Sunteți pe pagina 1din 9

On the Computation of Robust Control

Invariant Sets for Piecewise Affine Systems

T. Alamo1 , M. Fiacchini1 , A. Cepeda1 , D. Limon1 , J.M. Bravo2,


and E.F. Camacho1
1
Departamento de Ingenierı́a de Sistemas y Automática, Universidad de Sevilla,
Sevilla, Spain
{alamo, mirko, cepeda, limon, eduardo} @cartuja.us.es
2
Departamento de Ingenierı́a Electrónica, Sistemas Informáticos y Automática.
Universidad de Huelva, Huelva, Spain
caro@uhu.es

Summary. In this paper, an alternative approach to the computation of control in-


variant sets for piecewise affine systems is presented. Based on two approximation
operators, two algorithms that provide outer and inner approximations of the maximal
robust control invariant set are presented. These algorithms can be used to obtain a
robust control invariant set for the system. An illustrative example is presented.

1 Introduction
In the context of nonlinear MPC, the stable and admissible closed-loop behavior
is typically based on the addition of a terminal constraint and cost [1]. The
terminal constraint is chosen to be an admissible robust control invariant set of
the system. The size of this control invariant set determines, in many cases, the
feasibility region of the nonlinear MPC controller [2]. It is shown in [2] that the
domain of attraction of MPC controllers can be enlarged by means of a sequence
of controllable (not necessarily invariant) sets.
The stability analysis of piecewise affine systems (PWA systems) plays an
important role in the context of hybrid systems control. This is mainly due to
the fact that piecewise affine systems can model a broad class of hybrid systems
(see [3]). Therefore, it is of paramount relevance in the context of hybrid MPC
the computation of controllable sets for this class of nonlinear systems [4].
The estimation of the domain of attraction of piecewise affine systems has
been addressed by a number of authors. Quadratic (and piecewise quadratic)
Lyapunov functions for hybrid systems have been proposed in [5, 6, 7]. A poly-
hedral approach is presented in [8]. Piecewise affine Lyapunov functions are
considered in [9]. In [10] an algorithm to compute the maximal robust control
invariant set for a PWA system is presented. Moreover, sufficient conditions to
guarantee that the algorithm is finitely determined are also given.
It is well-known (see, for example, [8]) that the computation of the maxi-
mal robust control invariant set for a piecewise affine system requires such a
computational burden that it is difficult to obtain it in an exact manner.

R. Findeisen et al. (Eds.): Assessment and Future Directions, LNCIS 358, pp. 131–139, 2007.
springerlink.com 
c Springer-Verlag Berlin Heidelberg 2007
132 T. Alamo et al.

In this paper we propose an algorithm that circumvents the huge computa-


tional complexity associated to the obtainment of the maximal robust control
invariant set. Two new algorithms are proposed. The first one provides a convex
polyhedral outer bound of the maximal control invariant set for the piecewise
affine system. This outer estimation is used, by the second proposed algorithm,
to obtain a robust control invariant set for the system (not necessarily the max-
imal one). The algorithms are based on inner and outer approximations of a
given non-convex set.
The paper is organized as follows: Section 2 presents the problem statement.
Section 3 presents an algorithm that computes an outer bound of the maximal
robust control invariant set of the piecewise affine system. A procedure to obtain
a robust control invariant set is proposed in section 4. An illustrative example
is given in section 5. The paper draws to a close with a section of conclusions.

2 Problem Statement
Let us suppose that X is a bounded convex polyhedron. Suppose also that the
convex polyhedra Xi , i = 1, . . . , r, with disjoint interiors, form a partition of X.
r
That is, X = Xi .
i=1
We consider the following piecewise affine system:

x+ = f (x, u, w) = Ai x + Bi u + Ei w + qi if x ∈ Xi (1)

where x ∈ Rnx is the state vector; x+ denotes the successor state; u ∈ U =


{ u ∈ Rnu : u∞ ≤ umax } is the control input; w denotes a bounded additive
uncertainty: w ∈ W = { w ∈ Rnw : w∞ ≤  }.
In order to present the results of this paper it is important to refer to the
notion of the one step set [8].
Definition 1 (one step set). Given a region Ω, and system (1), the following
sets are defined:

Q(Ω) = { x ∈ X : there is u ∈ U such that f (x, u, w) ∈ Ω, ∀w ∈ W }

Qi (Ω) = { x ∈ Xi : there is u ∈ U such that Ai x+Bi u+Ei w+qi ∈ Ω, ∀w ∈ W}


The following well-known properties allow us to compute Q(Ω) for a piecewise
affine system [8]:

r
Property 1. Given a convex polyhedron Ω: Q(Ω) = Qi (Ω), where Qi (Ω),
i=1
i = 1, . . . , r are polyhedra.

s
Property 2. If Ω = Pj and P1 , P2 , . . . , Ps are convex polyhedra, then Q(Ω) =
j=1

r 
s
Qi (Pj ), where Qi (Pj ), i = 1, . . . , r, j = 1, . . . , s are convex polyhedra.
i=1 j=1
On the Computation of Robust Control Invariant Sets 133

Based on these definitions and properties, the maximal robust control invariant
set can be obtained by means of the following algorithm [8]:
Algorithm 1
(i) Set the initial region C0 equal to X.
(ii) Ck+1 = Q(Ck ).
(iii) If Ck+1 = Ck then Ck = C∞ . Stop. Else, set k = k + 1 and return to step
(ii).
Note that the evolution of any initial condition belonging to set Ck can be ro-
bustly maintained in X at least k sample times. Therefore, C∞ = lim Ck con-
k→∞
stitutes the set of initial condition for which the system is robustly controllable in
an admissible way. That is, C∞ is the maximal robust control invariant set.
Suppose that algorithm 1 converges to C∞ in kd steps. Then, applying prop-
erty 2 in a recursive way it is possible to state that Ckd = C∞ can be represented
by means of the union of rkd convex polyhedra. This worst-case estimation of
the number of convex polyhedra required to represent C∞ clearly shows that the
exact computation of C∞ for a piecewise-affine system might not be possible in
the general case: the number of sets required to represent the maximal robust
control invariant set grows in an exponential way with the number of iterations
of algorithm 1. Even in the case that algorithm 1 obtains (theoretically) the
maximal robust control invariant set in a finite number of steps, the complexity
of the representation might make it impossible to run the algorithm beyond a
reduced number of steps (normally insufficient to attain the greatest domain of
attraction). Therefore, it is compulsory to consider approximated approaches to
the computation of C∞ .
In this paper we propose an algorithm (based on convex outer (and inner)
approximations of the one step set) that can be used to compute a convex robust
control invariant set for the piecewise affine system.

3 Outer Bound of the Maximal Robust Control Invariant


Set
One of the objectives of this paper consists in providing a procedure to obtain
a convex outer approximation of C∞ for a piecewise affine system. This outer
bound has a number of practical and relevant applications:
(i) It captures the geometry of C∞ and makes the computation of a robust
control invariant set for the system easier (this use is explored in section
4). Moreover, it can be used as the initial set in algorithm 1. If the outer
bound is a good approximation, then algorithm 1 might require an (imple-
mentable) reduced number of iterations.
(ii) The constraints that define the outer bound can be included as hard con-
straints in a hybrid MPC scheme. Moreover, the inclusion of the aforemen-
tioned constraints can be used to improve the convex relaxations of the
nonlinear optimization problems that appear in the context of hybrid MPC.
134 T. Alamo et al.

(iii) The outer bound can be used as a measure of the controllable region of a
hybrid system. This can be used in the design of the hybrid system itself.
(iv) The obtained convex region can be also used to induce a control Lyapunov
function.
The following algorithm provides a convex polyhedron that serves as an outer
bound of the maximal robust control invariant set of a piecewise affine system:

Algorithm 2
(i) k = 0, Ĉ0 = X.

r >
(ii) Ĉk+1 = Co { Qi (Ĉk )} Ĉk
i=1
(iii) If Ĉk+1 = Ĉk , then Stop. Else, k=k+1, Go to step (ii).

Remark 1. Note that the convex hull operator (Co) required to implement the
algorithm can be substituted by any outer approximation of the convex hull. For
example, the envelope operator of [11], or the outer approximation provided in
[12]. The algorithm can be stopped when there is no significant improvement of
the outer bound. That is, when Ĉk is almost identical to Ĉk−1 . For example,
the algorithm could be stopped when (1 − s )Ĉk ⊆ Ĉk+1 , where s > 0 is a
arbitrarily small tuning parameter.

Property 3. Each one of the polyhedrons Ĉk obtained by means of algorithm 2


constitutes an outer bound of the maximal robust control invariant set of the
piecewise affine system. That is, C∞ ⊆ Ĉk , for all k ≥ 0.

Proof
It is clear that C∞ ⊆ X = Ĉ0 . To prove the property it suffices to show that
C∞ ⊆ Ĉk implies C∞ ⊆ Ĉk+1 , for every k ≥ 0. Suppose that C∞ ⊆ Ĉk :
 r 
) ) ) 1 )
C∞ = C∞ Ĉk = Q(C∞ ) Ĉk ⊆ Q(Ĉk ) Ĉk = Qi (Ĉk ) Ĉk ⊆ Ĉk+1
i=1

4 Inner Approximation of the Maximal Robust Control


Invariant Set
In this section, an algorithm that computes an inner approximation of C∞ is
presented. Such an algorithm is based on the complementary set of the one step
operator and on the notion of inner supporting constraint.

4.1 Complementary Set of Q(Ω)


The inner approximations of C∞ presented in this paper for Q(Ω) rely on the
computation of the complementary set of Q(Ω):
On the Computation of Robust Control Invariant Sets 135

Definition 2. Given set Ω, Qc (Ω) denotes the complementary set of Q(Ω) in


X. That is, Qc (Ω) = { x ∈ X : x ∈ Q(Ω) }. Given set Ω, Qci (Ω) denotes the
complementary set of Qi (Ω) in Xi . That is, Qci (Ω) = { x ∈ Xi : x ∈ Qi (Ω) }.

r
It is inferred from the previous definition that given set Ω: Qc (Ω) = Qci (Ω).
i=1
Therefore, in order to compute Qc (Ω), it suffices to compute Qci (Ω), i = 1, . . . , r.
The following property (see [13] for a proof) shows how to compute Qci (Ω).

Property 4. Suppose that Qi (Ω) = { x ∈ Xi : Gi x ≤ gi }, where Gi ∈ RLi ×nx


and gi ∈ RLi . Then,
1
Li
Qci (Ω) = Si,j (Gi , gi )
j=1

where Si,j (Gi , gi ) = { x ∈ Xi : Gi (j)x > gi (j), Gi (l)x ≤ gi (l) for l = 1, . . . ,


j − 1 }.

4.2 Inner Supporting Constraint


The construction of the proposed inner approximation of C∞ is based on the
notion of inner supporting constraint:
Definition 3. Suppose that S is a compact convex set that does not contain
the origin and that R is a bounded set. We say that { x : c x ≤ 1 } is
an inner supporting constraint of S over R if c is the solution of the following
minimization problem

min ρ
c,ρ

s.t. c x > 1, ∀x ∈ S
1
c x ≤ 1, ∀x ∈ R
ρ
The following property allows us to compute an inner supporting constraint of
r
a polyhedron S over Q(Ω) = Ti ( where Ti = Qi (Ω), i = 1, . . . , r) by means
i=1
of the solution of a linear optimization problem. The proof of the property is
similar to the proof of an analogous property in [12] and it is omitted because
of space limitations.
Property 5. Consider polyhedron S = { x : F x ≤ f } and the polyhedrons
Tl = { x : Ml x ≤ ml }, l = 1, . . . , r. Suppose that the scalar ρ and the vectors
c, λ, and βl , l = 1, . . . , r satisfy the following constraints:

ρ>0 (2)
λ≥0 (3)
βl ≥ 0, l = 1, . . . , r (4)
136 T. Alamo et al.

1 + f λ < 0 (5)
−ρ + m l βl ≤ 0, l = 1, . . . , r (6)
c + F λ = 0 (7)
c − Ml βl = 0, l = 1, . . . , r (8)

then

c x > 1, ∀x ∈ S (9)
1r
1
c x ≤ 1, ∀x ∈ Tl (10)
ρ
l=1

4.3 Robust Control Invariant Set: Proposed Algorithm


The following algorithm serves to compute a robust control invariant set for a
piecewise affine system:

Algorithm 3

(i) Set k = 0 and choose a contracting factor λ̃ ∈ (0, 1).


(ii) Make C̃0 equal to the outer approximation of C∞ obtained by means of
algorithm 2.
(iii) Given C̃k = { x : Hx ≤ h }, obtain Ti = Qi (λ̃C̃k ), i = 1, . . . , r.

r c
n
(iv) Obtain Qc (λ̃C̃k ) = Qci (λ̃C̃k ) = Sj by means of property 4.
i=1 j=1
(v) For every j = 1, . . . , nc obtain { x : c j x ≤ 1 }, the inner supporting

r
constraint of Sj over Q(λ̃C̃k ) = Ti . This can be achieved by means of
i=1
property 5.
>c
n
(vi) Make C̃k+1 = { x : c
j x ≤ 1 }.
j=1
(vii) If C̃k+1 ⊆ Q(C̃k+1 ) then C̃k+1 is a robust control invariant set. Stop. Else,
set k = k + 1 and go to step (iii).

Bearing in mind the λ contractive procedure of [14], a contracting factor


λ̃ ∈ (0, 1) has been included in the algorithm. Note that algorithm 3 finishes only
if C̃k+1 ⊆ Q(C̃k+1 ). In virtue of the geometrical condition of robust invariance
[14], it is inferred that C̃k is a robust control invariant set. That is, if algorithm 3
finishes then a robust control invariant set is obtained. This set serves as an inner
approximation of C∞ . Due to the approximate nature of the algorithm, it is not
guaranteed that algorithm 3 converges to a robust control invariant set. Note,
however, that it can be shown that each one of the obtained sets C̃k constitutes
an inner approximation of Ck . The proof of this statement is based on the fact
On the Computation of Robust Control Invariant Sets 137

>
that, by definition of inner supporting constraint, Sj { x : c
j x ≤ 1 } equals
the empty set. Thus,

) ) )nc
Qc (λ̃C̃k ) C̃k+1 = Qc (λ̃C̃k ) ( { x : c
m x ≤ 1 })
m=1
 
1
nc ) )nc 1
nc ) )nc
=( Sj ) ( { x : c
m x ≤ 1 }) = Sj (
{ x : cm x ≤ 1 })
j=1 m=1 j=1 m=1
1c
n )
⊆ (Sj { x : c
j x ≤ 1 }) = ∅.
j=1

>
That is, Qc (λ̃C̃k ) C̃k+1 = ∅. This implies that C̃k+1 ⊆ Q(λ̃C̃k ).

5 Numerical Example
In this example, region X = { x : x∞ ≤ 15 } is subdivided into the subregions
X1 , X2 and X3 . These subregions are defined as follows:

X1 = { x ∈ X : x1 − x2 ≤ 0 }
X2 = { x ∈ X : x1 − x2 > 0 and x1 + x2 ≥ 0 }
X3 = { x ∈ X : x1 − x2 > 0 and x1 + x2 < 0 }

Consider the following piecewise affine system:


⎧     

⎪ 11 0 1

⎪ x+ u+ w if x ∈ X1

⎪ 01


1 0





⎪     
⎨ 1 1 0 1
+
x = x+ u+ w if x ∈ X2

⎪ 0.5 1.5 1 0





⎪      



⎪ 1 −0.5 1 1

⎩ 0 1.5 x + 1.5 u + 0 w if x ∈ X3

In this example it is assumed that U = { u ∈ R : u∞ ≤ 2 } and


W = { w ∈ R : w∞ ≤ 0.1 }. The contracting factor for algorithm 3 has been
set equal to 0.95. In figure 1 the sequence of outer bounds Ĉk is displayed (the
outer approximation of the convex hull operator presented in [12] has been used).
The most inner polyhedron is used in algorithm 3 as initial guess to obtain a
robust control invariant set. In figure 2 a sequence of sets C̃k leading to a robust
control invariant set is displayed. The most inner polyhedron is a robust control
invariant set for the piecewise affine system.
138 T. Alamo et al.

15 15
X X
1 1

X2 X2
10 10

5 5

0 0

−5 −5

−10 −10

X3 X3
−15 −15
−15 −10 −5 0 5 10 15 −15 −10 −5 0 5 10 15

Fig. 1. Sequence of outer bounds Ĉk lead- Fig. 2. Sequence of sets C̃k leading to a
ing to an outer approximation of C∞ robust control invariant set

6 Conclusions
In this paper, an algorithm that provides an outer approximation of C∞ is
presented. This outer approximation has a number of practical and relevant ap-
plications. Based on the outer approximation of C∞ , an algorithm that provides
an inner approximation of C∞ is given. This algorithm can be used to obtain a
robust control invariant set for a piecewise affine system. An illustrative example
is presented.

References
[1] Mayne, D.Q., Rawlings, J.B., Rao, C.V. and Scokaert, P.O.M. ”Constrained model
predictive control: Stability and optimality”. Automatica, 36:789–814, (2000).
[2] Limon, D., Alamo, T. and Camacho, E.F. ”Enlarging the domain of attraction of
MPC controllers”. Automatica, 41(4):629–635, (2005).
[3] Heemels,W., De Schutter, B. and Bemporad, A., ”Equivalence of hybrid dynam-
ical models”. Automatica, 37:1085–1091, (2001).
[4] Lazar, M., Heemels, W.P.M.H., Weiland, S. and Bemporad, A. ”Stabilization
conditions for model predictive control of constrained PWA systems”. Proceedings
of the 43rd IEEE Conference on Decision and Control. 4595-4600, (2004).
[5] Johansson, M. and Rantzer, A., ”Computation of piecewise quadratic lya-
punov functions for hybrid systems”. IEEE Transactions on Automatic Control,
34(4):555–559, (1998).
[6] Mignone, D., Ferrari-Trecate, G., and Morari, M. ”Stability and Stabilization of
Piecewise Affine and Hybrid Systems: An LMI Approach. Proceedings of the 39th
IEEE Conference on Decision and Control. 504-509, (2000).
[7] Rodrigues, Luis. ”Stability analysis of piecewise-affine systems using controlled
invariant sets”. Systems and Control Letters, 53(2):157-169, (2004).
[8] Kerrigan, E.C. ” Robust Constraint Satisfaction: Invariant Sets and Predictive
Control”. PhD thesis, University of Cambridge, (2000).
On the Computation of Robust Control Invariant Sets 139

[9] Grieder, P., Kvanisca, M., Baotic, M. and Morari, M. ”Stabilizing low complexity
feedback control of constrained piecewise affine systems”. Automatica, 41:1683–
1694, (2005).
[10] Rakovic, S.V., Grieder, P., Kvasnica, M., Mayne, D.Q. and Morari, M. ”Com-
putation of invariant sets for piecewise affine discrete time systems subject to
bounded disturbances”. Proceedings of the 43rd IEEE Conference on Decision
and Control. 1418-1423, (2004).
[11] Bemporad, A., Fukuda, K. and Torrisi, F. ”Convexity recognition of the union of
polyhedra”. Computational Geometry, 18, 141-154, (2001).
[12] Alamo, T., Cepeda, A., Limon, D., Bravo, J.M., Fiacchini, M. and Camacho,
E.F. ”On the computation of robust control invariant sets for piecewise affine sys-
tems” Workshop on Assessment and Future Directions on NMPC. Freudentadt-
Lauterbad, Germany, August 26-30(2005).
[13] Bemporad, A., Morari, M., Dua, V. and Pistikopoulos, E.N. ”The Explicit Linear
Quadratic Regulator for Constrained Systems”. Automatica, 38(1), 3-20, (2002).
[14] Blanchini F., ”Set invariance in control”. Automatica, 35:1747–1767, (1999).

S-ar putea să vă placă și