Asian Journal of Control, Vol. 19, No. 6, pp.

1977–1990, November 2017

Published online 20 March 2017 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/asjc.1494

SYNTHESIS OF MIXED OBJECTIVE OUTPUT FEEDBACK ROBUST

MODEL PREDICTIVE CONTROL
Wei Jiang, Hongli Wang, Jinghui Lu, Weiwei Qin, and Guangbin Cai

ABSTRACT
Aiming at the constrained polytopic uncertain system with energy-bounded disturbance and unmeasurable states, a novel
synthesis scheme to design the output feedback robust model predictive control(MPC)is put forward by using mixed H2/H∞
design approach. The proposed scheme involves an offline design of a robust state observer using linear matrix
inequalities(LMIs)and an online output feedback robust MPC algorithm using the estimated states in which the desired mixed
objective robust output feedback controllers are cast into efficiently tractable LMI-based convex optimization problems. In ad-
dition, the closed-loop stability and the recursive feasibility of the proposed robust MPC are guaranteed through an appropriate
reformulation of the estimation error bound (EEB). A numerical example subject to input constraints illustrates the effective-
ness of the proposed controller.

Key Words: Model predictive control, output feedback, observer, estimation error bound, linear matrix inequality.

I. INTRODUCTION and synthesis are thus simplified based on the linear vertex
sub-models, while the implementation of the controller
Model predictive control (MPC), also known as reced- clearly depends on the weight functions accounting for the
ing horizon control (RHC), has developed from a kind of nonlinearity. In [5], Kothare et al. firstly presented an
heuristic control algorithm applied in industry into standard LMI–based approach for robust MPC synthesis that
control techniques with rich theoretical and practical con- allowed explicit incorporation of plant uncertainty, laying
tents over the last 30 years [1,2]. The reason for its success theoretical foundations for further researches. However,
has mainly depended on its ability to explicitly deal with such an approach might be conservative because it adopted
control problems for systems subject to physical constraints a constant Lyapunov function matrix and computationally
and improvement on the control performance via receding- expensive for its real-time iterative online optimization pro-
horizon optimization, and so on [3,4]. In practical applica- cedures. Thus, primary algorithms in [5] have been ex-
tions, it is always difficult to obtain the exact models, tended into linear parameter varying (LPV) systems and
especially in industrial processes. The study of robust improved by utilizing such ad hoc strategies as time-varying
MPC algorithms taking account of model uncertainties terminal constraint sets, parameter-dependent Lyapunov
and/or external disturbances together with satisfying certain functions, polyhedral invariant sets, tube invariant sets, ag-
control performances has been a hot topic (see [1–4] and gregation optimization scheme, and offline design/online
references therein). As one of the model uncertainty synthesis (see [6–14] and references therein).
descriptions, the polytopic system model has played an im- Since the system states are usually not fully available in
portant role in the uncertainty modeling for both linear practice, it is necessary to extend mature state feedback
time-invariant (LTI) and linear time-varying (LTV) systems. methods into output feedback methods to cope with wider
Through the polytopic presentation, the controller design ranges of physical systems. However, more effort has been
put into state feedback robust MPC (for example, see the
above cited works). On the one hand, the existing related
Manuscript received May 8, 2016; revised December 8, 2016; accepted January 11, output feedback robust MPC (OFRMPC) synthesis schemes
2017. can be classified into two categories, that is, the usual output
The authors are with Xi’an Institute of High-Tech, Xi’an 710025, People’s Repub-
lic of China
feedback based on the state estimator, and the dynamic out-
Wei Jiang is the corresponding author (email [email protected]) put feedback based on both the controller states and the out-
This research is partially supported by the Young Scientists Fund of the National put. On the other hand, according to the model descriptions,
Natural Science Foundation of China (Grant Nos. 61203007, 61304239,
61503392), and the Natural Science Foundation of Shaanxi Province, China (Grant the existing OFRMPC synthesis approaches can also be
No. 2015JQ6213). Our acknowledgement also goes to Professor Caisheng Jiang for classified into another three categories, that is, that with
his efficient refinement of this article and to Dr Dewei Li for his professional support.
The authors declare that there is no conflict of interests regarding the publication of
bounded disturbance, that with model parametric uncer-
this paper. tainty, and that with both. Without model parametric

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd
1978 Asian Journal of Control, Vol. 19, No. 6, pp. 1977–1990, November 2017

uncertainty, namely nominal model, OFRMPC has been of uncertain systems will be described via a state-space
successful [15]. Wan et al. in [16] have proposed an offline model with real time-varying parameter uncertainties and
OFRMPC scheme in which the state estimator is separately unknown energy-bounded disturbances. In essence, this
designed offline. However, the proposed controller in [16] paper generalizes the procedures in [24,25] to OFRMPC
has been proved to be much dependent on a cut-and-try for a discrete-time uncertain system described in polytopic
procedure and closed-loop stability is difficult to guaran- type with bounded disturbances and measurement noise,
tee. In [17], Huang et al. have presented a dynamic preserving the recursive feasibility and closed-loop
OFRMPC scheme for LPV systems based on a quasi- stability properties as in [15].
min-max algorithm. However, both the system and The remainder of this paper is organized as follows.
measurement disturbance are not considered, which often Section II presents the problem formulation and the control
results in performance deterioration or even loss of stabil- objectives. Section III states some preliminary lemmas, def-
ity in the case of small actual disturbances. Ding et al. in initions as well as a compact corollary related to the offline
[18–21] have addressed both online and offline approaches state observer. Section IV elaborates respectively several
to the dynamic OFRMPC for systems with both polytopic sufficient conditions and additional constraints guarantee-
uncertainty and bounded disturbance through several novel ing the mixed objectives, the recursive feasibility, and the
auxiliary optimization strategies, in which both the robust stability. Section V presents simulation results and
conservatism and the online computational burden can be analyses, with section VI drawing a conclusion.
reduced. Park et al. in [22] have extended the quasi-
min-max robust MPC algorithm in [6] based on state esti- Notations. The superscript ‘T’ stands for matrix transpo-
mator. Its main procedure includes an off-line design of a sition. ℝn represents the n-dimensional Euclidean space
robust state observer for LPV systems using LMIs and and ℝm × n denotes the set of all m × n real matrixes. I
an online robust OFRMPC algorithm using the estimated and 0 denote the identity matrix and zero matrix with
states. However, as shown in [23], the work in [22] is not compatible dimensions. diagf⋯g stands for a block-
able to guarantee the recursive feasibility of the optimiza- diagonal matrix. The symbol ‘*’ induces a symmetric
tion problem and the consequent closed-loop stability since
the separation principle does not hold for nonlinear sys-
·
structure in the matrix inequalities. k k refers to the
Euclidean vector norm or spectral matrix norm (e.g.,
tems, which have been further addressed and modified in ‖ x ‖P denotes xT Px for P > 0). εM : = {υ : υTMυ ≤ 1} de-
[15,23]. Though more freedom for optimization should notes the ellipsoid associated with the symmetric
be introduced by applying a dynamic output feedback con- positive-definite matrix M. The matrix inequality A > 0
troller, it will inevitably produce an extra computational means that A is positive definite. Matrixes are assumed
burden. Since introducing the state estimator is essential, to be compatible for algebraic operations if their dimen-
this paper tackles this topic for OFRMPC. sions are not explicitly stated.
Moreover, most of the conclusions in previous stud-
ies focus on closed-loop stability while providing little
control over dynamical behavior. In most practical cases,
a good controller should also deliver sufficiently fast and II. PRELIMINARIES AND PROBLEM
well-damped time response besides stability. Many classi- FORMULATION
cal control objectives, such as robust stabilization of uncer-
tain systems, or shaping of the open-loop response, can be Considering the following class of constrained
expressed in terms of the H∞ performance index and tack- polytopic discrete-time uncertain systems with bounded
led by H∞ synthesis techniques. Aside from the H∞ perfor- disturbances:
mance index, the H2 performance index has been

considered effective to deal with stochastic aspects, such xðk þ 1Þ ¼ Aðk Þxðk Þ þ Bðk Þuðk Þ þ Dðk Þwðk Þ
;
as measurement noise and random disturbances. One po- zðk Þ ¼ C ðk Þxðk Þ þ Eðk Þwðk Þ
tential way of simultaneously tackling H∞/H2 performance (1)
indexes is to design a mixed objective controller. To the
best of our knowledge, there only a few papers deal with where x(k) ∈ ℝn, u(k) ∈ ℝm, z(k) ∈ ℝn and w(k) ∈ ℝn are
mixed objective OFRMPC schemes for discrete-time un- the unmeasurable state, input, measured output and
certain systems, and the mixed objective OFRMPC scheme persistent unknown energy-bounded disturbances.
handling model parametric uncertainty, measurement A(k) ∈ ℝn × n and B(k) ∈ ℝn × m are the system matrix
noise, and energy-bounded disturbances simultaneously and input matrix. Assume that the pair (A(k), B(k)) is
still remains interesting and challenging [24,25], which stabilizable and (A(k), B(k), C(k), D(k), E(k)) ∈ Ω has the
motivates the current investigation. In this paper, the class following form:

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd
 W. Jiang et al.: Synthesis of Mixed Objective Output Feedback RMPC 1979

( )
  X
L   XL
Ω :¼ Co Aj ; Bj ; C j ; Dj ; Ej ¼ ½A; B; C; D; Ej½A; B; C; D; E ¼ λj Aj ; Bj ; C j ; Dj ; Ej ; λj ¼ 1; λj ∈½0; 1 ;
j¼1 j¼1

(2)

·
where Ω denotes the polytope and Cof g denotes the convex
hull. [Aj, Bj, Cj, Dj, Ej] denotes some known matrix pairs for
Remark 1. Note that H∞ performance guarantees global
asymptotic stability and H2 performance is useful to handle
j∈½1; 2; ⋯; L . The disturbances w(k) are assumed to be stochastic aspects, such as measurement noise and random
bounded as follows: disturbances. From equation (6), it is easy to verify that the
H∞ performance directly indicates the ability of the designed
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi controller to attenuate the influence of disturbance on out-
X ∞
wT ðk Þwðk Þ ≤ ϖ; kwðk Þk ≤ 1; (3) puts for uncertain systems. Generally, the definitions of
k¼0 H2 performance criteria for LTV systems have been pro-
posed based on system response to either a unit impulse
where ϖ denotes a known constant. The above disturbances or stationary white noise in frequency domain. The above
w(k) should represent any energy-bounded disturbances H2 performance is similar to the H∞ performance under a
since the coefficient matrixes D(k) and E(k) can be scaled zero-initial condition and with stationary white noise,
when necessary. while aiming at improving the dynamical behavior under
The constraints on system (1) are given by a nonzero-initial condition in particular. Although the H∞
control has been applied to various dynamic systems and
provides more robust results than H2 control, it is sensitive
jui ðk Þj ≤ d i ; d i > 0; i ¼ 1; 2; ⋯; m; k ≥ 0; (4)
to white disturbance or noise signals and leads to control
  performance degradation when a system is subject to
H j xðk Þ ≤ f
j; max ; f j; max > 0; j ¼ 1; 2; ⋯; n; k ≥ 0; white external disturbances. Therefore, to enjoy the merits
(5) of both H∞ control and H2 control, the mixed H∞/H2 per-
formance was introduced into the OFRMPC, which resem-
bles a guaranteed cost H∞ controller. Other relevant
where H 1 ; H 2 ; ⋯; H n are given known matrixes, di and
discussions and similar corresponding algorithms can be
fj , max are known component-wise peak bounds for control
found in [28,29]. □
inputs u(k) and states x(k), respectively.
For the system in (1), consider the following
performance requirements [24,25]:
III. OFFLINE ROBUST STATE OBSERVER
(i) H∞ performance requirement: under zero-initial DESIGN
condition, the controlled output z(k) satisfies
X
∞ X
∞ In this section, the states x(k) of system (1) are
zT ðk Þzðk Þ-γ2 wT ðk Þw ðk Þ ≤ 0; (6) assumed to be unmeasurable and adopt the following state
k¼0 k¼0 observer to estimate the uncertain states of the system.
^xðk þ 1Þ ¼ Aðk Þ ^x ðk Þ þ Bðk Þ u ðk Þ
(ii) H2 performance requirement: the controlled output (8)
z(k) satisfies þ LP ðzðk Þ  C ðk Þ x^ ðk ÞÞ;
where ^xðk Þ∈ℝn denotes the estimated state of x (k), and LP
X
∞
kzk2 ¼ zT ðk Þz ðk Þ ≤ α: (7) is the observer gain to be determined. Then, the estimation
k¼0 error e(k) and the quadratic function E (e(k)) are defined as
follows:
The goal of this paper is to design an OFRMPC to
eðk Þ ¼ xðk Þ  ^xðk Þ; (9)
meet the above performance requirements simultaneously.
Inspired by the idea of optimal guaranteed cost control Eðeðk ÞÞ ¼ eT ðk ÞP e eðk Þ; (10)
(OGCC) in [26,27], we use the H2 performance as the
online optimization objective of OFRMPC and deal with a where Pe > 0 is the weighting matrix. Then, from system (1)
given H∞ performance as a constraint. and state observer (8), the error dynamics can be derived as

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd
1980 Asian Journal of Control, Vol. 19, No. 6, pp. 1977–1990, November 2017

eðk þ 1Þ ¼ xðk þ 1Þ  ^xðk þ 1Þ By applying the schur complement on (14), one can get
¼ ðAðk Þ  LP C ðk ÞÞeðk Þ (11)
2 3
þ ðDðk Þ  LP Eðk ÞÞw ðk Þ: ð1  ρÞP e  
6 7
4 0 ρI  5≥0: (15)
P e ℏðk þ iÞ P e Γðk þ iÞ Pe
Based on the concept of quadratic boundedness
[30,31], which has been widely applied in the receding
horizon control and the design of observers, the following Since (15) is affine in (A(k), B(k), C(k), D(k), E(k)), the
corollary about the observer gain LP is given directly. inequality (15) is equal to (12). Therefore, the conclusion of
Corollary 1 can be obtained. □
Corollary 1. For the system described in (1) ~ (5), if there
exists a real scalar ρ ∈ (0, 1), a positive definite symmetric
matrix Pe and a positive definite matrix Ye defined as IV. MAIN RESULTS
Ye = PeLp with compatible dimensions satisfying the
following LMI constraints This section first lists the conditions that are sufficient
and recursively feasible for the mixed objective OFRMPC
2 3 scheme with the estimated states obtained from the
ð1  ρ Þ P e   designed state observer in Section III, and then elaborates
6 7 the design procedure of OFRMPC and its asymptotic
4 0 ρI  5≥ 0; j ¼ 1; 2; ⋯; L:
stability conditions.
P e Aj  Y e C j P e Dj  Y e Ej Pe
(12)
4.1 Sufficient conditions for mixed objectives
Then the observer gain LP will be obtained by
The purpose of this paper is to construct an admissible
LP ¼ P 1
e Y e and the estimation error e(k) will converge controller satisfying the mixed H2/H∞ performance require-
into εP e and stay in it thereafter, namely the estimated
ments. The future control sequence can be calculated as
states ^
xðk Þ will asymptotically converge to the system
follows:
states x(k) as k → ∞.
uðk þ ijk Þ ¼ F ðk Þ^xðk þ ijk Þ; (16)
Proof. According to the quadratic boundedness, the qua-
dratic function E(e(k)) in (10) is monotonously decreasing, where ^xðk þ ijk Þ and F(k) respectively denote the estimated
namely E(e(k + 1)) ≤ E(e(k)), if and only if the error dynami- states of x(k + i|k) and the feedback gain determined at step
cal system in (11) is quadratically bounded with Lyapunov k. Based on (8) and (11), it is natural that the prediction
matrix Pe and E(e(k)) ≥ 1. Define ℏ(k + i) := A(k + i)  LpC(k model of the augmented state is
+ i) and Γ(k + i) := (D(k + i)  LPE(k + i)). For the distur-
bances ‖w(k)‖ ≤ 1 (equal to w(k) ∈ εI), E(e (k + 1)) ≤ E(e(k)) " # " #" #
holds if and only if there exists a real scalar ρ ∈ (0, 1) such that ^xðk þ i þ 1Þ Φðk þ iÞ LP C ðk þ iÞ ^xðk þ iÞ
¼
eðk þ i þ 1Þ 0 ℏðk þ iÞ eðk þ iÞ
" # " #
ℏ ðk þ iÞP e ℏðk þ iÞ  P e þ ρP e
T
 L P E ðk þ i Þ
≤ 0: þ wðk þ iÞ;
ΓT ðk þ iÞP e ℏðk þ iÞ ΓT ðk þ iÞP e Γðk þ iÞ  ρI
Γðk þ iÞ
(13)
(17)

Further, the above inequality can be rewritten as

where

" # Φðk þ iÞ ≜ Aðk þ iÞ þ Bðk þ iÞF ðk Þ: (18)

ð1  ρÞP e 
0 ρI Let us define a quadratic functions
" #
ℏ ðk þ iÞP e
T
^ ði; k Þ ¼ k^xðk þ ijk Þk ; P ðk Þ > 0
 P 1 V (19)
e ½P e ℏðk þ iÞ P e Γðk þ iÞ ≥ 0: P ðk Þ
Γ ðk þ iÞP e
T

(14) for i ≥ 1, one can get

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd
 W. Jiang et al.: Synthesis of Mixed Objective Output Feedback RMPC 1981

^ ði þ 1; k Þ  V
^ ði; k Þ ¼ kΦðk þ iÞ^xðk þ ijk Þ þ LP C ðk þ iÞeðk þ ijk Þ þ LP Eðk þ iÞwðk þ iÞk
V P ðk Þ  k^
xðk þ ijk ÞkPðk Þ
2 3T 2 3
^xðk þ ijk Þ ^xðk þ ijk Þ
6 7 6 7
¼6 7 6 7 2 T
4 eðk þ ijk Þ 5 M ðk þ iÞ4 eðk þ ijk Þ 5 þ γ w ðk þ iÞwðk þ iÞ  z ðk þ ijk Þzðk þ ijk Þ;
T

w ðk þ i Þ w ðk þ i Þ
(20)

where 2 3 2 3 2 3T
P ðk Þ   ΦT ðk þ iÞ ΦT ðk þ iÞ
6 7 6 T 7 6 T 7
M 11 ðk þ iÞ ¼ ΦT ðk þ iÞP ðk ÞΦðk þ iÞ-P ðk Þ M ðk þ iÞ ¼ 6
4 0 0  7 6 T 7 6 T 7
5 þ 4 C ðk þ iÞLP 5P ðk Þ4 C ðk þ iÞLP 5
(21)
þ C T ðk þ iÞC ðk þ iÞ; 0 0 γ2 I ET ðk þ iÞLTP ET ðk þ iÞLTP

M 22 ðk þ iÞ ¼ C T ðk þ iÞLTP P ðk ÞLP C ðk þ iÞ þ ½ C ðk þ iÞ C ðk þ iÞ Eðk þ iÞ T ½ C ðk þ iÞ C ðk þ iÞ Eðk þ iÞ :

(22)
þ C T ðk þ iÞC ðk þ iÞ; (27)

For the robust stability, it is necessary to have

M 33 ðk þ iÞ ¼ ET ðk þ iÞLTP P ðk ÞLP Eðk þ iÞ
(23) ^xð∞jk Þ ¼ 0 and V ð^xð∞jk ÞÞ ¼ 0. Summing (20) from i = 0
 γ2 I þ ET ðk þ iÞEðk þ iÞ; to i = ∞, one can get

∞ n
X o
^
xT ðkjk ÞP ðk Þ^
xðkjk Þ ¼ zT ðk þ ijk Þzðk þ ijk Þ þ γ2 wT ðk þ iÞwðk þ iÞ þ k½^xðk þ ijk Þ eðk þ ijk Þ wðk þ iÞkM ðkþiÞ :
i¼0
(28)

The above equation can be transformed to

X
∞   ∞ n
X o
zT ðk þ ijk Þzðk þ ijk Þ ¼ ^xT ðkjk ÞP ðk Þ^xðkjk Þ þ γ2 wT ðk þ iÞwðk þ iÞ þ k½^
xðk þ ijk Þ eðk þ ijk Þ wðk þ iÞkM ðkþiÞ :
i¼0 i¼0
(29)

M 21 ðk þ iÞ ¼ C T ðk þ iÞLTP P ðk ÞΦðk þ iÞ From the performance requirements in (6) and (7), it is

(24)
þ C ðk þ iÞC ðk þ iÞ;
T clear that

(i) When ^xðkjk Þ ¼ 0, the goal (6) is equal to

M 31 ðk þ iÞ ¼ ET ðk þ iÞLTP P ðk ÞΦðk þ iÞ
(25)
þ ET ðk þ iÞC ðk þ iÞ; X
∞  
zT ðk þ ijk Þzðk þ ijk Þ  γ2 wT ðk þ iÞwðk þ iÞ ≤ 0:
i¼0
(30)
M 32 ðk þ iÞ ¼ ET ðk þ iÞLTP P ðk ÞLP C ðk þ iÞ
(26)
þ ET ðk þ iÞC ðk þ iÞ:
It can be directly derived from (28) that if M(k + i) ≤ 0
Through proper matrix factorization, M(k + i) can be holds, (30) will be satisfied. So the H∞ performance require-
rewritten as ment will be guaranteed.

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd
1982 Asian Journal of Control, Vol. 19, No. 6, pp. 1977–1990, November 2017

(ii) From (29), it is obvious that if M(k + i) ≤ 0, the H2 that the LMIs in (32) and (35) holds, the performance re-
performance requirement will also be satisfied when quirements in (6) and (7) can be satisfied.
the following inequality holds
X
∞   Proof. According to the above analysis, theorem 1 can be
zT ðk þ ijk Þzðk þ ijk Þ ≤ x^T ðk ÞP ðk Þ^xðk Þ þ γ2 ϖ 2 ≤ α: obtained intuitively. Thus the detailed proof is omitted for
i¼0
brevity. □
(31)
From this analysis, it is clear that if M(k + i) ≤ 0 from
i = 0 to i = ∞ and inequality (31) holds, the performance re- 4.2 LMIs for guaranteed stability
quirements in (6) and (7) can be guaranteed. By resorting to
Schur complement, the inequality (31) and M(k + i) ≤ 0 can Since the control input of the OFRMPC is obtained by
be rewritten as follows solving an optimization problem online at each step using
2 3 the estimated states rather than the true states, additional
1   stability conditions must be imposed to guarantee closed-
6 2 2 7
4 γ ϖ αγ2 ϖ 2  5≥ 0; (32) loop stability during the receding horizon implementation
of the controller. Generally, for both finite and infinite hori-
^
xðk Þ 0 Q ðk Þ
zon robust MPC, the same optimality argument is used that
the optimal cost at step k + 1 must be less than or equal to
2 3
P ðk Þ     the feasible cost at step k. This argument is valid for the case
6 0 0   7 where the optimal cost is based on the true state or based on
6 7
6 7 the estimated state the error dynamics of which are indepen-
6 0 0 γ2 I  7
6 7≥ 0; dent of and much faster than the system dynamics [5,15].
6 7
4 Φ ðk þ i Þ LP C ðk þ iÞ LP Eðk þ iÞ P ðk Þ1 5 To begin with, let us define a quadratic function
C ðk þ i Þ C ðk þ i Þ E ðk þ i Þ 0 I
ℜ ði; k Þ ¼ k^xðk þ ijk ÞkP ðk Þ þ keðk þ ijk ÞkPe ðk Þ ;
(33)
P e ðk Þ ¼ βðk ÞP e > 0; P ðk Þ > 0
where Q(k) is defined as αP(k) 1
and F(k) is defined (36)
as Y(k)Q(k)1. Pre- and post-multiplying (33) by
diag(α1/2P(k)1, α1/2I, α1/2I, α1/2I, α1/2I), (34) will be for i ≥ 1, where β(k) > 0 is a time-varying scalar. Generally,
equally obtained as follows the argument is usually adopted as follows
2 3  
Qðk Þ     ℜ ði þ 1; k Þ  ℜ ði; k Þ ≤  k^xðk þ ijk ÞkL þ kuðk þ ijk ÞkR ;
6 0 0    7
6 7
6
6
7 ∀k; i ≥ 0; (37)
6 0 0 αγ2
I   7
7≥0:
6 7
4 Aðk þ iÞQðk Þ þ Bðk þ iÞY ðk Þ αLP C ðk þ iÞ αLP Eðk þ iÞ Qðk Þ  5
where L and R are given positive definite symmetric
C ðk þ iÞQðk Þ αC ðk þ iÞ αEðk þ iÞ 0 αI
matrixes. Substituting (17) into (37), it follows that
(34)
ℜ ði þ 1; k Þ  ℜ ði; k Þ ¼ ‖Φðk þ iÞ^xðk þ ijk Þ
Since the inequality (34) is affine in(A(k), B(k),
C(k), D(k), E(k)) according to the polytopic description þ LP C ðk þ iÞeðk þ ijk Þ
in (2), it can be concluded that (34) holds if and only þ LP Eðk þ iÞwðk þ iÞ‖P ðk Þ
if there exist matrixes Y(k) and Q(k) such that
2 3  k^xðk þ ijk ÞkP ðk Þ
Q ðk Þ    
6 0 0    7 þ ‖ℏðk þ iÞeðk þ ijk Þ (38)
6 7
6 7
6 0 0 αγ2 I   7 þ Γðk þ iÞwðk þ iÞ‖P e ðk Þ
6 7≥ 0; j ¼ 1; 2; ⋯; L:
6 7
4 Aj Qðk Þ þ Bj Y ðk Þ αLP C j αLP Ej Q ðk Þ  5
 keðk þ ijk ÞkPe ðk Þ ≤
C j Qðk Þ αC j αEj 0 αI  
(35)  k^xðk þ ijk ÞkL þ kuðk þ ijk ÞkR
¼ k^xðk þ ijk ÞkLþFT ðk ÞRFðk Þ :
Theorem 1. Let all variables, definitions and assumptions
be as above. Considering the system described in (1) ~ (2), Similar to some fairly standard manipulations (i.e.,
if there exist matrixes Y(k) ∈ ℝm × n and Q(k) ∈ ℝn × n such removing ^xðk þ ijk Þ; e(k + i|k) and w(k + i), applying the

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd
 W. Jiang et al.: Synthesis of Mixed Objective Output Feedback RMPC 1983

Schur complement, using the congruence transformations, where Wii and Vjj denote the i - th and j - th diagonal element
substituting Q(k) = αP(k)1), which is also adopted in [29], of W and V, respectively.
the constraint (38) is satisfied if and only if
2 3 From the above analyses, the desired OFRMPC with
Q ðk Þ      
6
6 0 αP e ðk Þ      7
7
mixed H2/H∞ objectives can be designed as the following
6 7 theorem.
6 0 0 0     7
6 7
6 7
6 Φðk þ iÞQðk Þ αLP C ðk þ iÞ αLP Eðk þ iÞ Qðk Þ    7≥ 0:
6 7
6 0 αℏðk þ iÞ αΓðk þ iÞ 0 αP e ðk Þ1   7 Theorem 2. Considering the system described in (1) ~ (2)
6 7
6 7
4 L1=2 Qðk Þ 0 0 0 0 αI  5 with constraints in (4) and (5), if there exist matrixes
R1=2 Y ðk Þ 0 0 0 0 0 αI Y(k) ∈ ℝm × n and Q(k) ∈ ℝn × n such that the LMIs in (32),
(39) (35), (40), (42), (43) holds when H∞ performance index γ
is given, the performance requirements in (6) and (7) can
Since (39) is affine in (A(k), B(k), C(k), D(k), E(k)) be satisfied.
according to the polytopic description, it can be concluded
that (39) holds if and only if there exist matrixes Y(k) and The formulation of the online optimization problem of
Q(k) such that the OFRMPC in Theorem 2 can be summarized as follows
2 3 min α
Qðk Þ       Q;Y ;W;V
6
0 αβðk ÞP e     
7 (44)
6 7
6
6 0 0 0    
7
7 subject to ð32Þ; ð35Þ; ð40Þ; ð42Þ; ð43Þ:
6 7
6 7
6 Aj Qð k Þ þ Bj Y ð k Þ αLP C j αLP Ej Qð k Þ    7≥ 0;
6

7 Remark 3. In terms of a closed-loop asymptotic stability of
6 7
6 0 α Aj  L P C j α Dj  L P E j 0 αðβðk ÞP e Þ1   7
6 7 the OFRMPC based on Theorem 2, the procedure is to
6 7
4 L1=2 Qðk Þ 0 0 0 0 αI  5
R1=2 Y ðk Þ 0 0 0 0 0 αI prove that ℜ ði; k Þ ¼ k^xðk þ ijk ÞkPðk Þ þ keðk þ ijk ÞkPe ðk Þ is
(40) a strictly decreasing Lyapunov function for the closed-loop,
where P(k) > 0 and Pe(k) > 0 are obtained from the optimal
where j = 1 , 2 , ⋯ , L. solution at step k. □

Remark 2. In this paper, keð1ÞkP e ≤ ηð1Þ, where η(1) is a Remark 4. Strictly speaking, the variables in the above op-
user-specified constant. At each sampling instant k ≥ 1, the timization in (44) should be denoted by Qk,Yk,Wk,Vk etc. to
bound of e(k + 1) is refreshed as emphasize that they are computed at step k. For notational
convenience, the subscript is omitted here and in the next
keðk þ 1ÞkP e ≤ ηðk þ 1Þ: (41)
section. However, this notation will be briefly used in the re-
Applying corollary 1 yields keðk þ 1ÞkP e ≤1 þ ð1  ρÞ cursive feasibility proof. □
 
keðk ÞkP e  1 : By iteration it obtains keðk ÞkPe ≤1 þ
  Theorem 3. Considering the system described in (1) ~ (2)
ð1  ρÞk keð1ÞkP e  1 . Thus an admissible value for with constraints in (4) and (5), if there are feasible solu-
η(k + 1) in (41) is chosen as 1 + (1  ρ)(η(k)  1). Thus, for tions obtained from (44) at each step k ≥ 0, where η(k + 1)
computational simplicity, β(k) in (36) and (40) is chosen as is refreshed according to remark 2, suppose that (44) is
1/(1 + (1  ρ)k(η(1)  1)) for k ≥ 1 instead of a more sophisti- feasible at sampling instant k = 0, then (i) (44) is feasible
cated one. For more general forms of β(k), see [32,33]. □ for all k ≥ 0, and (ii) the feasible solution obtained from
(44) guarantees the asymptotic stability, that is, as k → ∞,
Lemma 3. (see [24,25]). The input and states constraints in {x(k), e(k)} enters into the neighborhood {0, 0} and stays
(4) and (5) will be guaranteed if there exist symmetric ma- in it thereafter.
trix W and V with compatible dimensions satisfying
 Proof. (i) It suffices to prove that the solution to the optimi-
W 
> 0; W ii ≤ d 2i ; i ¼ 1; 2; ⋯; m; (42) zation problem (44) at sampling instant k is a
Y T ðk Þ Q ðk Þ
feasible solution to the problem at sampling
 instant k + 1. Therefore, at step k + 1, we can
V 
> 0; V jj ≤ f 2j; max ; j ¼ 1; 2; ⋯; n; construct the same solution as that at step k.
ðHQðk ÞÞT Q ðk Þ Then, it is necessary to check (32), (35),
(43) (40), (42) and (43) at sampling instant k + 1,

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd
1984 Asian Journal of Control, Vol. 19, No. 6, pp. 1977–1990, November 2017

since they include α, x^ðk Þ , which might be 1α1 γ2 ϖ 2 ≥^xT ðk ÞQðk Þ1 x^ðk Þ ≥^xT ðk þ 1ÞQðk Þ1 x^ðk þ 1Þ
changed at different steps. come into existence. Then, the constructed solution will also
First, if the optimization problem (44) is feasible at satisfy equation (32). That is, the constructed solution is fea-
sampling instant k, the optimal solution can be assumed sible at sampling instant k + 1. Second, if (42) is satisfied at
to be {Q(k), Y(k), W(k), V(k), α*(k)}. At sampling instant k sampling instant k,u (k + 1|k + 1) satisfies the input
+ 1, one can construct the same solution as that at step k. constraints. Then, with uðk þ 1jk þ 1Þ ¼ uðk þ 1jk Þ ¼
Thus, it is clear that the state constraints (43), replacing k F ðk Þ^xðk þ 1jk Þ, it is clear that the input constraints, replac-
with k + 1, are directly satisfied. Since the disturbance is ing k with k + 1, are also satisfied. Third, adopting the similar
bounded as Equation (3) and the optimal solution satisfy congruence transformation on (35) and (40) as the feasibility
equations (35) and (40) at sampling instant k, the optimal proof of Theorem 8 in [32], the feasible solution of the opti-
solution will satisfy equation (32). In addition, from mization problem (35) and (40) at time instant k is also fea-
Corollary 1, the estimated states ^xðk Þ asymptotically sible at time instant k + 1 both the constraints (35) and (40)
converge to the system real states x(k) as k → ∞. Thus, are also satisfied replacing k with k + 1. This argument can
xT ðk ÞQðk Þ1 ^xðk Þ . Since equations (35)
1  α1 γ2 ϖ 2 ≥ ^ be continued for time instant k + 2 , k + 3 , … to complete
and (40) have been satisfied by the constructed solution, the proof.
(ii) At sampling instant k + 1, according to Theorem 2,
the bound of ℜ(i, k) will decrease with the speed
of no less than k^xðk þ ijk ÞkL whenever
k^xðk þ ijk ÞkL ≠0. For sufficiently large k, ℜ(i, k) will
evolve until ^xðk Þ ¼ 0 and ℜ(i, k) = 0, which means
that f^xðk Þ; uðk Þ; eðk Þg will converge to a neighbor-
hood of {0, 0, 0} and stay inside this neighborhood
thereafter. Moreover, as k → ∞, xðk Þ ¼ ^xðk Þ þ eðk Þ
converges since both ^xðk Þ and e(k) do. In conclusion,
the closed-loop system is robustly stable. □

Remark 5. For a mixed objective OFRMPC controller, if

the H2 performance index is given in advance, the robust
predictive controller with H∞ performance index as the ob-
jective by the similar method mentioned above can also be
designed. □

Fig. 1. Angular positioning system. [Color figure can be viewed at Remark 6. Theorem 2 can be extended to several standard
wileyonlinelibrary.com] problems encountered in practice, such as reference

Table I. Comparison results for proposed OFRMPC scheme.

Case ρ Pe LP αmin Ttotal(s)
 
0:92176 0:19595 0:94541
(a) 0.99 1011 0.3854 1.899
0:19595 0:22763  0:21196 
0:72958 0:16738 0:95213
(b) 0.95 1011 0.3981 1.977
0:16738 0:17895 0:25925
 
0:12404 0:03058 0:96546
(c) 0.90 1010 0.4012 1.891
0:03058 0:02997 0:30266
 
0:43471 0:12165 0:99460
(d) 0.80 1010 0.3958 1.868
0:12165 0:10368 0:38793
 
0:13905 0:04640 1:12631
(e) 0.60 109 0.3952 1.852
0:04640 0:02976 0:72642
 
0:92176 0:19595 0:94541
(f) 0.99 1011 0.3995 1.970
0:19595 0:22763 0:21196

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd
 W. Jiang et al.: Synthesis of Mixed Objective Output Feedback RMPC 1985

V. NUMERICAL SIMULATIONS

The effectiveness of the proposed RMPC scheme is

demonstrated by the following angular positioning system
adapted in [5]. The system (see Fig. 1) consists of a rotating
antenna at the origin of the plane, driven by an electric mo-
tor. The control problem is to use the input voltage to the
motor to rotate the antenna so that it always points to the di-
rection of a moving object in the plane. The angular posi-
tions of the antenna and the moving object (θ and θr, rad)
and the angular velocity of the antenna (θ, _ rad·s1) are as-
sumed to be unmeasurable.
Fig. 2. Applied disturbance. [Color figure can be viewed at The motion of the antenna can be described through
wileyonlinelibrary.com] the following discrete-time equations obtained from their
continuous-time counterparts by discretization, using a
sampling time of 0.1 s and Euler’s first-order approximation
for the derivative.
" #
θðk þ 1Þ
x ðk þ 1 Þ ¼ ¼ Aðk Þxðk Þ
θ_ ðk þ 1Þ
þ Bðk Þuðk Þ þ Dðk Þwðk Þ
" #" #
1 0:1 θ ðk Þ (45)
¼
1  0:1ηðk Þ θ_ ðk Þ
0
#" " #
0 0
þ uðk Þ þ wðk Þ;
0:5 0:1

Fig. 3. Control input. [Color figure can be viewed at

zðk Þ ¼ Cxðk Þ þ Eðk Þwðk Þ
wileyonlinelibrary.com] 
θ ðk Þ
¼ ½ 1 0:5  ˙ þ ½0:1wðk Þ; (46)
θ ðk Þ
trajectory tracking, constant set-point tracking, disturbance where, 0.1sec1 ≤ η(k) ≤ 10sec1 and |u(k)| ≤ 5volts. Parame-
rejection, and systems with multiple delays, under appropri- ter η(k) is assumed to be arbitrarily time-varying in the
ate assumptions. Besides referring to the concept of the as- range of variation. It can be concluded that

 
ðAðk Þ; Bðk Þ; C ðk Þ; Dðk Þ; Eðk ÞÞ∈ Ω ¼ Co Aj ; Bj ; C j ; Dj ; Ej
( )
X L   (47)
¼ ½A; B; C; D; Ej½A; B; C; D; E ¼ λj Aj ; Bj ; C j ; Dj ; Ej ;
j¼1

ymptotically stable invariant ellipsoid or polyhedral where L = 2, A1 ¼ ½ 1 0:1; 0 0:99 , A2 ¼ ½ 1 0:1; 0 0 ,

invariant sets, the efficient off-line formulation of mixed ob- B1 ¼ B2 ¼ ½ 0 0:5 T , C 1 ¼ C 2 ¼ ½ 1 0:5 , D1 ¼ D2 ¼
jective OFRMPC algorithms, which can be derived directly, ½ 0 0:1 T ; E1 = E2 = 0.1 and λ1(k) = (ηmax  η(k))/ℓ(η), λ2(k)
are omitted in this paper for the sake of brevity. □ = (η(k)  ηmin)/ℓ(η) where ℓ(η) = ηmax  ηmin, are convex
P
coordinates since 2j¼1 λj ¼ 1; λj ∈ [0,1]. In order to verify
At each time instant, the optimization problem men-
tioned in (44) is solved online to determine the optimal con- the effectiveness of the mixed-objective OFRMPC
trol sequence u(k + i|k) based on the measured output z(k) proposed in Theorem 2, the initial states of the system (1)
and the state observer (8). Then, subsequently only the first and the observer (8) are assumed as xð0Þ ¼ ½ 2 1 T and
control u(k|k) in this sequence will be implemented. ^xð0Þ ¼ ½ 4 3 T , respectively. For a fair comparison, five

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd
1986 Asian Journal of Control, Vol. 19, No. 6, pp. 1977–1990, November 2017

Fig. 4. Trajectories of true states, estimated states and EEBs. [Color figure can be viewed at wileyonlinelibrary.com]

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd
 W. Jiang et al.: Synthesis of Mixed Objective Output Feedback RMPC 1987

cases with different ρ have been considered. Applying the The effectiveness and tracking performance for an an-
above procedure, let us consider another case of the pro- gular positioning system is depicted in Fig. 5, where the
posed OFRMPC, i.e. case (f) where the disturbance is zero, blue dash line is the reference command, red bold line is
namely nominal case [34]. The state observer gain LP, re- for the responding true angular position and green dash-
lated coefficient matrix Pe, minimized H2 performance in- dotted line is for the estimated angular position. It is
dex αmin and the total computational time Ttotal are observed from Fig. 5 that the estimated angular position
obtained, as in Table I, via standard LMI Toolbox in the converges to the true ones over time and the proposed
MATLAB environment [35]. mixed objective OFRMPC achieves good performance for
The rationale of this turning procedure is not limited tracking problem. Figs 6 and 7 demonstrate the other state
to this example. Then, the proposed OFRMPC scheme and the control input of angular positioning system, respec-
has been applied with the following weighting matrices as tively. Since the input constraints are taken into consider-
L = diag(1,1) and R = 1. The H∞ performance index γ is ation, thus all the control moves are inside the boundary
set as γ = 0.5 and only control input constraint is consid- values as depicted in Fig. 7.
ered. In the simulation, Fig. 2 shows the applied unknown The above two examples illustrate that the mixed
norm-bounded disturbance, which is randomly generated in objective OFRMPC controllers in this article improve con-
the interval ½ 1 1 . The responses of control input sig- trol performance effectively and are more realistic from a
nals are shown in Fig. 3 determined by solving the optimi- practical point of view. In general (not limited to these
zation problem (44) at each step. It is clear that the examples), the control performance is affected in a
obtained control input satisfies the given constraint. The
trajectories of the true states, the estimated states, with
the corresponding ellipsoids of EEBs centered at the true
states, are depicted in Fig. 4, respectively. These figures
demonstrate that the estimated states converge to the true
states, which means the proposed observer (8) works well.
Note that the EEBs, as shown in Fig. 4, may not be as tight
as they really are. Hence, the tightness and convergence of
EEB can reflect the effectiveness of the method for com-
puting EEB.
By comparing Case (a) ~ Case (e), choosing smaller
decay rateρcould drive the states convergence faster and
also has certain effects on the state observer gain LP, but
not severely in this example. From Case (a) ~ Case (f), it
is obvious that the closed-loop system is asymptotically sta-
ble with the OFRMPC controller designed from Theorem 2
Fig. 5. Tracking trajectories of true states, estimated states and
when both model uncertainties and external disturbances reference of angular position. [Color figure can be
exist simultaneously. Using the same notions mentioned viewed at wileyonlinelibrary.com]
above, we modify the control objective into tracking a
predefined reference trajectory in spite of model uncertainty
and external disturbance. The benchmark reference trajec-
tory ℵr is chosen as follows, which is a time-varying piece-
wise step signal.

8
> 1:0 ; 1 ≤ k ≤ 50
>
>
< 2:0 ; 50 < k ≤ 100
ℵ r : x1 ¼ ; (48)
>
> 3:0 ; 100 < k ≤ 150
>
:
1:5 ; 150 < k ≤ 200
8
> 0 ; 1 ≤ k ≤ 50
>
>
< 0 ; 50 < k ≤ 100
x2 ¼ :
>
> 0 ; 100 < k ≤ 150
>
: Fig. 6. Trajectories of true states, estimated states of angular speed.
0 ; 150 < k ≤ 200 [Color figure can be viewed at wileyonlinelibrary.com]

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd
1988 Asian Journal of Control, Vol. 19, No. 6, pp. 1977–1990, November 2017

stability and optimality,” Automatica, Vol. 36, No. 6,

pp. 789–814 (2000).
2. Mayne, D. Q., “Model predictive control: recent devel-
opments and future promise,” Automatica, Vol. 50, No.
12, pp. 2967–2986 (2014).
3. Xi, Y. G., and D. W. Li, “Fundamental philosophy and
status of qualitative synthesis of model predictive con-
trol,” Acta Automatica Sinica, Vol. 34, No. 10, pp.
1225–1234 (2008).
4. Xi, Y. G., D. W. Li, and S. Lin, “Model predictive
control—status and challenges,” Acta Automatica
Sinica, Vol. 39, No. 3, pp. 222–236 (2013).
5. Kothare, M. V., V. Balakrishnan, and M. Morari,
Fig. 7. Control input. [Color figure can be viewed at
“Robust constrained model predictive control using lin-
wileyonlinelibrary.com] ear matrix inequalities,” Automatica, Vol. 32, No. 10,
pp. 1361–1379 (1996).
complicated manner by ρ and LP. There is no simple rule to 6. Lu, Y. H., and Y. Arkun, “Quasi-min–max MPC algo-
tune these offline selected parameters since OFRMPC rithms for LPV systems,” Automatica, Vol. 36, No. 4,
solves the optimization problem at each sampling instant re- pp. 527–540 (2000).
specting the min-max optimization and physical constraints. 7. Cuzzola, F. A., J. C. Geromel, and M. Morari, “An im-
The LMI Toolbox of MATLAB 2012a (Intel(R) Core(TM) proved approach for constrained robust model predic-
i3–2120 CPU 3.30Ghz, 4G Memory) is utilized for tive control,” Automatica, Vol. 38, No. 7, pp.
simulation. 1183–1189 (2002).
8. Wada, N., K. Saito, and M. Saeki, “Model predictive
control for linear parameter varying systems using pa-
rameter dependent Lyapunov function,” IEEE Trans.
VI. CONCLUSIONS Circuits Syst.—II: Express Briefs, Vol. 53, No. 12, pp.
1446–1450 (2006).
In this paper, we have proposed a novel mixed objective 9. Wan, Z. Y., and M. V. Kothare, “An efficient off-line
OFRMPC scheme for linear uncertain systems with state and formulation of robust model predictive control using
input constraints. The proposed method is developed on the linear matrix inequalities,” Automatica, Vol. 39, No.
basis of an infinite horizon quasi-min-max algorithm with 6, pp. 837–846 (2003).
LMI formulation. For this control scheme, we first intro- 10. Wan, Z. Y., and M. V. Kothare, “Efficient robust
duced the off-line design method for the robust state ob- constrained model predictive control with a time
server. Then, in terms of LMIs, we formulated the resulting varying terminal constraint set,” Syst. Control Lett.,
online optimization problem for the OFRMPC based on the Vol. 48, No. 5, pp. 375–383 (2003).
estimated states with H2 performance as objective function 11. Wang, M. H., G. Liu, P. T. Zhao, and S. H. Yang,
while satisfying a given H∞ control performance require- “RMPC algorithm for polytopic constrained LPV sys-
ment. At the same time, we also verified the recursive feasi- tems using polyhedral invariant sets,” Control Decis.,
bility and closed-loop asymptotic stability of proposed Vol. 28, No. 11, pp. 1661–1666 (2013).
OFRMPC through an appropriate formulation to refresh 12. Zhang, H. T., H. X. Li, and G. R. Chen, “Dual-mode
the estimation error bound. The numerical example verified predictive control algorithm for constrained Hammer-
the effectiveness of the proposed designs. For further im- stein systems,” Int. J. Control, Vol. 81, No. 10, pp.
provement, we will develop the off-line design of the pro- 1609–1625 (2008).
posed OFRMPC to reduce the online computation burden 13. Copp, D. A. and J. P. Hespanha, “Nonlinear output-
and make the design more practical. Besides, more efforts feedback model predictive control with moving horizon
to reduce the conservativeness and improve the spectrum of estimation,” 53rd IEEE Conference on Decision and
the synthesis approaches of OFRMPC should be undertaken. Control, Los Angeles, California, USA, pp.
3511–3517 (2014).
14. Zhang, H. T., Z. H. Chen, Y. J. Wang, M. Li, and T.
REFERENCES
Qin, “Adaptive predictive control algorithm based on
1. Mayne, D. Q., J. B. Rawlings, C. V. Rao, and P. O. M. Laguerre Functional Model,” Int. J. Adapt. Control Sig-
Scokaert, “Constrained model predictive control: nal Process., Vol. 20, pp. 53–76 (2006).

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd
 W. Jiang et al.: Synthesis of Mixed Objective Output Feedback RMPC 1989

15. Ding, B. C., and H. G. Pan, “Output feedback robust 29. Orukpe, P. E. and I. M. Jaimoukha, “Model predictive
MPC with one free control move for the linear control based on mixed H2/H∞ control approach,” Proc.
polytopic uncertain system with bounded disturbance,” of the 2007 American Control Conference, New York,
Automatica, Vol. 50, No. 11, pp. 2929–2935 (2014). USA, pp. 11–13 (2007).
16. Wan, Z. Y., and M. V. Kothare, “Robust output feed- 30. Alessandri, A., M. Baglietto, and Battistelli, G., “On
back model predictive control using off-line linear ma- estimation error bounds for receding horizon filters
trix inequalities,” J. Process Control, Vol. 12, No. 7, pp. using quadratic boundedness,” IEEE Trans. Autom.
763–774 (2002). Control, Vol. 49, No. 8, pp. 1350–1355 (2004).
17. Huang, H., He, D. F., and Q. X. Chen, “Quasi-min-max 31. Alessandri, A., M. Baglietto, and G. Battistelli, “De-
dynamic output feedback MPC for LPV systems,” Int. sign of state estimators for uncertain linear systems
J. Syst. Control Inf. Process., Vol. 1, No. 3, pp. using quadratic boundedness,” Automatica, Vol. 42,
233–245 (2014). No. 3, pp. 497–502 (2006).
18. Ding, B. C., “Constrained robust model predictive con- 32. Ding, B. C., and H. G. Pan, “Output feedback robust
trol via parameter-dependent dynamic output feedback,” MPC with one free control move for the linear
Automatica, Vol. 46, No. 9, pp. 1517–1523 (2010). polytopic uncertain system with bounded distur-
19. Ding, B. C., X. B. Ping, and H. G. Pan, “On dynamic bance,” Automatica, Vol. 50, No. 11, pp. 2929–2935
output feedback robust MPC for constrained quasi- (2014).
LPV systems,” Int. J. Control, Vol. 86, No. 12, pp. 33. Ding, B. C., and X. B. Ping, “Output feedback predic-
2215–2227 (2013). tive control with one free control move for nonlinear
20. Ping, X. B., and B. C. Ding, “Off-line approach to dy- systems represented by a Taksgi-Sugeno Model,”
namic output feedback robust model predictive con- IEEE Trans. Fuzzy Syst., Vol. 22, No. 2, pp.
trol,” Syst. Control Lett., Vol. 62, No. 11, pp. 249–263 (2014).
1038–1048 (2013). 34. Ding, B. C., and T. Zou, “Synthesizing output feedback
21. Ping, X. B., and B. C. Ding, “Dynamic output feedback predictive control for constrained uncertain time-
robust model predictive control based on ellipsoidal es- varying discrete systems,” Acta Automatica Sinica.,
timation error bound,” Acta Automatica Sinica, Vol. 40, Vol. 33, No. 1, pp. 78–83 (2007).
No. 2, pp. 219–226 (2014). 35. Gahinet, P., A. Nemirovski, A. J. Laub, et al., LMI
22. Park, J. H., T. H. Kim, and T. Sugie, “Output feedback control toolbox User’s Guide, The Math work Inc,
model predictive control for LPV systems based on Natick, MA (1995).
quasi-min-max algorithm,” Automatica, Vol. 47, No.
9, pp. 2052–2058 (2011). Wei Jiang received the B.S. and M.S. de-
23. Su, Y., and K. K. Tan, “Comments on ‘Output feedback grees from Xi’an Institute of High-Tech,
model predictive control for LPV systems based on Xi’an, China, respectively in 2010 and
quasi-min-max algorithm’,” Automatica, Vol. 48, No. 2012. He is currently studying for a Ph.
48, pp. 2385 (2012). D. in control science and engineering at
24. Huang, H., D. W. Li, and Y. G. Xi, “Synthesis of robust Xi’an Institute of High-Tech, Xi’an,
model predictive control based on mixed H2/H∞ control China. His research interests include
approach,” Control Decis., Vol. 25, No. 8, pp. model predictive control, linear parameter varying control,
1269–1272 (2010). and nonlinear control of hypersonic vehicles.
25. Huang, H., D. W. Li, and Y. G. Xi, “The improved ro-
bust model predictive control with mixed H2/H∞ con- Hong-li Wang received his B.E., M.E.,
trol approach,” Acta Automatica Sinica, Vol. 38, No. and Ph.D. degrees from Xi’an Institute of
6, pp. 944–949 (2012). High-Tech, Xi’an, China, respectively in
26. Li, Y., “Optimal guaranteed cost control of linear un- 1988, 1991, and 1999, all in control engi-
certain system: an LMI approach,” Control Theory neering. He is currently a Professor at the
Appl., Vol. 17, No. 3, pp. 423–428 (2000). Department of Control Engineering at
27. Huang, H., D. W. Li, and Y. G. Xi, “Mixed H2/H∞ robust Xi’an Institute of High-Tech, Xi’an,
model predictive control with saturated inputs,” Int. J. China, and is also a Technology Consulting Expert to the
Syst. Sci., Vol. 45, No. 12, pp. 2565–2575 (2014). Chinese Society of Inertial Technology. His research inter-
28. Scherer, C. W., “Multiobjective H∞/H2 control,” IEEE ests mainly include precision guidance, integrated naviga-
Trans. Autom. Control., Vol. 40, No. 6, pp. tion, nonlinear control, and optimal maintenance of
1054–1062 (1995). complex system.

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd
1990 Asian Journal of Control, Vol. 19, No. 6, pp. 1977–1990, November 2017

Jing-hui Lu received his Ph.D. from Guang-bin Cai received his Ph.D. from
Xi’an Institute of High-Tech, Xi’an, Xi’an Institute of High-Tech, Xi’an,
China, in 2011. He is currently a Lecturer China, in 2012. He is a Lecturer at the De-
at the Department of Control Engineering partment of Control Engineering at Xi’an
at Xi’an Institute of High-Tech, Xi’an, Institute of High-Tech, Xi’an, China. His
China. His research interests include non- research interests include linear parame-
linear control and star navigation. ter-varying control, nonlinear control and
flight control of hypersonic vehicles.

Wei-wei Qin received his Ph.D. from

Xi’an Institute of High-Tech, Xi’an,
China, in 2012. He is a Lecturer at the De-
partment of Aerospace Engineering at
Xi’an Institute of High-Tech, Xi’an,
China. His research interests include
model predictive control and design of
flight control system.

© 2017 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd