Available online at www.sciencedirect.

com

Fuzzy Sets and Systems 212 (2013) 78 – 96

www.elsevier.com/locate/fss

T–S model-based nonlinear moving-horizon H ∞

control and applications
Ping Wang, Shuyou Yu, Hong Chen∗
Department of Control Science and Engineering, Jilin University, NanLing Campus, 130025 Changchun, PR China

Received 11 April 2011; received in revised form 7 June 2012; accepted 7 June 2012
Available online 16 June 2012

Abstract
Using a T–S fuzzy model, we propose a formulation for a moving horizon H∞ control problem for nonlinear systems. Our main
contribution is combination of T–S fuzzy models and a moving horizon H∞ control scheme by Chen and Scherer to address the
disturbance attenuation of nonlinear constrained system. A sufficient condition for the disturbance amplitude is given to guarantee
the feasibility of the optimization problem at each time. A parameter-dependent state feedback control law is adopted, and the
corresponding optimization problem is reduced to a convex optimization problem involving linear matrix inequalities (LMIs). The
H∞ attenuation index for the nonlinear moving horizon H∞ control problem is adapted online to satisfy time-domain constraints.
Finally, the effectiveness of the proposed scheme is successfully demonstrated by control of a continuous stirred tank reactor and
active queue management for routers.
© 2012 Elsevier B.V. All rights reserved.

Keywords: Fuzzy control; T–S fuzzy models; Nonlinear moving horizon control; H∞ performance; Time-domain constraints; Linear matrix
inequality (LMI)

1. Introduction

In recent years, Takagi–Sugeno (T–S) fuzzy control [1] has become one of the most popular and promising approaches
in model-based fuzzy control because it can combine the flexibility of fuzzy logic theory and the rigorous mathematical
theory of linear or nonlinear system in a unified framework. Numerous successful applications of fuzzy control have
sparked analysis and design of fuzzy control systems described by a set of If–Then rules [2].
There are many results in the literature on stability analysis and design. For continuous-time T–S fuzzy control
systems, three different types of Lyapunov function have been used for stability analysis and fuzzy controller synthesis:
quadratic, nonquadratic, and piecewise. Among these, nonquadratic Lyapunov functions (NQLFs), also called fuzzy
Lyapunov functions, usually allow more relaxed stability conditions [3–5]. However, the problem of finding the least
conservative sufficient conditions for fuzzy summations remains open. Many results for relaxed stabilization conditions
have been published [6–8].
For design problems involving other performance requirements, such as response speed, constraints on control input
and output and robust stabilization, some interesting results for T–S fuzzy control have been published. Kruszewski et al.

∗ Corresponding author. Tel.: +86 431 85094831; fax: +86 431 85095243.
E-mail addresses: [email protected] (P. Wang), [email protected] (S. Yu), [email protected] (H. Chen).

0165-0114/$ - see front matter © 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.fss.2012.06.002
 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96 79

considered discrete-time uncertain nonlinear models in a T–S form and studied their stabilization using a nonquadratic
Lyapunov function [9]. Qiu et al. investigated delay-dependent robust H∞ filtering for a class of uncertain discrete-time
state-delayed T–S fuzzy systems [10]. Most H∞ control formulations based on T–S fuzzy models concentrate on robust
stability conditions and robust performance analysis, but without taking time-domain constraints into account.
Model predictive control (MPC), or receding horizon control (RHC), is one of the most powerful optimizing control
approaches for a large class of constrained nonlinear industrial processes in which a control sequence is determined by
optimizing a finite horizon cost at each sampling instant, based on an explicit process model and state measurement
[11]. Many successful applications of MPC using fuzzy models have been reported. An efficient offset-free predictive
control approach with output feedback for nonlinear processes based on their approximate fuzzy models was developed
by Zhang et al. [12]. Considering the output and input constraints, Khairy et al. used a fuzzy model to predict future
behavior and designed a stabilizing model-based predictive controller [13].
For robust MPC, Zhang et al. proposed two stable fuzzy-model predictive controllers based on piecewise Lyapunov
functions and the min–max optimization of a quasi-worst-case infinite-horizon objective function [14]. However, for
strong constraints and/or uncertainty, the monotonicity required for the value function needs further discussion.
Building on the work of Chen and Scherer [15], moving-horizon constrained H∞ control based on T–S fuzzy models
for a constrained discrete-time nonlinear system is presented. In the nominal case, MPC and T–S fuzzy models have
been combined to achieve exponential stability of nonlinear systems subject to output and input constraints [13].
For robust MPC, this paper combines moving-horizon constrained H∞ control and T–S fuzzy models to address the
disturbance attenuation issue. First, the nonlinear system considered is represented by local models according to fuzzy
rules, and the local models are blended into an overall single model through fuzzy membership functions. For this T–S
fuzzy model, following Chen and Scherer [15], a parameter-dependent state feedback control law is adopted and the
corresponding optimization problem is reduced to a convex optimization problem involving linear matrix inequalities
(LMIs). This leads to a tractable nonlinear fuzzy moving-horizon H∞ control problem for which the H∞ attenuation
index is adapted online to satisfy time-domain constraints.
The remainder of the paper is organized as follows. Section 2 introduces the constrained nonlinear system to be
controlled and the T–S fuzzy models. Section 3 provides some preparatory results on LMI-based fuzzy constrained
H∞ control problems. An efficient algorithm for moving-horizon nonlinear H∞ control is introduced in Section 4. The
corresponding synthesis problem can be formulated as an LMI problem. In Section 5, the proposed controller is applied
to a continuous stirred tank reactor and active queue management at routers. The simulation results demonstrate the
effectiveness of the proposed algorithm. Section 6 concludes with a brief summary.

2. Preliminaries

2.1. System representation

Consider the smooth nonlinear control system

⎧
⎨ x(k + 1) = f (x(k), u(k), v(k)), x(k) = xk , k≥0,
⎪
z 1 (k) = g1 (x(k), u(k), v(k)), (1)
⎪
⎩
z 2 (k) = g2 (x(k), u(k))

subject to time-domain constraints

− z 2s,max ≤z 2 (k)≤z 2s,max for all k ≥ 0, s = 1, 2, . . . , p, (2)

where x(k) ∈ Rn and u(k) ∈ Rm are the state and control inputs, respectively, v(k) ∈ Rq is the disturbance, which
belongs to l2 [0, ∞), and z 1 (k) ∈ Z1 ⊂ R p1 and z 2 (k) ∈ Z2 ⊂ R p2 are the performance output and the output to be
constrained, respectively. The origin (0, 0) belongs to the set Z2 , and z 2,max is the given constant-compatible vector. The
nonlinear functions f (·, ·, ·), g1 (·, ·, ·) and g2 (·, ·) are twice continuously differentiable in all arguments. For simplicity
of notation, we assume that the origin is an equilibrium in (1) without disturbances, i.e. f (0, 0, 0) = 0, g(0, 0) = 0.
The task is to design a controller for (1) such that the closed-loop system is asymptotically stable and the H∞ norm
from disturbance v to performance output z 1 is minimized while the time-domain constraints (2) are satisfied.
80 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96

2.2. T–S fuzzy models and parallel distributed compensation

It has been proven that any twice differentiable continuous nonlinear function can be approximated, to any degree
of accuracy, using T–S fuzzy models [2]. The main feature of the T–S fuzzy model is to express the local dynamics of
each fuzzy rule by a linear system model. The overall fuzzy model of the system is obtained by fuzzy blending of the
linear system models.
The ith rule of the T–S fuzzy model of the nonlinear system (1) is of the following form:
Ri :
IF 1 (k) is M1i , 2 (k) is M2i , . . . , and r1 (k) is Mri1 ,
THEN x(k + 1) = Ai x(k) + B1i v(k) + B2i u(k),
z 1 (k) = C1i x(k) + D11i v(k) + D12i u(k),
z 2 (k) = C2i x(k) + D22i u(k),
i = 1, 2, . . . , r,

where Ri represents the ith fuzzy rule, M ij is the fuzzy set and r is the number of model rules. For (k) = [1 (k),
2 (k), . . . , r1 (k)] ∈ Rr1 , premise variables  j (k) may be functions of the state variables, external inputs, and/or time.
Nonlinear system (1) can be approximated in the form of a T–S fuzzy model as
⎧
⎪ r
⎪
⎪ x(k + 1) = h i ()[Ai x(k) + B1i v(t) + B2i u(k)],
⎪
⎪
⎪
⎨ i=1
r
z 1 (k) = h i ()[C1i x(k) + D11i v(k) + D12i u(k)], (3)
⎪
⎪
⎪
⎪
i=1

⎪
⎪
r
⎩ z 2 (k) = h i ()[C2i x(k) + D22i u(k)],
i=1

where
 ((k)) 
r1
h i ((k)) = r i , i ((k)) = M ij ( j (k)).

i=1 i ((k))
j=1

The term M ij ( j (k)) is the grade of membership for  j (k) in M ij .

Since

r
i ((k))≥0, i ((k)) > 0, i = 1, 2, . . . , r,
i=1

we have

r
h i ((k))≥0, h i ((k)) = 1, i = 1, 2, . . . , r.
i=1

In the parallel distributed compensation (PDC) design [2,16], each control rule is designed from the corresponding
rule of a T–S fuzzy model. The fuzzy controller designed shares the same fuzzy sets as the fuzzy model in the premise
parts. Hence, we construct the following fuzzy controller via PDC:
Ci :
IF 1 (k) is M1i , 2 (k) is M2i , . . . , and r1 (k) is Mri1 , (4)
THEN u(k) = K i x(k), i = 1, 2, . . . , r,

where K i ∈ Rm×n is a constant feedback matrix. The whole controller can be expressed as
u(k) = ((k))x(k), (5)
 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96 81


where ((k)) = ri=1 h i ((k))K i . Substituting (5) into (3), we obtain the closed-loop system model
⎧
⎨ x(k + 1) = Acl ((k))x(k) + Bcl ((k))v(k),
⎪
z 1 (k) = C1cl ((k))x(k) + D1cl ((k))v(k), (6)
⎪
⎩
z 2 (k) = C2cl ((k))x(k),

where

r 
r
Acl ((k)) = h i ((k))h j ((k))(Ai + B2i K j ), (7a)
i=1 j=1


r 
r
C1cl ((k)) = h i ((k))h j ((k))(C1i + D12i K j ), (7b)
i=1 j=1


r 
r
C2cl ((k)) = h i ((k))h j ((k))(C2i + D22i K j ), (7c)
i=1 j=1


r
Bcl ((k)) = h i ((k))B1i , (7d)
i=1

r
D1cl ((k)) = h i ((k))D12i . (7e)
i=1

Our control problem has been converted to finding a parameter-dependent control law ((k)) such that closed-loop
system (6) is asymptotically stable, the H∞ norm from disturbance v to performance output z 1 is minimized, and the
constrained output z 2 of the system meets the time-domain constraints.
Synthesis and analysis of a fuzzy control system can usually be reduced sufficiently to a feasibility problem with
parameter-dependent matrix inequalities [2]. Here, we introduce a less conservative scheme to convert the parameter-
dependent matrix inequalities to LMIs.

Lemma 1 (Gao [17]). If there exist matrices Tii = TiiT , Ti j = T jiT (i  j = 1, 2, . . . , r ) such that Li j (1≤i, j≤r )
satisfies

Lii ≤Tii , i = 1, 2, . . . , r, (8a)

Li j + L ji ≤Ti j + Ti Tj , j < i, (8b)
[Ti j ]r ×r ≤0, (8c)

then the parameter-dependent matrix inequalities


r 
r
h i ((k))h j ((k))Li j ≤0 (9)
i=1 j=1
r
are feasible, where h i ((k))≥0, i=1 h i ((k)) = 1, ∀(k) and
⎛ ⎞
T11 · · · T1r
⎜ . . ⎟
=⎜ . . . ... ⎟ .
[Ti j ]r ×r ⎝ . ⎠
Tr 1 · · · Trr

Remark 2.1. For the case “< 0”, we obtain similar results.
82 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96

3. Constrained H∞ control based on T–S models

For the constrained H∞ control problem, we can state the following result for closed-loop system (6).

Theorem 1 (Constrained nonlinear H∞ control based on T–S models). For a given , suppose that the optimization
problem

min  (10)
,Q,Y1 ,Y2 , ...,Yr

subject to


r 
r
h i ()h j ()Li j ≥0, (11)
i=1 j=1


r 
r
h i ()h j ()Fi j,s ≥0, s = 1, 2, . . . , n z2 (12)
i=1 j=1

has an optimal solution (, Q, Y1 , Y2 , . . . , Yr ). In (11) and (12), Li j and Fi j,s are given as
⎡ ⎤
Q ∗ ∗ ∗
⎢ I ∗ ∗ ⎥
⎢ 0 ⎥
Li j = ⎢ ⎥,
⎣ Ai Q + B2i Y j B1i Q 0 ⎦
C1i Q + D12i Y j D11i 0 I
 
1 2
 z 2s,max esT (C2i Q + D22i Y j )
Fi j,s = ,
∗ Q

where es is the element of the basis vector in R p2 and ∗ denotes expressions that can be deduced by symmetry. Then
(1) controlled with the parameter-varying state feedback control u = ()x has the following properties.
(I) The H∞ norm from disturbance
∞ v to performance output z 1 is less than (or equal to) .
(II) If the disturbance energy i=0 v(i) 2 ≤ and the initial state x(0) satisfy  + V (x(0))≤, then
(i) All perturbed state trajectories remain in an ellipsoid defined as

(P, ) := {x ∈ Rn |V (x)≤}. (13)

(ii) The constraints (2) are respected.


In the above, () = rj=1 h j ()K j with K j = Y j Q −1 , V (x) = x T P x with P = Q −1 .

Proof. (I) Consider the Lyapunov function candidate V (x) = x T P x with P = P T > 0. Then the H∞ performance
of closed-loop system (6) is bounded by  if the following dissipation inequality is satisfied [18]:

V (x(k + 1)) − V (x(k)) + −1 z 1 (k)T z 1 (k) − v(k)T v(k)≤0. (14)

By taking the parameter-varying state-feedback control (5) and substituting (6) into (14), we have
 T  
x(k) T (·)P A (·) − P + −1 C T (·)C (·)
Acl T (·)P B (·) + −1 C T (·)D (·)
Acl
cl 1cl 1cl cl 1cl 1cl
v(k) BclT (·)P Acl (·) + −1 D1cl
T (·)C (·)
1cl BclT (·)P Bcl (·) + −1 D1cl
T (·)D (·) − 
1cl
 
x(k)
≤0, (15)
v(k)
 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96 83

where Acl (·) = Acl ((k)), Bcl (·) = Bcl ((k)), C1cl (·) = C1cl ((k)) and D1cl (·) = D1cl ((k)). Then (15) reduces to the
condition that P satisfies
 
T (·)P A (·) − P + −1 C T (·)C (·)
Acl AclT (·)P B (·) + −1 C T (·)D (·)
cl 1cl 1cl cl 1cl 1cl
≤0 (16)
BclT (·)P Acl (·) + −1 D1cl
T (·)C (·)
1cl BclT (·)P Bcl (·) + −1 D1cl
T (·)D (·) − 
1cl

and the inequality (16) is equivalent to

⎡ ⎤
P ∗ ∗ ∗
⎢ I ∗ ∗ ⎥
⎢ 0 ⎥
⎢ ⎥ ≥0 (17)
⎣ P Acl ((k)) P Bcl ((k)) P 0 ⎦
C1cl ((k)) D1cl ((k)) 0 I
according to the classical Schur complement. By substituting (7) into (17), we infer that (14) can be guaranteed by the
existence of P > 0 and K j ( j = 1, 2, . . . , r ) satisfying
⎡ ⎤
P ∗ ∗ ∗
r  r ⎢ I ∗ ∗ ⎥
⎢ 0 ⎥
h i ()h j () ⎢ ⎥ ≥0. (18)
⎣ P(Ai + B2i K j ) P B1i P 0 ⎦
i=1 j=1
C1i + D12i K j D12i 0 I

With substitution of Q = P −1 and Y j = K j Q, we observe that the equivalence between (18) and (11) is obtained by
performing a congruence
∞ transformation with diag{Q, I, Q, I }.
(II-i) Assuming i=0 v(i) 2 dt≤, (14) implies that


k−1
V (x(k)) + −1 z 1 (k) 2 ≤V (x(0)) +  (19)
i=0

for all k≥0. Given any x(0), (19) shows that the state trajectory starting from x(0) stays in the ellipsoid defined by (13)
with  :=  + V (x(0)). This means that the ellipsoid (P, ) contains the set of all reachable states for the closed-loop
system.
(II-ii) Following [19], we infer for all k ≥ 0 that
|z 2s (k)|2 = |esT C2cl ((k))x(k)|2
 2
 
 T r  r

= es h i ((k))h j ((k))(C2i + D22i K j )x(k)
 i=1 j=1 

r 
r
≤ h i ((k))h j ((k))|esT (C2i + D22i K j )x(k)|2
i=1 j=1

(by virtue of x(k) ∈ (P, ))


r 
r
≤ h i ((k))h j ((k)) max |esT (C2i + D22i Yi Q −1 )x|2
x∈(P,)
i=1 j=1

(using the Cauchy–Schwarz inequality)


r 
r
≤ h i ((k))h j ((k)) esT (C2i + D22i Yi )Q −1/2 22 .
i=1 j=1
Using (12), we arrive at
|z 2s (k)| ≤ z 2s,max . (20)
Hence, the feasibility of (12) implies that the time-domain constraints in (2) are respected. 
84 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96

From (12), we know that the larger the value of , the smaller is the feasible set of (Q, Y1 , Y2 , . . . , Yr ) and hence
the larger is the optimal value . This implies worse performance. However, property II(ii) reveals that the smaller
the value of , the smaller is the disturbance energy that guarantees satisfaction of the time-domain constraints. This
contradiction motivates us to exploit the moving-horizon strategy to online manage the trade-off between constraint
satisfaction and good performance.

4. Moving-horizon nonlinear H∞ control based on a T–S model

At time k, nonlinear moving-horizon H∞ control [15] based on a T–S model can be formulated as
min  s.t. (11), (12), and (21a)
,Q,Y1 ,Y2 , ...,Yr
 
 −  x(k)T
≥0, (21b)
x(k) Q
 
p0 − pk−1 + x T (k)Pk−1 x(k) x(k)T
≥0, (21c)
x(k) Q
where (, Q, Y1 , Y2 , . . . , Yr ) are the independent variables. That is, the values of x(k),  and  are given for the solution
at time k = 0 (constraint (21c) drops), and the values of x(k), , , p0 and Pk−1 are given for the solution at time
k > 0. Then the independent variables , Q, Y1 , . . . , Yr in (21) are to be determined at time k. Constraint (21b) forces
the actual state x(k) to be in the interior of the ellipsoid (P,  − ) and dissipation constraint (21c) is required to
guarantee that the closed-loop moving-horizon system is dissipative. With the actual state x(k), if (21) has an optimal
solution, denoted as (k , Q k , Y1k , Y2k , . . . , Yr k ), we then define a feedback control law according to the MPC principle
as follows:

r
u(k) = h j ()K jk x(k) (22)
j=1

with K jk = Y jk Q −1
k . The scalar pk in (21c) is recursively updated as

pk := pk−1 − (x T (k)Pk−1 x(k) − x T (k)Pk x(k)) (23)

with Pk = Q −1
k .
A more detailed discussion on the dissipation constraint can be found in the literature [15,20].
To solve the parameter-dependent optimization problem (21) efficiently online, we reduce inequality constraints (11)
and (12) to LMIs according to Lemma 1. That is, (11) and (12) are feasible if there exist Ti j and Mi j (i, j = 1, 2, . . . , r )
such that
⎧
⎪ L ≥T , i = 1, 2, . . . , r,
⎨ ii ii
Li j + L ji ≥Ti j + Ti Tj , j < i, (24a)
⎪
⎩
[Ti j ]r ×r ≥0.
⎧
⎪ F ≥Mii,s , i = 1, 2, . . . , r,
⎨ ii,s
Fi j,s + F ji,s ≥Mi j,s + MiTj,s , j < i, (24b)
⎪
⎩
[Mi j,s ]r ×r ≥0.
Hence, optimization problem (21) becomes
min  s.t. (24a), (24b), (21b), and (21c) hold. (25)
,Q,Y j

Here, the independent variables are the same as in (21). By virtue of the moving horizon principle of MPC, optimization
problem (21) is solved at each sampling time k, updated with the actual state x(k) [11]. This is implementable, since
pk−1 and Pk−1 have been determined at the previous sampling time k − 1 and are held fixed. The procedure for solving
the moving-horizon H∞ control problem online based on T–S models can be summarized as follows.
 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96 85

Algorithm 1.
• Step 0. Obtain the T–S fuzzy model for the controlled system and choose (determine) parameters  and  for LMI
optimization problem (21b).
• Step 1. At time k = 0, get x(0) and solve (25) without (21c) to obtain (0 , P0 , Y j0 ). Compute p0 = x(0)T P0 x(0)
and go to step 3.
• Step 2. At time k > 0, get x(k) and solve (25) to obtain an optimal solution (k , Pk , Y jk ).
• Step 3. Set K jk = Y jk Q −1
k and pk = pk−1 − (x (k)Pk−1 x(k) − x (k)Pk x(k)). Apply closed-loop control (22) to
T T

the plant, replace k by k + 1 and go to step 2.

Theorem 2 (Moving-horizon H∞ control based on T–S models). For given (, ), suppose that semi-definite program-
ming problem (25) with actual state x(k) has an optimal solution (k , Q k , Y jk ) at any time k. Then the parameter-varying
state feedback (22) guarantees the following.
(I) The dissipation inequality


k
(¯−1 z 1 (i) 2 −¯ v(i) 2 )≤x(0)T P0 x(0) (26)
i=0

is satisfied, with ¯ := max{0 , 1 , . . . , k }.

(II) The H∞ norm from disturbance v to performance output z 1 is less than ¯ .
(III) The constraints (2) are satisfied.

Proof. (I) Because of the Schur complement, (21c) is equivalent to

p0 − pk−1 + x(k)T Pk−1 x(k) − x(k)T Pk x(k)≥0. (27)

Substituting (23) into the above matrix inequality recursively, we obtain that dissipation constraint (21c) enforces


k
(x(i)T Pi−1 x(i) − x(i)T Pi x(i))≥0. (28)
i=1

The feasibility of (24a) at sampling time k implies that inequality (14) is satisfied with V (x) = x T Pk x and k , i.e.

x(k + 1)T Pk x(k + 1) − x(k)T Pk x(k) + −1

k z(k) z(k) − k w(k) w(k) < 0.
T T
(29)

Hence, cumulating (29) from i = 0 to k leads to


k 
k
¯ −1 z 1 (i) 2 − ¯ w(i) 2 ≤x(0)T P0 x(0) − x(k + 1)T Pl x(k + 1) − (x(i)T Pi−1 x(i) − x(i)T Pi x(i)), (30)
i=0 i=1

where ¯ = max{0 , 1 , . . . , k }. Considering (28), the above inequality becomes


k
¯ −1 z 1 (i) 2 − ¯ v(i) 2 ≤x(0)T P0 x(0) − x(k + 1)T Pk x(k + 1). (31)
i=0

Thus, dissipation inequality (26) is satisfied, since x(k + 1)T Pk x(k + 1)≥0.
(II) Since x(0)T P0 x(0)≥0, we furthermore obtain through (31) that the H∞ norm from disturbance v to performance
output z 1 is less than ¯ .
(III) From the proof of Theorem 1 and Lemma 1, we know that constraints (2) are satisfied at any time k. Thus, the
argument can be continued as time elapses. 
86 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96

Remark 4.1. In robust MPC, we strive in general to solve the following minimax optimization problem for system (3)
and actual state x(k) in the moving-horizon fashion:
∞

min max z 1 (i) 2 .
u v
i=k

Similar to the approach of Chen et al. [21], the above minimax problem for the fuzzy control system can be relaxed
to a Lagrange (weak) dual problem. By a similar derivation, the Lagrange (weak) dual problem is equivalent to
inequality constraint (11) in optimization problem (21) with the feedback law u = K x. Then, solving (21) gives a
solution of the primary minimax problem. Moreover, it is solved repeatedly, with the actual state used to update (11). It
should be pointed out that the proposed approach is performed implicitly an infinite horizon prediction with feedback
parameterization of u = K x, which is clearly different from classical MPC.

Remark 4.2. Compared to studies in which a static state-feedback law is applied during the sampling interval
[15,22,23], we adopt a linear parameter variable control law [Eq. (22)], which introduces extra degrees of freedom in
our optimization problem.

Remark 4.3. Dissipation constraint (21c) can automatically adapt the performance level while respecting time domain
constraints. Moreover, because of recursion of (23) with state pk , the level of dissipation does not have to decrease
in each step. This makes it possible to reduce the infeasibility of (25). However, if the system is affected by large
disturbances, (25) may be still infeasible. In this case, the controller has to maintain the previous values as remedial
action. At the next sampling time, (25) is solved with the new states.

5. Illustrative examples

5.1. Nonlinear stirred tank reactor

In this section, the proposed robust MPC algorithm is applied to a benchmark system, a continuous stirred tank reactor
(CSTR) [24,25]. Assuming a constant liquid volume, the CSTR for an exothermic, irreversible reaction A −→ B is
described by the following dynamic model based on a component balance for reactant A and an energy balance:
⎧  
⎪
⎪ q E
⎪
⎨ AĊ = (C Af − C A ) − k 0 exp − C A,
V RT
  (32)
⎪
⎪ q H E UA
⎪
⎩ Ṫ = (T f − T ) − k0 exp − CA + (Tc − T ) + vT ,
V C p RT V C p
where C A is the concentration of reactant A, T is the reactor temperature, and Tc is the temperature of the coolant
steam. The parameters are q = 100 l/min, V = 100 l, C A f = 1 mol/l, T f = 350 K, = 103 g/l, C p = 0.239 J/(g K),
k0 = 7.2 × 1010 min−1 , E/R = 8750 K, H = −5 × 104 J/mol, and U A = 5 × 104 J/(min K). Under these
eq eq
conditions the steady state is C A = 0.5 mol/l, Tc = 300 K, and T eq = 350 K, which is an unstable equilibrium.
The temperature of the coolant steam is constrained to 250 K≤Tc ≤350 K, the reactant concentration is constrained to
0.4 mol/l≤C A ≤0.6 mol/l, and the reactor temperature is constrained to 335 K≤T ≤365 K.
Following Cao and Frank [26], the model is discretized using a sampling period of Ts = 0.083 min (5 s). The maxi-
mum uncertainty is bounded by vmax , vmax = 10 K/min. The control objective is to minimize H∞ performance from
disturbance vT to concentration C A and reactor temperature T around the steady state, manipulating the temperature
of the coolant in its admissible range for any possible disturbance.
In this example, we treat the term exp(−E/RT ) as a parameter-varying term, where T is the only varying parameter.
Accordingly, nonlinear system (32) can be rewritten as the linear parameter-varying system
⎧   
⎪
⎪ q E q
⎨ Ċ A = − V − k0 exp − RT
⎪ CA + CAf ,
V
    (33)
⎪
⎪ H E q UA UA q
⎪
⎩ Ṫ = − k0 exp − CA − + T+ Tc + vT + T f ,
C p RT V V C p V C p V
 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96 87

where the varying term exp(−E/RT )∈[exp(−E/335R), exp(−E/365R)] since R and E are constant, and the function
exp(−E/RT ) monotonically increases in the set T ∈ [335, 365].
eq eq
We normalize system (33), that is, we define states x1 = (C A − C A )/C A , x2 = (T − T eq )/T eq , control input
eq eq
u = (Tc − Tc )/Tc and external v = vT /vmax , and then the nonlinear system can be approximated by the following
two-rule T–S fuzzy model:

R1 : IF reactor temperature is low (i.e., x2 (t) is about −1),

THEN x(t) = A1 x(t) + B11 v(t) + B21 u(t)
z 1 (k) = C11 x(k) + D111 v(k) + D121 u(k),
z 2 (k) = C21 x(k) + D221 u(k),
R2 : IF reactor temperature is high (i.e., x2 (t) is about 1),
THEN ẋ(t) = A2 x(t) + B12 v(t) + B22 u(t)
z 1 (k) = C12 x(k) + D112 v(k) + D122 u(k),
z 2 (k) = C22 x(k) + D222 u(k),
with
   
0.8957 0 0.7299 0
A1 = , A2 = ,
0.0007 0.7736 0.0005 0.7736
   
0 0
B11 = B21 = , B12 = B22 = .
0.1531 0.0732

For CSTR control, we choose [x1 x2 ]T as performance outputs, and hence

     
1 0 0 0
C1i = , D11i = , D12i = , i = 1, 2.
0 1 0 0

Moreover, the constrained outputs are [x1 x2 u]T and hence

⎛ ⎞ ⎛ ⎞
1 0 0
⎜ ⎟ ⎜ ⎟
C2,i = ⎝ 0 1 ⎠ , D22i = ⎝ 0 ⎠ , i = 1, 2.
0 0 1
The membership functions shown in Fig. 1 are given by
         
−8750 −8750 −8750 −8750
h 1 () = exp − exp exp − exp ,
T 335 365 335
         
−8750 −8750 −8750 −8750
h 2 () = exp − exp exp − exp
365 T 365 335
and (k) = T is the premise variable.
We choose  = 1.5 and  = 0.2 in Algorithm 1. Fig. 2 shows the system state responses under nonlinear moving-
horizon H∞ control based on a T–S model (solid lines) with the zero initial condition; the corresponding disturbance
input is plotted in Fig. 1. The proposed solutions are compared with those from a linear state feedback controller termed
constrained H∞ control (dotted line) [27]. The control input u and the performance level  are plotted in Fig. 3.
From the plot of the input variables it is clear that the control actions calculated using constrained H∞ control violate
the input constraint. Thus, the closed-loop system under constrained H∞ control will be unstable. By contrast, it is
clear that moving-horizon H∞ control respects the time-domain constraints by online relaxation of the performance
level .
The computation time for this example needs to be addressed. Table 1 summarizes the computer configuration
parameters used for simulation and Fig. 4 shows the computation time for solving (25) versus sampling time. The results
88 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96

Fig. 1. Membership functions of the fuzzy CSTR model and disturbance input v(k).

Fig. 2. Responses of x1 (k) and x2 (k).

show that the computation time increases when disturbance exists in the system. Moreover, the average computation
time is much less than the sampling period.

5.2. Active queue management

The proposed moving-horizon H∞ approach was also used to design a controller for AQM routers [28]. Ignoring
the TCP timeout mechanism, a dynamic model of TCP behavior is described by the following coupled and nonlinear
 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96 89

Fig. 3. Control input u(k) and H∞ performance index (k).

Table 1
Computer configuration parameters.

Manufacturer and model Intel Core 2 Duo

Processor speed 2.5 GHz
RAM size 2.00 GB
Hard drive size 160 GB

differential equations:
⎧
⎪ w(t − R(t)) w(t)w(t − R(t))
⎪
⎨ ẇ(t) = (1 − p(t − R(t))) − p(t − R(t)),
w(t)R(t − R(t)) 2R(t − R(t))
(34)
⎪
⎪ w(t)
⎩ q̇(t) = −C(t) + N ,
R(t)

where w denotes the average TCP window size (packets), q the average queue length (packets), C the link capacity
(packets/s), R(t) = q(t)/C(t) + T p the transmission RTT (s), T p the propagation delay (s), N the number of TCP
sessions and p the drop probability of a packet.
The queue length, window size and drop probability are positive and bounded quantities, q ∈ [0, qmax ], w ∈
[0, wmax ], and q ∈ [0, 1], where qmax and wmax denote the buffer capacity and maximum window size, respectively.
Probability p belongs to the interval [0, 1]. In practical networks, the available link capacity changes with time and is
difficult to measure. Here, we suppose that the nominal value, say C0 , of C(t) is known, while
C(t) = C(t) − C0 is
unknown and considered as a disturbance to the system [29].
Take (w, q) as the state, p as the input and q as the output. For a given triplet of network parameters (N , C0 , T p ),
any operating point (w∗ , q∗ , p∗ ) is defined by ẇ = 0 and q̇ = 0, so that R∗ = q∗ /C0 + T p , w∗ = R∗ C0 /N , q∗ =
2N 2 /(2N 2 + C02 R∗2 ).
90 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96

Fig. 4. Computation time for Algorithm 1 versus sampling time.

The equilibrium of model (34) is defined as (w0 , q0 , p0 ) and then we set

q = q − q0 ,
p = p − p0 ,
w = w − w0
and
C = C − C0 . System (34) has the form
˙ = F(x̃,
C, u) = f (x̃,
C) + g(x̃,
C)ũ(t − R0 ),
x̃(t) (35)
where
       
x̃1 (t)
w(t) f 1 (x̃,
C) g1 (x̃,
C)
x̃(t) = = , f (x̃,
C) = , g(x,
C) = ,
x̃2 (t)
q(t) f 2 (x̃,
C) g2 (x̃,
C)

(x̃1 (t − R0 ) + w0 )(2 − p0 (2 + (x̃1 (t) + w0 )2 ))

f 1 (x̃,
C) =   ,
x̃2 (t − R0 ) + q0
2(x̃1 (t) + w0 ) + Tp

C(t − R0 ) + C0
x̃1 (t) + w0
f 2 (x̃,
C) = −(
C(t) + C0 ) + n  ,
x̃2 (t) + q0
+ Tp

C(t) + C0
(x̃1 (t − R0 ) + w0 )(2 + (x̃1 + w0 )2 )
g1 (x̃,
C) = −  ,
x̃2 (t − R0 ) + q0
2(x̃1 + w0 ) + Tp

C(t − R0 ) + C0
g2 (x̃,
C) = 0, ũ(t − R0 ) =
p(t − R0 ).
Obviously, the equilibrium point of system (35) is (x̃0 = 0, ũ 0 = 0). Ignoring the dependence of the time-delay
argument t − R(t) on queue length, the result of Taylor’s linearization of model (35) about the equilibrium (x̃0 , ũ 0 ) is
obtained as [30]
 −2C n     
0
0 2n 2 +R02 C02 0
˙ = 2 +C 2 R 2 −
x̃(t) 2n 0 0 x̃(t) + 2
2n R0 ũ(t − R0 ) + T
C(t). (36)
n
R0 − R10 0 − Rp0
About an operating non-equilibrium point (x̃s , ũ s ), the linearized model [31] has the form
     
T 2n 2 +Rs2 C02 0
˙x̃(t) = a1 x̃(t) + − 2n 2 Rs ũ(t − R0 ) +
C(t), (37)
T
a2T 0 − Rps
 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96 91

where
f i (x̃s , 0) − x̃sT ∇ f i (x̃s , 0)
ai = ∇ f i (x̃s , 0) + x̃s , x̃s  0, i = 1, 2. (38)
x̃s 2
Here, the change for the input delay is ignored. However, models (36) and (37) with an input time delay should be
approximated by a common linear model. A simple method is to approximate the time delay with rational terms
using an all-pole approximation, i.e., the transfer function of time delay e−s ≈ 1/(1 + s). Then model (36) is
approximated by
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤ ⎡ ⎤

ẇ(t) 0 1 0
w(t) 0 0
⎢ ⎥ ⎢ a a R0 −1 ⎥⎢ ⎥ ⎢ b ⎥ ⎢ ⎥
⎣
ẅ(t) ⎦ = ⎢⎣ R0 R0 0 ⎥ ⎢ 0 ⎥
C(t),
⎦ ⎣
ẇ(t) ⎦ + ⎣ R0 ⎦ ũ(t) + ⎣ ⎦ (39)
Tp

q̇(t) n
R0 0 − R0
1
q(t) 0 − R0

where a = −2C0 n/(2n 2 + C02 R02 ) and b = −(2n 2 + R02 C02 )/2n 2 R0 ; model (37) can be approximated by the same
way. According to the above linear local dynamic models, T–S fuzzy models can be immediately constructed. We
choose q as the premise variable and use three fuzzy rules to construct the T–S fuzzy models.
Here, the parameters are N = 120, C0 = 3750 packets/s, T p = 0.05 s, qmax = 600 packets, wmax = 20. The
maximum link bandwidth disturbance is approximately 6.4% of its nominal value. The control objective is to minimize
H∞ performance from the disturbance
C(t) to the queue length q(t) around the steady state q0 = 300 packets and
adapt the performance online to satisfy time-domain constraints.
Here, the equilibrium point of model (34) for the state transformation is

q0 = 300, w0 = 4.0625, p0 = 0.1081, R0 = 0.13.

Then the other two operating points can be obtained as follows:

q + = 450, w + = 5.3125, p + = 0.0012, R + = 0.17;

q − = 150, w − = 2.8125, p − = 0.2018, R − = 0.09.

Three coordinate linear subsystems are obtained and reduced to dimensionless forms using the notation x1 (t) =
(w(t) − w0 )/(wmax − w0 ), x2 (t) = ẇ(t)/(wmax − w0 ), x3 (t) = (q(t) − q0 )/(qmax − q0 ), control input u(t)= ( p(t) −
p0 /(1 − p0 ) and external disturbance v(t) = (C(t) − C0 )/0.064C0 . Finally, we discretize these using a sampling
period Ts = 0.15 s. We define the membership functions for state variable x3 (k) as in Fig. 5.
Then the nonlinear TCP network with AQM routers can be obtained by blending the following three linear subsys-
tems.
R1 : IF the queue length x3 (k) is M11
THEN x(k + 1) = A1 x(k) + B11 v(k) + B21 u(k)
(40a)
z 1 (k) = C11 x(k) + D111 v(k) + D121 u(k),
z 2 (k) = C21 x(k) + D221 u(k),
R2 : IF the queue length x3 (k) is M21
THEN x(k + 1) = A2 x(k) + B12 v(k) + B22 u(k)
(40b)
z 1 (k) = C12 x(k) + D112 v(k) + D122 u(k),
z 2 (k) = C22 x(k) + D222 u(k),
R3 : IF the queue length x3 (k) is M31
THEN x(k + 1) = A3 x(k) + B13 v(k) + B23 u(k)
(40c)
z 1 (k) = C13 x(k) + D113 v(k) + D123 u(k),
z 2 (k) = C23 x(k) + D223 u(k),
92 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96

M31 1 M11
M21 M21

-0.5 0 0.5
x3 (k)

Fig. 5. Membership functions of x3 (k).

with
⎛ ⎞ ⎛ ⎞
0.8526 0.0925 0 0.7392 0.0682 0
⎜ ⎟ ⎜ ⎟
A1 = ⎝ −1.1255 0.1170 0 ⎠ , A2 = ⎝ −1.7715 −0.0156 0 ⎠,
4.0966 0.3033 0.3084 4.3123 0.2868 0.2147
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
0.5129 0.0364 0 0 −6.3360
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
A3 = ⎝ −2.5511 −0.1212 0 ⎠ , B11 = ⎝ 0 ⎠ , B21 = ⎝ −48.3767 ⎠ ,
3.9485 0.2172 0.1084 −0.0346 −13.4356
⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞
0 −5.4957 0 −4.2526
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
B12 = ⎝ 0 ⎠ , B22 = ⎝ −43.7970 ⎠ , B13 = ⎝ 0 ⎠ , B23 = ⎝ −22.2706 ⎠ .
−0.0393 −14.6205 −0.0446 −15.1520
For design of a controller for AQM routers, we choose x3 as the performance output and hence

C1i = (0 0 1), D11i = 0, D12i = 0, i = 1, 2, 3.

Moreover, the constrained outputs are [x1 x3 u]T and hence

⎛ ⎞ ⎛ ⎞
1 0 0 0
⎜ ⎟ ⎜ ⎟
C2,i = ⎝ 0 0 1 ⎠ , D22i = ⎝ 0 ⎠ , i = 1, 2, 3.
0 0 0 1
For T–S fuzzy model (40), the T–S fuzzy controller can be designed by applying Theorem 2. Thus, the T–S fuzzy
controller has the form:
C1 : IF the queue length x3 (k) is M11
THEN u(k) = K 1 x(k), (41a)
C2 : IF the queue length x3 (k) is M21
THEN u(k) = K 2 x(k), (41b)
3
C : IF the queue length x3 (k) is M31
THEN u(k) = K 3 x(k). (41c)
The corresponding closed-loop system maintains consistence with (6).
We choose  = 5.5 and  = 1 in Algorithm 1. Figs. 6–8 show the system state responses and control input under the
proposed moving-horizon H∞ control (solid line) for the initial condition x(0) = [0.308 0 0.5]T ; the corresponding
disturbance input is plotted in Fig. 9. The proposed solutions are compared with those given by the controller proposed
by Chen and Scherer [15] (dotted line).
 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96 93

460
440
420

queue length (packets)

400
380
360
340
320
300
280
260
0 1 2 3 4 5 6 7
time (s)

Fig. 6. Response of the queue length.

5.4
5.2
5
window size (packets)

4.8
4.6
4.4
4.2
4
3.8
3.6
3.4
0 1 2 3 4 5 6 7
time (s)

Fig. 7. Response of the TCP window size.

The states for the proposed controller converge to the equilibrium point more rapidly if the states are far away from
the equilibrium point and are disturbed. When the system states are close to the equilibrium point, the performance
of the two control approaches is similar, although the Chen and Scherer control (dotted line) causes slight oscillation.
The average drop probability of 10.96% calculated for moving-horizon H∞ control is less than the 11.83% calculated
for the Chen and Scherer approach. The same computer parameters as for the CSTR case were used. The computation
time versus sampling time is plotted in Fig. 10. It is clear to see that the influence of disturbance on the computation
time still exists. Compared to the CSTR example, the average computation time for AQM control is greater, which
can be attributed to the number of rules for the controller. Moreover, the average computation time is greater than the
sampling time.

Remark 5.1. The above two examples indicate that the computational effort for (25) with LMI constraints is strongly
related to the number of controller rules. The computation time for Algorithm 1 is longer than the sampling time in
the second example. Thus, it is clear that there is still much work to be done for systems with fast dynamics. One
possibility for speeding up the online calculation is to implement the proposed controller on a field-programmable gate
array (FPGA) chip [32]. A second possibility for fast computation is to develop a fast optimization algorithm [33].

6. Conclusions

For nonlinear systems with time-domain constraints, we proposed a moving-horizon H∞ control algorithm based
on T–S fuzzy models. The nonlinear dynamics are represented by local models according to fuzzy rule, and the local
94 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96

0.22

0.2

0.18

drop probability
0.16

0.14

0.12

0.1

0.08

0.06
0 1 2 3 4 5 6 7
time(s)

Fig. 8. The drop probability used as control input.

250
200
150
C(t)−C0 (packets/s)

100
50
0
−50
−100
−150
−200
0 1 2 3 4 5 6 7 8
time (s)

Fig. 9. The disturbance profile.

14

13

12

11
computation time (s)

10

4
0 1 2 3 4 5 6 7
time (s)

Fig. 10. The computation time for Algorithm 1 versus sampling time.
 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96 95

models are blended into an overall single model through fuzzy membership functions. A convex optimization problem
involving LMIs is repeatedly solved online to guarantee disturbance attenuation. Similar to the moving-horizon H∞
control for a constrained linear system [15], the proposed algorithm can trade off system performance against constraint
satisfaction in an adaptive manner. Two application examples were presented to show the merits and design procedures
of the proposed controller.

Acknowledgment

This work were supported by the National Science Fund of China for Distinguished Young Scholars under Grant
(No. 60725311), the Program for Changjiang Scholars and Innovative Research Team in University (No. IRT1017) and
the 973 Program (No. 2012CB821202).

References

[1] T. Takagi, M. Sugeno, Fuzzy identification of systems and its applications to modeling and control, IEEE Trans. Syst. Man Cybern. 15 (1985)
116–132.
[2] K. Tanaka, H.O. Wang, Fuzzy Control Systems Design and Analysis: A Linear Matrix Inequality Approach, 1st ed., John Wiley & Sons,
New York, 2002.
[3] D. Lee, J. Park, Y. Joo, A new fuzzy Lyapunov function for relaxed, stability condition of continuous-time Takagi–Sugeno fuzzy systems, IEEE
Trans. Fuzzy Syst. 19 (2011) 785–791.
[4] B. Rhee, S. Won, A new fuzzy Lyapunov function approach for a Takagi–Sugeno fuzzy control system design, Fuzzy Sets Syst. 157 (2006)
1211–1228.
[5] M. Bernal, T.M. Guerra, Generalized non-quadratic stability of continuous-time Takagi–Sugeno models, IEEE Trans. Fuzzy Syst. 18 (2010)
815–822.
[6] C. Fang, Y. Liu, S.W. Kau, A new LMI-based approach to relaxed quadratic stabilization of T–S fuzzy control systems, IEEE Trans. Fuzzy
Syst. 14 (2006) 386–397.
[7] A. Kruszewski, T. Sala, T. Guerra, C. Ariño, A triangulation approach to asymptotically exact conditions for fuzzy summations, IEEE Trans.
Fuzzy Syst. 17 (2009) 985–994.
[8] A. Sala, C. Ariño, Asymptotically necessary and sufficient conditions for stability and performance in fuzzy control: applications of Polya’s
theorem, Fuzzy Sets Syst. 158 (2007) 2671–2686.
[9] A. Kruszewski, R. Wang, T. Guerra, Non-quadratic stabilization conditions for a class of uncertain non linear discrete-time T–S fuzzy models:
a new approach, IEEE Trans. Autom. Control 53 (2008) 606–611.
[10] J.B. Qiu, G. Feng, J. Yang, A new design of delay-dependent robust H∞ filtering for discrete-time T–S fuzzy systems with time-varying delay,
IEEE Trans. Fuzzy Syst. 17 (2009) 1044–1058.
[11] D.Q. Mayne, J.B. Rawlings, C.V. Rao, P.O.M. Scokaert, Constrained model predictive control: stability and optimality, Automatica 36 (2000)
789–814.
[12] T.J. Zhang, G. Feng, X.J. Zeng, Output tracking of constrained nonlinear processes with offset-free input-to-state stable fuzzy predictive control,
Automatica 45 (2009) 900–909.
[13] M. Khairy, A.L. Elshafei, H.M. Emara, LMI based design of constrained fuzzy predictive control, Fuzzy Sets Syst. 161 (2010) 893–918.
[14] T.J. Zhang, G. Feng, J.H. Lu, Fuzzy constrained min–max model predictive control based on piecewise Lyapunov functions, IEEE Trans. Fuzzy
Syst. 15 (2007) 686–698.
[15] H. Chen, C.W. Scherer, Moving horizon H∞ control with performance adaptation for constrained linear systems, Automatica 42 (2006)
1033–1040.
[16] H.O. Wang, K. Tanaka, M. Grifin, Parallel distributed compensation of nonlinear systems by Tanaka–Sugeno fuzzy model, in: Proceedings of
FUZZ-IEEE/IFES’95, 1995, pp. 531–538.
[17] X. Gao, Control for T–S Fuzzy Systems Based on LMI Optimization, Dissertation, Jilin University, 2006 (in Chinese).
[18] W. Lin, C.I. Byrnes, H∞ control of discrete time nonlinear systems, IEEE Trans. Autom. Control 4 (1996) 494–510.
[19] S. Boyd, L. El Ghaoui, E. Feron, V. Balakrishnan, Linear Matrix Inequalities in System and Control Theory, SIAM, Philadelphia, 1994.
[20] H. Chen, C.W. Scherer, Disturbance attenuation with actuator constraints by moving horizon H∞ control, in: Proceedings of IFAC ADCHEM’03,
Hong Kong, China, 2003, pp. 447–452.
[21] H. Chen, X.-Q. Gao, H. Wang, An improved moving horizon H∞ control scheme through Lagrange duality, Int. J. Control 79 (2006) 239–248.
[22] B. Kouvaritakis, J.A. Rossiter, J. Schuurmans, Efficient robust predictive control, IEEE Trans. Autom. Control 45 (2000) 1545–1549.
[23] M.V. Kothare, V. Balakrishnan, M. Morari, Robust constrained model predictive control using linear matrix inequalities, Automatica 32 (1996)
1361–1379.
[24] L. Magni, G.D. Nicolao, R. Scattolini, A stabilizing model-based predictive control algorithm for nonlinear systems, Automatica 37 (2001)
1351–1362.
[25] D. Limon, J.B. Brovo, T. Alamo, E.F. Camacho, Robust MPC of constrained nonlinear systems based on interval arithmetic, IEEE Proc. Control
Theory Appl. 152 (2005) 325–332.
[26] Y.Y. Cao, P.M. Frank, Analysis and synthesis of nonlinear time-delay systems via fuzzy control approach, IEEE Trans. Fuzzy Syst. 8 (2000)
200–211.
96 P. Wang et al. / Fuzzy Sets and Systems 212 (2013) 78 – 96

[27] H. Chen, K.-H. Guo, Constrained H∞ control of active suspensions: an LMI approach, IEEE Trans. Control Syst. Technol. 13 (2005)
412–421.
[28] V. Misra, W.B. Gong, D. Towsley, Fluid-based analysis of a network of AQM routers supporting TCP flows with an application to RED, in:
Proceedings of ACM/SIGCOMM, 2000.
[29] F. Zheng, J. Nelson, An H∞ approach to the controller design of AQM routers supporting TCP flows, Automatica 45 (2009) 757–763.
[30] K.B. Kim, Design of feedback controls supporting TCP based on the state-space approach, IEEE Trans. Autom. Control 7 (2006) 1086–1099.
[31] M.C.M. Teixeira, S.H. Zak, Stabilizing controller design for uncertain nonlinear systems using fuzzy models, IEEE Trans. Fuzzy Syst. 7 (1999)
133–142.
[32] H. Chen, F. Xu, Y. Xi, FPGA/SOPC based implementation of model predictive controller, Control Theory Appl. http://dx.doi.org/10.1049/
iet-cta.2010.0443, in press.
[33] Y. Wang, S. Boyd, Fast model predictive control using online optimization, IEEE Trans. Control Syst. Technol. 18 (2010) 267–278.