Вы находитесь на странице: 1из 38

Template complex zonotopes for computing

invariants of affine hybrid systems

Arvind Adimoolam and Thao Dang


VERIMAG, Grenoble, France

GT-VERIF 2017

1 / 22
Invariant

I All trajectories starting inside the invariant remain inside


the invariant.

2 / 22
System description: General model

I Specified as hQ, A, Ui, Q is a discrete set of locations,


A : Q Mnn (R) maps to a finite set of matrices,
U Rn is a bounded input set.
I State: Pair (q, x) Q Rn .
I Trajectory: (q, x) : Z0 Q Rn if
x(k + 1) = Aq(k) x(k) + u : u U
3 / 22
Invariant computation problem

Let T Mkn (R), d Rk and I Q Rn . Then, compute


S Q Rn such that q, q 0 Q,

Aq Sq U Sq0 (Invariance)
x Sq , Tx d (Safety)
I S (Containment of initial set)

4 / 22
Features of a good set representation and
computation

1 Eigenstructure based contraction.


2 Closure under system operations.
3 Avoid wrapping effect.
4 Reasonable complexity of invariant computation.

5 / 22
1) Eigenstructure based contraction
The representation should capture contraction along stable
eigenvectors of linear sub-systems.

6 / 22
1) Eigenstructure based contraction
The representation should capture contraction along stable
eigenvectors of linear sub-systems.
Example: Quadratic Lyapunov or Ellipsoids: Consider sub-level
set of n
X
eiT x e Ti x ,
 
f (x) =
i=1

where ei is eigenvector and i is eigenvalue of matrix A.

6 / 22
1) Eigenstructure based contraction

The representation should capture contraction along stable


eigenvectors.
Counterexample: Polytopes and usual zonotopes can not
utilize complex eigenstructure.

7 / 22
2) Closure under system operations

Basic system operations should preserve set representation for


good approximation quality.
I Example: Zonotopes, Convex polytopes etc are closed
under linear transformation and Minkowski sum.

8 / 22
2) Closure under system operations

Basic system operations should preserve set representation for


good approximation quality.
I Example: Zonotopes, Convex polytopes etc are closed
under linear transformation and Minkowski sum.
I Counterexample: Ellipsoids are not closed under
Minkowski sum.

8 / 22
3) Avoid wrapping effect

Wrapping effect: Overapproximation error accumulates in


step-by-step computation.

I Step-by-step computation based on Finite set of Support


Vectors or Template polyhedra can suffer from
wrapping effect.

9 / 22
3) Avoid wrapping effect

Wrapping effect: Overapproximation error accumulates in


step-by-step computation.

I Step-by-step computation based on Finite set of Support


Vectors or Template polyhedra can suffer from
wrapping effect.
I Avoids wrapping effect: Quadratic lyapunov using LMIs
and BMIs.

9 / 22
4) Reasonable complexity of computation

The complexity of computation should not be too high.

10 / 22
4) Reasonable complexity of computation

The complexity of computation should not be too high.


I Counterexample: Iterative convex hull approximation.

Size increases in
each step propor-
tional to dimen-
sion.

10 / 22
Analysis of previous approaches

Quadratic Template Convex


Property lyapunov polyhedra or polyhedra
(Ellipsoid) Support vectors
Eigenstructure
based X 7 7
contraction
Minkowski sum
7 X X
closure
Avoid wrapping
X 7 X
effect
Complexity of
reasonable reasonable high
computation

11 / 22
Template complex zonotopes

Extension of usual zonotopes by


1 Complex valued generators and complex combining
coefficients
2 Variable bounds on absolute values of combining coefficients.

Definition
Let V Mnm (C) (template) and s Rm 0 (scaling factors)
n
and c R (center). Then the following is a template
complex zonotope:
C (V , c, s) = {V  :  Cm , |i | si i {1, ..., m}} .

12 / 22
Geometric visualization of template complex
zonotopes
Minkowski sums of line segments and some ellipsoids, hence can
have non-polyhedral real projections.

13 / 22
Properties of TCZ: contraction based on
eigenstructure

AC ([e1 , ..., en ], 0, s) = C ([e1 , ..., en ], 0, D (|i |) s)

14 / 22
Properties of TCZ: Closure under system
operations

I Linear transformation: AC (V , c, s) = C (AV , Ac, s).

15 / 22
Properties of TCZ: Closure under system
operations

I Linear transformation: AC (V , c, s) = C (AV , Ac, s).


I Minkowski sum:   
0 0 0
 0
 0 s
C (V , c, s) C (V , c , s ) = C V V , c + c , .
s0

15 / 22
Properties of TCZ: Closure under system
operations

I Linear transformation: AC (V , c, s) = C (AV , Ac, s).


I Minkowski sum:   
0 0 0
 0
 0 s
C (V , c, s) C (V , c , s ) = C V V , c + c , .
s0
I Projection: Let z Rn . Then
maxxC(V ,c,s) z T x = z T c + z T V s.

15 / 22
Inclusion checking: A convex condition

A relation v defined as follows is a sufficient condition for


inclusion.
I C V0 0 0

nm0 , c , s v C (Vnm , c, s) if

X Mmm0 (C) and y Cm s.t.


VX = V 0 D (s 0 ) , Vy = c 0 c, and
m0
!
m X
max |yi | + |Xij | si 0
i=1
j=1

I Above is equivalent to a set of second order conic


constraints on the center and scaling factors.

16 / 22
Convex condition for computing invariant
Let additive input set be overapproximated by C V in , c in , s in and


initial set by C (G , 0, h). A set of states S where for each q Q,


Sq is defined as C (V , cq , sq ) is a safe invariant if q, q 0 Q,

/ inclusion after affine transformation/


  
 in
 in sq
C Aq V V , cq + c , v C (V , cq0 , sq0 ) (1)
s in
/ inclusion of initial set/
C (G , 0, h) v C (V , cq , sq ) (2)
/ satisfaction of linear constraints (safety)/
Tcq + |TV | sq d (3)

17 / 22
Template selection

18 / 22
Template selection

1 Collect all eigenvectors all matrices Aq , q Q and their


products. This captures contractive directions.

18 / 22
Template selection

1 Collect all eigenvectors all matrices Aq , q Q and their


products. This captures contractive directions.
2 Append input set overapproximation template W and its
transformations the linear matrices Aq . This helps in
good quality overapproximation of Minkowski sum with
input set after template based abstraction.

18 / 22
Template selection

1 Collect all eigenvectors all matrices Aq , q Q and their


products. This captures contractive directions.
2 Append input set overapproximation template W and its
transformations the linear matrices Aq . This helps in
good quality overapproximation of Minkowski sum with
input set after template based abstraction.
3 Any arbitrary set of vectors can be added. Increases
accuracy of verification, but at a computational expense.

18 / 22
Experiment: Networked vehicle platoon

Benchmark 2014
ARCH Work-
shop [3]

Fast switching:
Integer dwell times
Method
Comp.
e1 e2 e3
time (s)
octagon
UB UB UB NT
template
SpaceEx
100 support
UB UB UB NT
vectors
TCZ invariant 46 54 57 12.6

UB: >1000, NT: Not terminating in more than 180s,

n/a: Not applicable/not available, TCZ: Tempalte complex zonotope.


19 / 22
Experiment: Networked control system

Stability verification problem. Example from [Wittenmark,


Astrom et.al.]

Reference tmin tmax


Value recommended in [4] 0.08 0.22
NCS toolbox [2] 0.08 0.4
Template complex zonotope 0.08 0.53

20 / 22
Conclusion: Advantages of Template Complex
Zonotopes

1 Ability to utilize complex eigenstructure.

21 / 22
Conclusion: Advantages of Template Complex
Zonotopes

1 Ability to utilize complex eigenstructure.


2 Closure under system operations.

21 / 22
Conclusion: Advantages of Template Complex
Zonotopes

1 Ability to utilize complex eigenstructure.


2 Closure under system operations.
3 Circumvent wrapping effect (i.e., compute invariant in
single step by solving convex constraints).

21 / 22
Conclusion: Advantages of Template Complex
Zonotopes

1 Ability to utilize complex eigenstructure.


2 Closure under system operations.
3 Circumvent wrapping effect (i.e., compute invariant in
single step by solving convex constraints).
4 Complexity of computation is reasonable (polynomial
number of SOCP constraints to solve).

21 / 22
Future work

I Next problem: Extending this method to handle general


affine hybrid systems, where switching is controlled by
linear constraints.

22 / 22
Future work

I Next problem: Extending this method to handle general


affine hybrid systems, where switching is controlled by
linear constraints.
I Upcoming research: Augmented complex zonotopes to
handle controlled switching in affine systems.

22 / 22
Future work

I Next problem: Extending this method to handle general


affine hybrid systems, where switching is controlled by
linear constraints.
I Upcoming research: Augmented complex zonotopes to
handle controlled switching in affine systems.
I Challenge: Extending these ideas to non-linear hybrid
systems. How to use the feature of TCZ to incorporate
complex eigenstructure for Non-Linear hybrid systems?

22 / 22
Arvind Adimoolam and Thao Dang.
Using complex zonotopes for stability verification.
In American Control Conference.
sites.google.com/site/cztopepubs/, 2016.
N.W. Bauer, S.J.L.M. van Loon, M.C.F. Donkers,
N van de Wouw, and W.P.M.H. Heemels.
Networked control systems toolbox: Robust stability
analysis made easy.
In IFAC Workshop on Distributed Estimation and Control
in Networked Systems (NECSYS), pages 5560, 2012.
Ibtissem Ben Makhlouf and Stefan Kowalewski.
Networked cooperative platoon of vehicles for testing
methods and verification tools.
In ARCH@ CPSWeek, pages 3742, 2014.

22 / 22
Bjorn Wittenmark, Karl Johan Astrom, and Karl-Erik
Arzen.
Computer control: An overview.
IFAC Professional Brief, 1 2002.

22 / 22

Вам также может понравиться