2010 American Control Conference WeB16.

2
Marriott Waterfront, Baltimore, MD, USA
June 30-July 02, 2010

An optimal control L2-gain disturbance rejection

design for networked control systems
Pablo Millán, Luis Orihuela, Carlos Vivas and Francisco R. Rubio

Abstract— This paper proposes an optimal control solution of the initial value of the functional (V (x0 ,t0 )), so it depends
for the L2 -gain disturbance rejection problem for networked on the initial conditions. To remove this dependence, they
control systems. The problem, usually referred as mixed H2 /H∞ , propose some conservative bounds of the guaranteed cost,
aims at designing a linear stabilizing controller minimizing
a given cost function subject to a L2 -gain disturbance rejec- formulating an optimization problem to minimize the cost.
tion constraint. The problem is formulated assuming network The method proposed in this work deals with linear
induced time-varying delays and/or packet losses. An LMI- NCSs subject to L2 bounded disturbances, where signals are
based minimization problem is derived based on a Lyapunov- affected by time-varying delays and packet dropouts. No a
Krasovskii approach, considering a polytopic covering of the
time-delay range. Simulation results are provided to verify the priori knowledge of the statistical distribution of the delay
performance of the methodology. is assumed, though upper and lower bounds of the delay
interval are required. Additionally, it is assumed that the
I. INTRODUCTION maximum number of consecutive packet losses is bounded.
This work addresses the design of an optimal L2 -gain Thus, given a cost function J and a controlled output z(t),
disturbance rejection control for linear systems controlled the proposed problem can be informally stated as designing
through a communication network. These control schemes, a linear controller such that:
usually termed as NCS (Networked Control Systems), have
captured the attention of many researches in last years. There • Stabilizes the unperturbed system (ω (t) ≡ 0), as the cost
are a huge range of potential applications of these techno- function J is minimized.
logies and, also, challenging control problems have arisen • The controlled output satisfies kz(t)k2 ≤ γ kω (t)k2 for
from the presence of unreliable communication channels in any nonzero disturbance ω (t).
the control loops, see [1]. This problem is usually referred in the literature as the
Many authors have studied the conditions which must mixed H2 /H∞ control problem (see, for instance, [13] for
be verified to stabilize the whole system using different time-varying delay systems), as the H2 part accounts for
controllers and under specific conditions of the network. To the optimization of a performance index with an L2 -gain
name some of them, we can find works which study the disturbance rejection constraint in the H∞ component.
stability of NCSs with delays, [2]; packet losses, [3]; or with
To find a suitable controller, the Lyapunov-Krasovskii
limited bandwidth, [4].
approach is used, yielding an optimization problem involving
In this framework, some remarkable results have been
Nonlinear Matrix Inequalities (NLMI). Then, a well-known
obtained based on the Lyapunov-Krasovskii approach, [5].
procedure is employed to reformulate the optimization pro-
Using an appropriate Lyapunov-Krasovskii functional, it is
blem in terms of Linear Matrix Inequalities (LMIs). In
possible to take into account delays and packet dropouts
order to reduce the conservatism of the stability criterion,
which affect the communication. This technique has been
a polytopic covering of the time-delay range is employed.
widely used to study the stability of time-delay systems,
This allows us to retain the maximum and minimum value
see for instance [6], [7] and references therein. It is from
of the delay to very last formulation of the NLMI problem,
the work [8], and the input delay approach to sampled-data
thus avoiding unnecessary sources of conservatism.
system, when the contributions of time-delay systems were
widely applied to networked control system framework, see Finally, we present simulation results for a vehicle tracking
[9]. problem in which the above mentioned characteristics are
Several works are available in the Literature in the context featured. We will show that the presented method offers a
of optimal control for NCS. In [10], some modifications of satisfactory solution for the proposed problem.
Kalman filter are introduced to design an optimal controller The paper is organized as follows. Section II is devoted
dealing with packet dropouts. However, few works resort to to the description of the NCS model to be considered.
the powerful Lyapunov-Krasovskii approach, see [11], [12]. Section III is concerned with the main results. In Section
Both works consider time delays in their formulations finding III-A the optimal L2 -gain control problem is formulated. A
a guaranteed cost control. The guaranteed cost is a function particular solution for this problem is proposed in Section
III-B. Finally, Section IV applies the previous result to an
P. Millán, L. Orihuela, C. Vivas and F.R. Rubio are with Dpto. de scenario of vehicle tracking. Conclusions are summarized
Ingenierı́a de Sistemas y Automática. Escuela Superior de Ingenieros.
Universidad de Sevilla, Spain {pmillan, orihuela, vivas, in Section V.
rubio}@cartuja.us.es

978-1-4244-7425-7/10/$26.00 ©2010 AACC 1344

II. NCS MODEL AND PRELIMINARIES III. MAIN RESULT
Consider the following LTI system given by: A. Problem Formulation
ẋ(t) = Ax(t) + Bu(t) + Bω ω (t), (1) In this section the optimal control problem with L2 -gain
z(t) = Cx(t) + Du(t), (2) disturbance rejection for Networked Control System is
x(t) = φ (t), t ∈ [t0 − τM ,t0 ], (3) formulated. Next, a general solution is proposed. As we will
prove, the design of a suitable controller for the proposed
where x(t) ∈ Rn , u(t) ∈ Rm and z(t) ∈ Rq are the state
problem can be stated as an optimization problem.
vector, control input vector and controlled output, respec-
tively; ω (t) ∈ L2 [t0 , ∞) denotes the external disturbances; A,
Definition 2. The Optimal L2 -gain Control Problem
B, Bω , C and D are some constant matrices of appropriate
Consider that the LTI system described by (5)-(7) is
dimensions and φ (t) denotes the initial conditions.
controlled over a communication network, which satisfies
Consider system (1)-(3) being controlled through a net-
(9)-(10). Given:
work. The inclusion of such a network in the control loop
• A desired level of disturbance attenuation γ .
induces time-varying delays, and possibly packet dropouts. R∞
• A quadratic cost function J = t [xT (t)Qx(t) +
Assume that the sensor samples measurements from the plant 0

in a clock-driven manner, at time instants t = jk h, with h uT (t)Ru(t)]dt, with Q, R > 0,

being the sampling time, and jk (k = 1, 2, 3, ..) integers such The Optimal L2 -gain problem consists in finding an sta-
that { j1 , j2 , j3 , ...} ⊆ {1, 2, 3, ...} and jk < jk+1 . With these bilizing linear controller K such that:
assumptions packet losses are contemplated, while out-of- 1) The closed-loop system is asymptotically stable with
order packets are rejected by the controller or the actuator. ω (t) ≡ 0,
Let us define t ∈ [tk ,tk+1 ) as the time intervals where the 2) The controller minimizes the cost function J with
control input applied to the system is constant, where tk are ω (t) ≡ 0,
the time instants when the control signal, corresponding to 3) Under the assumption of zero initial condition, the
the plant state at t = jk h, reaches the plant. controlled output z(t) satisfies kz(t)k2 ≤ γ kω (t)k2 for
Therefore, the control input can be written as: any nonzero ω (t) ∈ L2 [0, ∞).
u(t) = Kx(tk − τsc (k) − τca (k)), t ∈ [tk ,tk+1 ), (4)
The following assumptions will be additionally needed to
where τsc (k) and τca (k) are the network induced delays of
provide a solution to the problem.
the data corresponding to the measured plant state at t = jk h, Assumption 4. Given a continuous quadratic Lyapunov-
from sensor to controller and from controller to actuator, Krasovskii functional (LKF) V (t), it is assumed that the
respectively. The round-trip delay τsa (k) can also be defined derivative for t ∈ [tk ,tk+1 ) can be written in the following
as τsa (k) = τsc (k) + τca (k). way:
Thus the system (1)-(3) under the control law (4) can be V̇ (t) ≤ ξ T (t)Ξ(K)ξ (t) − z(t)T z(t) + γ T ω T (t)ω (t), (11)
rewritten as:
ẋ(t) = Ax(t)+ BKx(t − τ (t))+ Bω ω (t), (5)
where ξ (t) ∈ Rnξ is an augmented state vector and
Ξ(K) ∈ Rnξ ×nξ is a symmetric matrix which depends,
z(t) = Cx(t) + DKx(t − τ (t)), ∀t ∈ [tk ,tk+1 ), (6)
among others, on the controller matrix.
x(t) = φ (t), t ∈ [t0 − τM ,t0 ], (7)
where τ (t) = t − tk + τsc (k) + τca (k) and τM is the upper Assumption 5. The cost function J can be written as:
Z ∞
bound of τ (t) (see Definition 1). J= [ξ T (t)Φ(K)ξ (t)]dt, (12)
The following assumptions, which are fairly common in t0
the NCS framework, are also made in this paper. where Φ(K) is a positive definite matrix of appropriate
Assumption 1. The sensor is clock-driven. The controller dimensions which possibly depends on K.
and actuator are event-driven. The actuator is a zero-order The following lemma is required for further developments
holder. in this work
Assumption 2. Two constants τ sa , τ sa ≥ 0, exist such that Lemma 1. Suppose that Assumptions 4-5 hold. Then, the
the following inequality holds: Optimal L2 -gain control problem can be solved by finding a
controller matrix K such that:
τ sa ≤ τsa (k) ≤ τ sa , ∀k ∈ N. (8)
min α, (13)
K
Assumption 3. The maximum number of consecutive data
subject to α Ξ(K) < −Φ(K) (14)
dropouts from sensor to actuator is bounded by n p ∈ N.
α > 0, α ∈ R (15)
Furthermore, this definition will be needed in the following
sections. Proof. In the proof of this lemma, we will demonstrate that
Definition 1. Regarding to Assumptions 2 and 3, it is a controller which solves the optimization problem (13)-(15)
possible to define two constants τm ≥ 0 and τM > τm such
that: also satisfies all the issues of Definition 2.
1) For ω (t) ≡ 0, considering (11) for t ∈ [tk ,tk+1 ) the
τ (t) ≥ τ sa = τm , (9) following holds:
τ (t) ≤ (1 + n p )h + τ sa = τM . (10) V̇ (t) ≤ ξ T Ξ(K)ξ (t) − z(t)T z(t). (16)

1345
From (16) and regarding to the assumption 5 and the dimensions solve the following optimization problem for the
equations (14)-(15), one can obtain that V (t) decrease for two vertices of the polytope τ (t) defined by (9)-(10),
t ∈ [tk ,tk+1 ). Due to V (t) is continuous in [t0 , ∞), then min α
V̇ (xt ) ≤ −ρ kx(t)k2 for a sufficient small ρ > 0, which ensure subject to (20),
asymptotic stability of system (5)-(7), see e.g. [14].
2) For ω (t) ≡ 0 and the condition (15):
1 then, the optimal L2 -gain controller for the system (5)-(7)
V̇ (t) ≤ ξ T Ξ(K)ξ (t) < −ξ T (t) Φ(K)ξ (t). (17) with a control network satisfying Assumptions 1-3 is given
α
by K = Y X −1 .
Integrating both sides of (17) from tk to t ∈ [tk ,tk+1 ), yields,
Z t
Proof. To prove the previous theorem it suffices to show
1 that the derivative of the LKF (19) can be written in the
V (t) −V (tk ) < − [xT (s)Qx(s) + uT (s)Ru(s)]ds.
α tk form (11) and that the aforementioned optimization problem
Obviously ∞ is equivalent to that in (13)-(15).
k=1 [tk ,tk+1 ) = [t0 , ∞). Provided that V (t) is con-
S
tinuous in t, one can see that,
Z t Taking the time derivative of V (t) along the trajectory of
V (t) −V (t0 ) ≤ − [xT (s)Qx(s) + uT (s)Ru(s)]ds. (5) yields that, for t ∈ [tk ,tk+1 ):
t0
V̇ (t) = 2xT (t)Pẋ(t) + xT (t)(Q1 + Q2 )x(t) − xT (t − τm )Q1 x(t − τm )
When t → ∞, the asymptotic stability of the system implies
that V (t) → 0, so that, − xT (t − τM )Q2 x(t − τM ) + ẋT (t)(τM Z1 + ∆τ Z2 )ẋ(t)
Z t Z t−τm
1
Z ∞
− ẋT (s)Z1 ẋ(s)ds − ẋT (s)Z2 ẋ(s)ds. (21)
−V (t0 ) < − [xT (s)Qx(s) + uT (s)Ru(s)]ds ⇒ J < αV (t0 ). t−τM t−τM
α t0
The augmented state vector is defined as: ξ T (t) =
The value of V (t0 ) depends on the initial condition φ (t) and x (t), xT (t − τ (t)), xT (t − τm ), xT (t − τM ), ω T (t) . Then,
T
it is a measure of its norm. Therefore, minimizing α the cost the following null terms are added to the right-hand side
function J is minimized regardless of the initial conditions. of (21):
3) For ω 6= 0 and under zero initial conditions: Z t
T T T 0 = 2ξ T (t)N̄ x(t) − x(t − τ (t)) − ẋ(s)ds ,
V̇ (t) ≤ −z(t) z(t) + γ ω (t)ω (t). (18) t−τ (t)
Z t−τ (t)
Integrating both sides of (18) and using the same continuity T
0 = 2ξ (t)S̄ x(t − τ (t)) − x(t − τM ) − ẋ(s)ds ,
argument as before, one can see that, t−τM
Z t Z t
Z t−τm
T
V (t) −V (t0 ) ≤ − zT (s)z(s)ds + γ 2 ω T (s)ω (s)ds. 0 = 2ξ (t)M̄ x(t − τm ) − x(t − τ (t)) − ẋ(s)ds ,
t−τ (t)
t0 t0
2 T 2 T
0 = γ ω (t)ω (t) − γ ω (t)ω (t),
Then, letting t → ∞, taking into account that under zero initial
condition V (t0 ) = 0 and the positive definitiveness of the 0 = ξ T (t)C̄C̄T ξ (t) − zT (t)z(t).
functional, it can be shown that, where:
Z ∞ Z ∞
T 2 T
z (s)z(s)ds ≤ γ ω (s)ω (s)ds ⇒ kz(t)k2 ≤ γ kω (t)k2 N̄ T = [N1T N2T 0 0 0] , M̄ T = [M1T M2T 0 0 0];
t0 t0
T
S̄ = [S1T S2T 0 0 0] , C̄ = [C T
DK 0 0 0].

With these terms, equation (21) can be rewritten as:
Lemma 1 proposes a general solution to the proposed control
problem. It can be used for different LKFs and for different V̇ (t) = ξ T (t)(Γ + C̄C̄T )ξ (t) + ẋT (t)(τM Z1 + ∆τ Z2 )ẋ(t)
Z t Z t
network constraints. − ẋT (s)Z1 ẋ(s)ds − 2ξ T (t)N̄ ẋ(s)ds
t−τ (t) t−τ (t)
B. A Solution to the optimal L2 -gain control problem Z t−τm Z t−τm
− ẋT (s)Z2 ẋ(s)ds − 2ξ T (t)M̄ ẋ(s)ds
In this section the previous result is applied to a particular t−τ (t) t−τ (t)
LKF and a solution to the optimal L2 -gain control problem Z t−τ (t) Z t−τ (t)
is given. Particularly, we use the following LKF: − ẋT (s)(Z1 + Z2 )ẋ(s)ds − 2ξ T (t)S̄ ẋ(s)ds
t−τM t−τM
Z t Z t
2 T
V (t) = xT (t)Px(t) + xT (s)Q1 x(s)ds + xT (s)Q2 x(s)ds + γ ω (t)ω (t) − zT (t)z(t). (22)
t−τm t−τM
Z 0 Z t Z −τm Z t Now, using the well-known upper bound for the inner
+ ẋT (s)Z1 ẋ(s)dsd θ + ẋT (s)Z2 ẋ(s)dsd θ .(19) product of two vectors:
−τM t+θ −τM t+θ
−aT Xa − 2bT a ≤ bT X −1 b, X > 0,
where all the involved matrices are required to be definite
positive. the following upper bounds for the integral terms in (22) can
The following theorem offers a particular solution for the be found:
proposed problem. Z t Z t
Theorem 1. Given scalars τm , τM , γ , ε > 0 and the weigh- − ẋT (s)Z1 ẋ(s)ds − 2ξ T (t)N̄ ẋ(s)ds
t−τ (t) t−τ (t)
ting matrices Q and R, if the matrices X, Q̃1 , Q̃2 , Z̃1 , Z̃2 >
0 and any matrices Y, Ñi , M̃i , S̃i , (i = 1, 2) of appropriate ≤ (τ (t) + ε )ξ T (t)N̄Z1−1 N̄ T ξ (t),

1346
Γ̃ (τ (t) + ε )Ñ (τ (t) + ε − τm )M̃ (τM + ε − τ (t))S̃ τM Ã ∆τ Ã C̃ Q̃ R̃
 
 ∗ −(τ (t) + ε )Z̃1 0 0 0 0 0 0 0 

 ∗ ∗ −(τ (t) + ε − τm )Z2 0 0 0 0 0 0 


 ∗ ∗ ∗ −(τM + ε − τ (t))(Z̃1 + Z̃2 ) 0 0 0 0 0 


 ∗ ∗ ∗ ∗ −τM X Z̃1−1 X 0 0 0 0  < 0,

(20)
∗ ∗ ∗ ∗ ∗ −∆τ X Z̃2−1 X 0 0 0
 
 

 ∗ ∗ ∗ ∗ ∗ ∗ −I 0 0 

 ∗ ∗ ∗ ∗ ∗ ∗ ∗ −α Q−1 0 
∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ −α R−1
where
Ñ T = [Ñ1T Ñ2T 0 0 0];   Q̃T = [X 0 0 0 0];
M̃ T
= [M̃1T M̃2T 0 0 0]; θ11 θ12 M̃1 −S̃1 Bω R̃T = [0 Y 0 0 0];
 ∗ θ22 M̃2 −S̃2 0 
S̃T = [S̃1T S̃2T 0 0 0]; 
Γ̃ =  ∗ ∗ −Q̃1 0 0

, θ11 = AX + XAT + Q̃1 + Q̃2 + Ñ1 + Ñ1T ;
C̃T = [CX T DY 0 0 0]; BY − Ñ1 + S̃1 − M̃1 + Ñ2T ;
 
 ∗ ∗ ∗ −Q̃2 0  θ12 =
ÃT = [AX BY 0 0 Bω ]; ∗ ∗ ∗ ∗ −γ 2 I θ22 = −Ñ2 − Ñ2T + S̃2 + S̃2T − M̃2 − M̃2T ;

Z t−τm Z t−τm the next section. Otherwise, some null matrices appears
− ẋT (s)Z2 ẋ(s)ds − 2ξ T (t)M̄ ẋ(s)ds in the diagonal of the LMIs. It is worth to mention that
t−τ (t) t−τ (t)
≤ (τ (t) − τm + ε )ξ T (t)M̄Z2−1 M̄ T ξ (t), this modification does not introduce any conservatism, since
ε > 0 can be chosen as small as necessary, i.e., ε → 0+ .
Z t−τ (t) Z t−τ (t)
− ẋT (s)(Z1 + Z2 )ẋ(s)ds − 2ξ T (t)S̄ ẋ(s)ds C. Algorithm for Controller Design
t−τM t−τM
≤ (τM − τ (t) + ε )ξ T (t)S̄(Z1 + Z2 )−1 S̄T ξ (t). (23) Notice that (20) is not an LMI. There is procedure (see
[15]) which let us to address the nonlinear matrix inequality
where the constant ε ∈ R+ have been introduced in the (20) by introducing some new matrix variables. First, define
bounding terms for design convenience. Then, combining two variables T1 , T2 such that,
(22) with (23), it can be shown that, for t ∈ [tk ,tk+1 ),
X Z̃1−1 X ≥ T1 > 0, X Z̃2−1 X ≥ T2 > 0, (27)
V̇ (t) ≤ ξ (t) Γ + (τ (t) + ε )N̄Z1−1 N̄ T + (τ (t) − τm + ε )M̄Z2−1 M̄ T
T
which is equivalent to:
+ (τM − τ (t) + ε )S̄(Z1 + Z2 )−1 S̄T + C̄C̄T + ĀτM Z1 ĀT
−T1−1 X −1

−T2−1 X −1

≤ 0, ≤ 0. (28)
+ Ā∆τ Z2 ĀT ξ (t) + γ 2 ω T (t)ω (t) − zT (t)z(t). (24) X −1 −Z̃1−1 X −1 −Z̃2−1

where ĀT = [A BK 0 0 Bω ]. So that, if we define Ξ(K) Now, introducing some new variables,
as: X̄ = X −1 , T̄i = Ti−1 , Z̄i = Z̃i−1 i = 1, 2, (29)
Ξ(K) , Γ + (τ (t) + ε )N̄Z1−1 N̄ T + (τ (t) − τm + ε )M̄Z2−1 M̄ T + (25)
equation (28) can be rewritten as,
(τM − τ (t) + ε )S̄(Z1+Z2 )−1 S̄T+ C̄C̄T + ĀτM Z1 ĀT + Ā∆τ Z2 ĀT ,
−T̄1 X̄ −T̄2 X̄
it yields that the derivative of the LKF (19) can be written ≤ 0, ≤ 0. (30)
X̄ −Z̄1 X̄ −Z̄2
as in (11). It remains to prove that the optimization problem
given in Lemma 1 is analogous to that in Theorem 1. Con- This way, instead of using the original condition (20), the
sidering the equation (15) in Lemma 1 and the Assumption following nonlinear minimization p
5, yields,
−R −Q ¯
α Ξ(K) < −Φ(K) ⇔ Ξ(K) − K̄ T K̄ − I¯T I<0 (26)
α α
where K̄ T = [0 K 0 0 0]T and I¯T =
T
[I 0 0 0 0] . From equation (25) and (26), applying
Schur complements leads to a matrix inequality with
the same structure as (20). To obtain (20) is sufficient
to introduce the definitions X = P−1 , Q̃i = XQi X,
M̃i = XMi X, Ñi = XNi X, S̃i = XSi X, Z̃i = XZi X, i = 1, 2
and pre- and post-multiply the matrix inequality by
diag = [X, X, X, X, I, X, X, X, I, I, I, I, I].

Remark. The scalar parameter ε > 0 needs to be in-
troduced in order to make feasible the LMIs appearing in

1347
In order to solve the aforementioned minimization vehicle are required (position, speed). This data is nonet-
problem (31) the algorithm introduced in [15] can be heless subject to delays and losses due to unreliable data
implemented. Theorem 2.1 in [15] proved the convergence links, or hardware limitations (bus access policy).
of this algorithm. Thus, the control signal F2 (t) can be obtained as
 R 
e(t − τ (t))
F2 (t) = K  e(t − τ (t)) 
IV. APPLICATION OF THE OPTIMAL L2 -GAIN y(t − τ (t))
CONTROLLER
where transmission delays are implicit through the term τ (t).
This section describes simulations on a particular problem Simulations are performed assuming system parameters
to show the performance of the optimal L2 -gain controller. as m = 1200 Kg, AT = 1.5 m2 , ca = 0.15, cr = 0.015 and a
Consider a vehicle tracking system intended to keep a nominal speed vn = 120 Km/h.
constant distance between two moving objects. More preci- Network induced delays are assumed to lay in the range
sely, assume a master vehicle that moves at unknown time- from 0ms to 200ms, and a sample period h = 500ms.
varying speed, which is followed by a slave vehicle trying The following simulations consider an inter-vehicle set-
to keep a constant distance from the first. point distance of 20 meters, with a master vehicle moving
The equations of the dynamical system can be written as at constant speed of 120Km/h, except around t = 45s where
follows:
the vehicle slows down temporarily before recovering the
e(t) = p1 (t) − p2 (t) − lr , nominal speed.
ė(t) = v1 (t) − v2 (t) = y(t), In Figure 1 the behavior of the proposed L2 -gain control
ẏ(t) = a1 (t) − a2 (t). is compared for different disturbance attenuation levels when
network-induced delays are considered. It appears as expec-
where p(t), v(t) and a(t) denote positions, velocities and ted that deviations from the setpoint are reduced as parameter
accelerations; and e(t) and y(t) denote the errors in posi- γ is decreased. This is achieved at the cost of more aggressive
tion and velocity, respectively. Moreover, forces equilibrium control actions that must be properly weighted to not exceed
yields: actuator limits. The trade-off between the L2 -gain of the
Fi (t) − Fa,i(t) − Fr,i(t) = mi ai (t), i = 1, 2. system and the optimality can be also checked in the Figure
1. The controller with free γ results in a cost J1 = 659.1, the
Where Fi is the motor force, Fa,i is the aerodynamical drag controller with γ < 1.8 achieves a cost J2 = 670.9 and the
and Fr,i is the rolling friction, given by: controller with γ < 0.8 has a cost J3 = 2478.
1 Similarly, the performance of the system can be modu-
Fa (t) = ca AT ρaire v2 (t), Fr (t) = cr mg cos(α (t)).
2 lated acting on the weighting policy in matrices Q and
where, R. Figure 2 shows the effect of this trade-off. In both
cases an attenuation factor γ < 1.8, and a weighting matrix
• ca , cr : aerodynamic and tire-road drag coefficients,
Q = diag(100, 100, 0.01) are chosen. Matrix R is set in
respectively.
one case R = 10−5 prioritizing tracking performance, and
• AT : vehicle’s aerodynamic cross-section.
R = 10−2 assuming more stringent control action constraints.
• ρair : air density.
As expected, tracking error is reduced in the first case, since
• m: vehicle mass.
control effort is limited in the latest case using the weighting
• g: gravity constant.
matrix R.
• α (t): road slope angle.
These simulations show how the method can be used to
For the sake of simplicity, both vehicles will be assu-
med with identical characteristics. The linearization of the adjust different performance requirements and disturbance
aerodynamical force using a Taylor serie expansion yields rejection capabilities for network control systems, where
the following dynamic equations for the system: parameter γ together with matrices Q and R can be tuned
 R    R  for such purpose.
d  e(t)   0 0 1 e(t)
e(t) = 0 1 0   e(t)  V. C ONCLUSIONS
dt y(t) 0 −c3 c00 y(t)
    This paper deals with the problem of designing optimal
0 0 control laws for network control systems with L2 -gain dis-
+  0  F2 (t) +  0  F1 (t) (32)
−c3 c3 turbance rejection constraints.
Linear time invariant systems are considered to be contro-
with c0 = ρ AT cd and c3 = 1/m. lled through a communication network that imposes delays
Assume that the slave vehicle receives position and speed and data losses in the communications.
from the master vehicle at a sampling rate h. The objective The problem, also referred as mixed H2 /H∞ , is derived
is computing a control action F2 (t), such that the distance based on a Lyapunov-Krasovskii approach, considering a
between both vehicles remains as close as possible to a polytopic covering of the time-delay range. Solution for the
constant set point lr . The controller is implemented on the problem is provided in terms of a set of LMIs together with
slave vehicle, though dynamic variables from the master an optimization procedure.

1348
24
Networked optimal L2−gain
controller with free γ
23 Networked optimal L −gain
2
controller with γ<1.8
22 Networked optimal L2−gain

Inter−vehicle distance (m)

controller with γ<0.8

21

20

19

18

17
0 10 20 30 40 50 60 70 80
time (s)

Fig. 1. System response with network effects for various disturbance attenuation levels

R EFERENCES
24 [1] J. Hespanha, P. Naghshtabrizi, and Y. Xu, “A survey of recent results in
Networked optimal
−5 networked control systems,” Proceedings of the IEEE, special edition,
L −gain controller with R=10
Inter−vehicle distance (m)

2 vol. 95, no. 1, pp. 138–162, 2007.

22 Networked optimal [2] P. Naghshtabrizi and J. Hespanha, “Designing an observer-based
−2
L −gain controller with R=10 controller for a network control system,” in Proceedings of the 44th
2
IEEE Conference on Decision and Control, and the European Control
20
Conference, Seville, Spain, December 2005, pp. 848–853.
[3] J. Xiong and J. Lam, “Stabilization of linear systems over networks
18
with bounded packet loss,” Automatica, vol. 43, no. 1, pp. 80–87,
2007.
[4] S. Tatikonda and S. Mitter, “Control under communication cons-
16 traints,” IEEE Transactions on Automatic Control, vol. 49, no. 7, pp.
0 10 20 30 40 50 60 70 80 1056–1068, 2004.
time (s) [5] M. S. Mahmoud, Robust Control and Filtering for Time-delay Systems.
New York: Marcel Dekker, Inc., 2000.
[6] D. Zhang and L. Yu, “Equivalence of some stability criteria for linear
2000 time-delay systems,” Applied Mathematics and Computation, vol. 202,
no. 1, pp. 395–400, 2008.
1000
[7] S. Xu and J. Lam, “On equivalence and efficiency of certain stability
Control action u (N)

criteria for time-delay systems,” IEEE Transactions on Automatic

0
Control, vol. 52, no. 1, pp. 95–101, 2007.
−1000
[8] Y. V. Mikheev, V. A. Sobolev, and E. Fridman, “Asymptotic analysis
of digital control systems,” Automatic and Remote Control, vol. 49,
−2000 pp. 1175–1180, 1988.
[9] D. Yue, Q. L. Han, and J. Lam, “Network-based robust H∞ control of
−3000 systems with uncertainty,” Automatica, vol. 41, no. 6, pp. 999–1007,
2005.
−4000 [10] V. Gupta, B. Hassibi, and R. M. Murray, “Optimal LQG control across
0 10 20 30 40 50 60 70 80 packet-dropping links,” Systems and Control Letters, vol. 56, no. 6,
time (s)
pp. 439–446, 2007.
[11] J. Qiu, H. Yang, and J. Zhang, “Optimal guaranteed cost control for
a class of uncertain systems with both state and input delays via
delta operator approach,” in Eighth ACIS International Conference
Fig. 2. Trading off performance vs. control effort. Effect of the weighting on Software Engineering, Artificial Intelligence, Networking, and
policy on R and Q Parallel/Distributed Computing, The Haier Training Center, Qingdao,
China, July-August 2007, pp. 417–422.
[12] O. I. Kosmidou and Y. S. Boutalis, “A linear matrix inequality
approach for guaranteed cost control of systems with state and input
Simulation results show the effectiveness of the method delays,” IEEE Transactions on Systems, Man, and CyberneticsPart A:
to trade-off with a unified design methodology, performance, Systems and Humans, vol. 36, no. 5, pp. 936–942, 2006.
[13] J. H. Kim, “Robust mixed H2 /H∞ control of time-varying delay
control effort and disturbance rejection capabilities. systems,” International Journal of Systems Science, vol. 32, no. 11,
pp. 1345–1351, 2001.
[14] J. C. Hale and S. M. Verduyn Lunel, Introduction of functional
VI. ACKNOWLEDGMENTS differential equations. New York: Springer, 1993.
[15] L. El Ghaoui, F. Oustry, and M. AitRami, “A cone complementary li-
nearization algorithm for static output-feedback and related problems,”
The authors gratefully acknowledge CICYT (DPI2007- IEEE Transactions on Automatic Control, vol. 42, no. 8, pp. 1171–
64697), and the European Commission (FeedNetBack Pro- 1176, 1997.
ject, grant agreement 223866), for funding this work.

1349