlinearity:
ẋ = Ax + Bw, w = ∆(z)
Nominal
System
z = Cx
6/48
Carsten Scherer
How to Handle Neglected Effects?
Consider static nonlinearity:
∆
ẋ = Ax + B∆(Cx)
w
z
Pull out nonlinearity:
ẋ = Ax + Bw, w = ∆(z)
Nominal
System
z = Cx
Linear Fractional Representation (LFR)
• Naturally related to modeling of system interconnections
• Very flexible: Easy extension to dynamic uncertainties
6/48
Carsten Scherer
Nominal Stability
ẋ = Ax asymptotically stable iff exists solution X of LMI
X 0 and AT X + XA ≺ 0.
7/48
Carsten Scherer
Nominal Stability
ẋ = Ax asymptotically stable iff exists solution X of LMI
X 0 and AT X + XA ≺ 0.
Proof with Lyapunov function V (x) := xT Xx since
d
V (x(t)) = x(t)T AT X + XA x(t) < 0.
dt
7/48
Carsten Scherer
Nominal Stability
ẋ = Ax asymptotically stable iff exists solution X of LMI
X 0 and AT X + XA ≺ 0.
Proof with Lyapunov function V (x) := xT Xx since
d
V (x(t)) = x(t)T AT X + XA x(t) < 0.
dt
All control systems must at least be stabilized ...
... LMI stability characterization for LTI systems ...
... first reason for success of SDP’s in control.
7/48
Carsten Scherer
Robust Stability
System with linear fractional representation
ẋ = Ax + Bw, z = Cx + Dw, w = ∆(z).
is exponentially stable if ...
8/48
Carsten Scherer
Robust Stability
System with linear fractional representation
ẋ = Ax + Bw, z = Cx + Dw, w = ∆(z).
is exponentially stable if ...
... exists X 0 and Hermitian multiplier P with
"
# "
#T "
#
AT X + XA XB
0 I
0 I
+
P
≺0
BT X
0
C D
C D
and such that
"
∆(z)
z
#T "
P
∆(z)
z
#
≥ 0 for all z.
8/48
Carsten Scherer
Proof
Fully elementary Lyapunov arguments:
"
#T "
# "
#T "
# "
#

T
x(t)
A X + XA XB
0 I
0 I  x(t)
+
P
≤0

w(t)
BT X
0
C D
C D  w(t)
⇐⇒
d
x(t)T Xx(t) +
dt
"
w(t)
z(t)
#T "
P
w(t)
z(t)
#
≤0
9/48
Carsten Scherer
Proof
Fully elementary Lyapunov arguments:
"
#T "
# "
#T "
# "
#

T
x(t)
A X + XA XB
0 I
0 I  x(t)
+
P
≤0

w(t)
BT X
0
C D
C D  w(t)
⇐⇒
d
x(t)T Xx(t) +
dt
"
w(t)
z(t)
#T "
P
w(t)
z(t)
#
≤0
⇐⇒
d
x(t)T Xx(t) +
dt
"
|
∆(z(t))
z(t)
#T "
P
∆(z(t))
z(t)
#
≤0
{z
}
can be dropped since ≥ 0
9/48
Carsten Scherer
Dissipativity (Willems 1972)
"
∃ X:
AT X + XA XB
BT X
0
#
"
+
0 I
C D
#T "
P
0 I
C D
#
≺0
⇐⇒
ẋ = Ax + Bw
z = Cx + Dw
)
"
w
dissipative for supply s(w, z) = −
z
#T "
P
#
w
.
z
No loss of generality: Quadratic storage function V (x) = xT Xx
Z
t2
s(w(t), z(t)) dt ≤ V (x(t1 ))
V (x(t2 )) −
t1
Initial energy at time t1
Final energy at time t2
Supplied energy
10/48
Carsten Scherer
Dissipativity
System Movement
Initial Energy
Final Energy
Supplied Energy
• Is generalization of physical law of conservation of energy
• Extends Lyapunov theory to open dynamical systems
• Fundamental concept behind LMI’s in control
11/48
Carsten Scherer
Energy-Gain = Dissipation
Energy gain is minimal γ with
Z ∞
Z ∞
2
kz(t)k dt ≤ γ
kw(t)k2 dt
0
w
z
System
y
0
u
Controller
Closed-loop system description:
ẋ = Ax + Bw
z = Cx + Dw
z
Closed-Loop
System
w
12/48
Carsten Scherer
Energy-Gain = Dissipation
Energy gain is minimal γ with
Z ∞
Z ∞
2
kz(t)k dt ≤ γ
kw(t)k2 dt
0
w
z
System
y
0
u
Controller
Closed-loop system description:
ẋ = Ax + Bw
z = Cx + Dw
z
Closed-Loop
System
w
Minimize γ such that exists X with
"
# "
#T"
#"
#
AT X + XA XB
0 I
−γI 0
0 I
+
≺ 0.
BT X
0
C D
0
I
C D
12/48
Carsten Scherer
Robust Stability = Separation by Dissipation
Uncertainty anti-dissipative


T
∆(z) 
P
z



∆(z) 
0
z
∆
z
w
Nominal
System
Separation with
same Multiplier P
Nominal system dissipative




T
AT X + XA XB   0 I 
+
P
BT X
0
C D



0 I 
≺0
C D
13/48
Carsten Scherer
Dissipativity – A colorful concept
Lyapunov
stability
backstepping
integral quadratic
constraints
theory of MPC
small gain
theorem
Hamilton-Jacobi-Bellman
structured
singular value
circle criterion
Outline
• Dissipativity and Linear Matrix Inequalities
• Linear parameter-varying system analysis and synthesis
• Robust LMI problems
• Matrix sum-of-squares relaxations
• Verification of relaxation exactness
• Examples
14/48
Carsten Scherer
Linear Parameter Varying Systems
Bike dynamics model:
"
#
" #
0 1
1
ẋ =
x+
u
γ 0
0
i
h
2
y =
αv βv x.
feedback
Depends on time-varying velocity.
dynamical system
15/48
Carsten Scherer
Linear Parameter Varying Systems
Bike dynamics model:
"
#
" #
0 1
1
ẋ =
x+
u
γ 0
0
i
h
2
y =
αv βv x.
feedback
Depends on time-varying velocity.
Feedback controller ... adapted to
on-line measured value of parameter.
dynamical system
15/48
Carsten Scherer
Wafer-Stage: The System
Waferscanners
Lens system
Silicon wafer
on wafer-stage
Positioning
16/48
Carsten Scherer
Wafer-Stage: Performance Limitations
wafer
Moving wafer-stage …
… has to accurately track surface
Movements … oscillations … performance limitations
17/48
Carsten Scherer
Wafer-Stage: Illustration of Dimensions
Accuracy
Layer with 17 silicon atoms
18/48
Carsten Scherer
Wafer-Stage: Parameter Dependence of Model
Position-dependent
resonances
... long-term collaboration DCSC (TU-Delft) and CFT (Philips)
19/48
Carsten Scherer
Linear Parameter Varying Systems
parameter
curve ¼(t)
LPV system
¼2
¼3
ẋ(t) = A(π(t))x(t)
with value & velocity bounds
¼1
π(t) ∈ Π and π̇(t) ∈ Φ.
¼5
¼4
20/48
Carsten Scherer
Linear Parameter Varying Systems
parameter
curve ¼(t)
LPV system
¼2
¼3
ẋ(t) = A(π(t))x(t)
¼1
with value & velocity bounds
π(t) ∈ Π and π̇(t) ∈ Φ.
¼5
¼4
Exponentially stable if exists smooth X(π) with
X
X(π) 0,
∂j X(π)φj + A(π)T X(π) + X(π)A(π) ≺ 0
j
for all δ = (π, φ) ∈ Π × Φ.
20/48
Carsten Scherer
Proof
For arbitrary parameter trajectory π(t) and system trajectory x(t):
d
x(t)T X(π(t))x(t) =
dt
= 2x(t)T X(π(t))ẋ(t) +
X
x(t)T [∂j X(π(t))π̇j (t)] x(t)
j
Algebraic inequality guarantees strict negativity ...
... standard Lyapunov arguments finish proof.
21/48
Carsten Scherer
Performance
Input-output LPV system
ẋ(t) = A(π(t))x(t) + B(π(t))w(t)
z(t) = C(π(t))x(t) + D(π(t))w(t)
L2 -gain smaller than γ if exists smooth X(π) 0 with
T P



X(π)φ
X(π)
0
0
I
0
∂
I
0
j
j
 j



 A(π) B(π) 
 A(π) B(π)  
X(π)
0
0
0


≺ 0


 0

 0
2
I
0
0
−γ
I
0
I




C(π) D(π)
0
0
0 I
C(π) D(π)
for all δ = (π, φ) ∈ Π × Φ.
22/48
Carsten Scherer
Synthesis of LPV Controller
System and Controller ...
z¢
¢(±(t))
w¢
... Linear Fractional Representations
z
Design Goals
• exponential stability
• minimal energy gain
w
System
y
u
Controller
Convex Optimization ...
... gain-scheduled controller
zc
¢c(±(t))
wc
23/48
Carsten Scherer
Reduction to Finite Dimensions
Stability inequality
X
∂j X(π)φj + A(π)T X(π) + X(π)A(π) ≺ 0
j
24/48
Carsten Scherer
Reduction to Finite Dimensions
Stability inequality
X
∂j X(π)φj + A(π)T X(π) + X(π)A(π) ≺ 0
j
Search X(π) in span of X1 (π), . . . , Xn (π).
Results in semi-infinite LMI problem
!
X
k
xk
X
∂j Xk (π)φj + A(π)T Xk (π) + Xk (π)A(π)
≺ 0.
j
24/48
Carsten Scherer
Reduction to Finite Dimensions
Stability inequality
X
∂j X(π)φj + A(π)T X(π) + X(π)A(π) ≺ 0
j
Search X(π) in span of X1 (π), . . . , Xn (π).
Results in semi-infinite LMI problem
!
X
k
xk
X
∂j Xk (π)φj + A(π)T Xk (π) + Xk (π)A(π)
≺ 0.
j
With δ = (π, φ) ∈ Π × Φ = δ results in generic formulation
F0 (δ) + x1 F1 (δ) + · · · + xn Fn (δ) ≺ 0.
24/48
Carsten Scherer
Outline
• Dissipativity and Linear Matrix Inequalities
• Linear parameter-varying system analysis and synthesis
• Robust LMI problems
• Matrix sum-of-squares relaxations
• Verification of relaxation exactness
• Examples
24/48
Carsten Scherer
Robust LMI Problems
Infimize cT x over LMI-constraint
F0 (δ) + x1 F1 (δ) + · · · + xn Fn (δ) ≺ 0 for all δ ∈ δ.
• Have seen concrete engineering examples ... there are many more!
• Is convex problem. Semi-infinite constraints render it NP-hard.
• Computationally tractable approximations: Relaxations
Polytope δ and affine dependence on parameter δ ...
... easy for scalar constraints
... tough for matrix constraints !
25/48
Carsten Scherer
Linear Fractional Representation
Suppose F (δ) is rational in δ = (δ 1 , . . . , δ k ) without pole at zero.
Can construct A, B, C, D such that
"
#
h
i ∆(δ) (I − A∆(δ))−1 B
F (δ) = C D
I
with diagonal and affine
"
∆(δ) =
δ 1 I1 .
0
..
0
δ k Ik
#
.
State-space realization of multi-dimensional system.
26/48
Carsten Scherer
Example: Velocity in Bike Model
With v = v0 + δ consider
"
#
" #
0 1
1
ẋ =
x+
u
γ 0
0
h
i
y =
α(v0 + δ) β(v0 + δ)2 x
27/48
Carsten Scherer
Example: Velocity in Bike Model
With v = v0 + δ consider
"
#
" #
0 1
1
ẋ =
x+
u
γ 0
0
h
i
y =
α(v0 + δ) β(v0 + δ)2 x
∆(δ)
w
z
y
Has Linear Fractional Representation
"
#
" #

0 1
1



ẋ =
x+
u


γ 0
0



h
i
h
i 
2
y =
v0 α v0 β x + α 2v0 α w


"
#
"
#




1 0
0 1


z =
x+
w

0 β/α
0 0
Nominal
System
"
w=
δ 0
0 δ
u
#
z
27/48
Carsten Scherer
Alternative Description of Uncertain LMI
Construct linear fractional representation



C0
 .. 

−1
 .  ∆(δ)(I − A∆(δ)) B + 
Cn
and define affine
"
W (x) =
C0 +
0
Pn
j=1
xj Cj



D0
F0 (δ)
..  1  .. 
. =  . 
2
Dn
Fn (δ)
#
P
C0T + nj=1 xj CjT
P
(D0 + D0T ) + nj=1 xj (Dj + DjT )
Alternative description of LMI:
" #T
"
#
n
X
∆(δ)(I − A∆(δ))−1 B
∗
≺0
F0 (δ) +
xj Fj (δ) =
W (x)
∗
I
j=1
28/48
Carsten Scherer
Infimize cT x such that for all δ ∈ δ:
"
"
#T
#
∆(δ)(I −A∆(δ))−1 B
∆(δ)(I −A∆(δ))−1 B
W (x)
≺ 0.
I
I
ROB
Robust LMI Problem
• Huge range of applications in robust optimization and control:
Uncertain LP’s, robust performance, Lyapunov design, ...
• Captures affine dependence on parameters: A = 0.
• Only few problem classes computationally tractable.
Goals: Compute upper bounds by efficiently solvable relaxation.
Numerically check quality for specific problem instance.
29/48
Carsten Scherer
How to Construct Relaxations?
Choose linear mappings G(y) and H(y) such that
"
#T
"
#
∆(δ)
∆(δ)
G(y) 4 0 =⇒
H(y)
< 0 for all δ ∈ δ.
I
I
30/48
Carsten Scherer
How to Construct Relaxations?
Infimum of cT x such that there exists y with
"
#T
"
#
I 0
I 0
H(y)
+ W (x) ≺ 0
G(y) ≺ 0,
A B
A B
REL
Choose linear mappings G(y) and H(y) such that
"
#T
"
#
∆(δ)
∆(δ)
G(y) 4 0 =⇒
H(y)
< 0 for all δ ∈ δ.
I
I
is upper bound on value of ROB.
Separation by dissipation ... for LMI-parametrized multipliers!
30/48
Carsten Scherer
Proof
Feasibility of REL
"
G(y) ≺ 0,
I 0
A B
#T
"
H(y)
I 0
A B
#
+ W (x) ≺ 0
31/48
Carsten Scherer
Proof
Feasibility of REL
"
G(y) ≺ 0,
I 0
A B
#T
"
H(y)
I 0
A B
#
+ W (x) ≺ 0
implies feasibility of ROB since
"
#T
"
#
∆(δ) (I − A∆(δ))−1 B
∆(δ) (I − A∆(δ))−1 B
W (x)
+
I
I
"
#T
"
#
∆(δ)
∆(δ)
+[(I − A∆(δ))−1 B]T
H(y)
(I − A∆(δ))−1 B ≺ 0.
I
I
|
{z
}
<0
31/48
Carsten Scherer
Example: Convex Hull Relaxation
Suppose parameter set is finitely generated: δ = co{δ 1 , . . . , δ N }.
"
∆(δ)
I
#T
"
H(y)
∆(δ)
I
#
< 0 for δ ∈ co{δ 1 , . . . , δ N }
⇑
32/48
Carsten Scherer
Example: Convex Hull Relaxation
Suppose parameter set is finitely generated: δ = co{δ 1 , . . . , δ N }.
"
∆(δ)
I
#T
"
H(y)
∆(δ)
I
#
< 0 for δ ∈ co{δ 1 , . . . , δ N }
⇑
#T "
#
" #T " #
"
∆(δ j )
∆(δ j )
I
I
< 0, j = 1, . . . , N
H(y) = y,
y
4 0,
y
I
I
0
0
|
{z
}
G(y) 4 0
32/48
Carsten Scherer
Example: Convex Hull Relaxation
Suppose parameter set is finitely generated: δ = co{δ 1 , . . . , δ N }.
"
∆(δ)
I
#T
"
H(y)
∆(δ)
I
#
< 0 for δ ∈ co{δ 1 , . . . , δ N }
⇑
#T "
#
" #T " #
"
∆(δ j )
∆(δ j )
I
I
< 0, j = 1, . . . , N
H(y) = y,
y
4 0,
y
I
I
0
0
|
{z
}
G(y) 4 0
Value of ROB = Value of REL in case of δ ⊂ R (one parameter).
32/48
Carsten Scherer
Outline
• Dissipativity and Linear Matrix Inequalities
• Linear parameter-varying system analysis and synthesis
• Robust LMI problems
• Matrix sum-of-squares relaxations
• Verification of relaxation exactness
• Examples
32/48
Carsten Scherer
Key Point
Uncertainty set δ = {δ ∈ Rk : g1 (δ) ≥ 0, . . . , gm (δ) ≥ 0}
33/48
Carsten Scherer
Key Point
Uncertainty set δ = {δ ∈ Rk : g1 (δ) ≥ 0, . . . , gm (δ) ≥ 0}.
δ region
Compact polytope ... gj (δ) affine
33/48
Carsten Scherer
Key Point
Uncertainty set δ = {δ ∈ Rk : g1 (δ) ≥ 0, . . . , gm (δ) ≥ 0}.
δ region
Compact polytope ... gj (δ) affine
δ region
Semi-algebraic ... gj (δ) polynomial
33/48
Carsten Scherer
Key Point
Uncertainty set δ = {δ ∈ Rk : g1 (δ) ≥ 0, . . . , gm (δ) ≥ 0}.
δ region
Compact polytope ... gj (δ) affine
δ region
Semi-algebraic ... gj (δ) polynomial
Positivity of matrix-valued polynomial for scalar constraints:
"
#T
"
#
∆(δ)
∆(δ)
H(y)
< 0 if g1 (δ) ≥ 0, . . . , gm (δ) ≥ 0.
I
I
33/48
Carsten Scherer
Matrix Sum-of-Squares (SOS)
The polynomial matrix S(δ) is SOS if there exists a rectangular
polynomial matrix T (δ) such that
S(δ) = T (δ)T T (δ).
• S(δ) is scalar-valued and tj (δ) denote elements of column T (δ):
X
S(δ) =
tj (δ)2 .
j
Our definition is generalization of standard scalar one.
• If S(δ) is SOS then S(δ) < 0 for all δ ∈ Rn .
Converse not true. (Hilbert 1888, Motzkin 1967)
34/48
Carsten Scherer
Checking SOS Property is Linear SDP
Fix scalar monomial basis u1 (δ), . . . , ut (δ) and search T1 , . . . , Tt in
T (δ) =
t
X
i=1
Ti ui (δ) =
t
X
Ti [ui (δ) ⊗ I].
i=1
Collect u1 (δ), . . . , ut (δ) into column vector u(δ).
35/48
Carsten Scherer
Checking SOS Property is Linear SDP
Fix scalar monomial basis u1 (δ), . . . , ut (δ) and search T1 , . . . , Tt in
T (δ) =
t
X
i=1
Ti ui (δ) =
t
X
Ti [ui (δ) ⊗ I].
i=1
Collect u1 (δ), . . . , ut (δ) into column vector u(δ).
S(δ) is SOS with respect to monomial basis u(δ) if and only if
S(δ) = [I ⊗ u(δ)]T W [I ⊗ u(δ)] for some W < 0.
• Linear affine and semi-definite constraints: Standard LMI problem.
• Full flexibility in choice of monomials. Can be made explicit.
35/48
Carsten Scherer
Constrained Positivity
It is trivial that
S(δ) < 0 for all δ ∈ Rk with g1 (δ) ≥ 0, . . . , gm (δ) ≥ 0
36/48
Carsten Scherer
Constrained Positivity
It is trivial that
S(δ) < 0 for all δ ∈ Rk with g1 (δ) ≥ 0, . . . , gm (δ) ≥ 0
⇑
S(δ) −
m
X
Sj (δ)gj (δ) is SOS
for
Sj (δ) which are SOS
j=1
36/48
Carsten Scherer
Constrained Positivity
It is trivial that
S(δ) < 0 for all δ ∈ Rk with g1 (δ) ≥ 0, . . . , gm (δ) ≥ 0
⇑
S(δ) −
m
X
Sj (δ)gj (δ) is SOS
for
Sj (δ) which are SOS
j=1
Latter is LMI problem for SOS matrices with respect to finite basis.
36/48
Carsten Scherer
Constraint Qualification Implies Equivalence
Suppose g1 (δ), . . . , gm (δ) satisfy constraint qualification. Then
S(δ) 0 for all δ ∈ Rk with g1 (δ) ≥ 0, . . . , gm (δ) ≥ 0
⇓
S(δ) −
m
X
Sj (δ)gj (δ) is SOS for Sj (δ) which are SOS
j=1
• Extends scalar results of Schmüdgen (1991) and Putinar (1993)
Scherer and Hol (2004)
• Complementary version: S(δ) scalar-valued and gj (δ) matrix-valued
Kojima (2003), Hol and Scherer (2003,2004)
37/48
Carsten Scherer
Constraint Qualification
2
r − kδk −
m
X
sj (δ)gj (δ) is SOS for some r > 0 and SOS sj (δ).
j=1
38/48
Carsten Scherer
Constraint Qualification
2
r − kδk −
m
X
sj (δ)gj (δ) is SOS for some r > 0 and SOS sj (δ).
j=1
If affine g1 (δ), . . . , gm (δ) define a compact polytope δ then the
SOS polynomials s1 (δ),..., sm (δ) can be taken of at most degree 2.
• Obtained by Jacobi & Prestel (2001) using deep results from real
algebraic geometry without degree information.
• Our proof: Elementary but tricky application of semi-definite duality.
Scherer and Hol (2004)
38/48
Carsten Scherer
New Matrix SOS Relaxation
"
∆(δ)
I
#T
"
H(y)
∆(δ)
I
#
< 0 if g1 (δ) ≥ 0, . . . , gm (δ) ≥ 0
⇑
"
∆(δ)
I
#T
H(y)
"
∆(δ)
I
#
−
m
X
yj (δ)gj (δ) is SOS ...
j=1
... for SOS polynomial matrices y1 (δ), . . . , ym (δ)
Search for SOS matrices w.r.t. finite basis: Fits in our framework!
Value of REL converges to value of ROB: Exhaustive basis sequence!
39/48
Carsten Scherer
Infimize cT x such that for all δ ∈ δ:
"
"
#T
#
−1
−1
∆(δ)(I −A∆(δ)) B
∆(δ)(I −A∆(δ)) B
≺ 0.
W (x)
I
I
ROB
Infimize cT x such that there exists y with
"
#T
"
#
I 0
I 0
G(y) ≺ 0,
H(y)
+ W (x) ≺ 0.
A B
A B
REL
Summary
• Can construct relaxations with suitable G(.) and H(.) such that
value of REL is an upper bound of value of ROB.
• Can systematically construct sequence to guarantee convergence.
40/48
Carsten Scherer
Discussion
• Undeniable Problems
Growth in computational complexity to reduce gap
Unavoidable for worst-case scenarios
• Practical Observations
Good approximation with small-sized relaxations ...
... for concrete problem instances in control
How can we determine approximation quality? Detect exactness?
41/48
Carsten Scherer
Outline
• Dissipativity and Linear Matrix Inequalities
• Linear parameter-varying system analysis and synthesis
• Robust LMI problems
• Matrix sum-of-squares relaxations
• Verification of relaxation exactness
• Examples
41/48
Carsten Scherer
Infimize cT x such that there exists y with
"
#T
"
#
I 0
I 0
G(y) ≺ 0,
H(y)
+ W (x) ≺ 0.
A B
A B
REL
Dualization
42/48
Carsten Scherer
Infimize cT x such that there exists y with
"
#T
"
#
I 0
I 0
G(y) ≺ 0,
H(y)
+ W (x) ≺ 0.
A B
A B
REL
Maximize hW0 , M i such that there exist M < 0, N < 0 with
hWj , M i + cj = 0, j = 1, . . . , n, and
"
# "
#T 
I 0
I 0
 + G∗ (N ) = 0.
H∗ 
M
A B
A B
DUAL
Dualization
G∗ (.) and H ∗ (.) are usual Hilbert adjoints.
42/48
Carsten Scherer
Characterizing Exactness
The following statements are equivalent:
• Exists a worst δ0 ∈ δ and the relaxation is exact.
• Exists an optimal solution M of DUAL with
"
#
i I 0
h
M = 0 for some δ ∈ δ.
I −∆(δ)
A B
43/48
Carsten Scherer
Characterizing Exactness
The following statements are equivalent:
• Exists a worst δ0 ∈ δ and the relaxation is exact.
• Exists an optimal solution M of DUAL with
"
#
i I 0
h
M = 0 for some δ ∈ δ.
I −∆(δ)
A B
If δ has LMI description ...
• Solve REL and DUAL to compute optimal value and optimal M .
• Try to determine solution δ ∈ δ (LMI problem).
• If solvable then the values of ROB and REL are equal.
43/48
Carsten Scherer
Outline
• Dissipativity and Linear Matrix Inequalities
• Linear parameter-varying system analysis and synthesis
• Robust LMI problems
• Matrix sum-of-squares relaxations
• Verification of relaxation exactness
• Examples
43/48
Carsten Scherer
Illustrative Academic Example
δ unit box ... choose g1,2 (δ) = ±δ1 + 1 and g3,4 (δ) = ±δ2 + 1.
3.5
3
Degrees of SOS factor:
2.5
y1 (δ1 )
y2 (δ1 )
y3 (δ2 )
y4 (δ2 )
2
1.5
1
0
0
1
1
1
1
1
1
0.5
0
0.5
0.6
0.7
0.8
0.9
1
44/48
Carsten Scherer
High-Performance Aircraft System
nz
®
q
u
u: Control input
α: Measurable parameter
nz : Tracked output
Nonlinear system description with aerodynamic effects:
α̇ = KM an α2 +bn α+cn (2−M /3) α+dn u +q
q̇ = M 2 am α2 +bm α−cm (7−8M /3) α+dm u
nz = M 2 an α2 +bn α+cn (2−M /3) α+dn u
45/48
Carsten Scherer
Main Idea
Rewrite as linear parameter-varying system
α̇ = Kδ1 an δ2 2 +bn δ2 +cn (2−δ1 /3) α+dn u +q
q̇ = δ1 2 am δ2 2 +bm δ2 −cm (7−8δ1 /3) α+dm u
nz = δ1 2 an δ2 2 +bn δ2 +cn (2−δ1 /3) α+dn u
with bounds 2 ≤ δ1 (t) ≤ 4 and −20 ≤ δ2 (t) ≤ 20.
Design good controller scheduled with δ1 (t), δ2 (t)
→ Is good controller for nonlinear system
46/48
Carsten Scherer
Application to Aircraft Model
Synthesis with Convex Hull Relaxation
M
decreases
5 seconds
M(t)
(t) decreases
in 5inseconds
from 4 from
to 2. 4 to 2.
40
Normal
Reference
Response
Reference
Response
30
Normal acceleration
Normal
acceleration
acceleration
20
10
0
−10
−20
0
1
2
3
4
5
Time
47/48
42/
Carsten Scherer
Conclusions
• Dissipativity ... Linear Matrix Inequalities in Control
• Range of applications: stability, performance, robustness
• Linear Parameter Varying systems ... robust LMI’s
• Novel matrix SOS relaxations with convergence
• Numerical verification of exactness
How to exploit system theoretic structure ...
... to improve reliability of LMI solvers ...
... to reduce computational complexity
?
48/48
Carsten Scherer