Automatica 117 (2020) 108974

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Brief paper

Adaptive model predictive control for a class of constrained linear

systems with parametric uncertainties✩
∗
Kunwu Zhang, Yang Shi
Department of Mechanical Engineering, University of Victoria, Victoria, B.C., Canada, V8W 3P6

article info a b s t r a c t

Article history: This paper investigates adaptive model predictive control (MPC) for a class of constrained linear
Received 19 January 2019 systems with unknown model parameters. We firstly propose an online strategy for the estimation
Received in revised form 27 February 2020 of unknown parameters and uncertainty sets based on the recursive least square technique. Then the
Accepted 16 March 2020
estimated unknown parameters and uncertainty sets are employed in the construction of homothetic
Available online xxxx
prediction tubes for robust constraint satisfaction. By deriving non-increasing properties on the
Keywords: proposed estimation routine, the resulting tube-based adaptive MPC scheme is recursively feasible
Adaptive model predictive control under recursive model updates, while providing the less conservative performance compared with the
Parameter identification robust tube MPC method. Furthermore, we theoretically show that the perturbed closed-loop system
Multiplicative uncertainties is asymptotically stable under standard assumptions. Finally, numerical simulations and comparisons
are given to illustrate the efficacy of the proposed method.
© 2020 Elsevier Ltd. All rights reserved.

1. Introduction or slowly changing. To improve the performance, a general so-

lution is to reformulate the robust MPC scheme manually based
Model predictive control (MPC) has become one of the most on different description of uncertainties (Joe Qin & Badgwell,
successful methods for multivariable control systems since it pro- 2003), however, which is relatively resource-intensive and time-
vides an effective and efficient methodology to handle complex consuming. Alternatively, a promising solution is to allow for
and constrained systems (Li & Shi, 2016). The main insight of online model adaptation in the MPC framework, which is termed
MPC is to obtain a sequence of optimal control actions over as adaptive MPC.
the prediction horizon by solving an optimization problem. The In recent years, adaptive MPC has drawn increasing attention
prediction employed in MPC is conducted based on an explicit since it provides a promising solution to reduce conservatism
system model. Therefore, having an accurate model is critical of robust MPC by incorporating system identification into the
for achieving the desirable performance. However, various cat- robust MPC framework. An early work on adaptive MPC is pro-
egories of uncertainties, such as the measurement noise and the posed in Mayne and Michalska (1993) for input-constrained non-
model mismatch, are inevitable in practical control problems. Al- linear uncertain systems, where the convergence of parameter
though standard MPC, which is designed for the nominal system estimates can be guaranteed if the MPC problem is recursively
model without considering uncertainties, has certain inherent feasible. Later in Fukushima, Kim, and Sugie (2007), a data se-
robustness against sufficiently small disturbances under certain lection mechanism is considered to improve the convergence
conditions (Yu, Reble, Chen, & Allgöwer, 2014), its performance performance of parameter estimates for linear systems, then the
may be unacceptable for many practical applications due to the estimated error bound is employed to construct the compari-
limited robustness, robust MPC therefore has attracted consider- son model used in the robust MPC framework. But this method
able attention in recent years (e.g., Li & Shi, 2014; Liu, Li, Gao, & relies on the system model represented in a controllable canon-
Xu, 2018; Mayne, 2014). Yet, the robust MPC method is generally ical form. An alternative of fulfilling the persistent excitation
developed based on the given bound of uncertainties, its perfor- (PE) condition is to impose an additional constraint on system
mance is relatively conservative if the uncertainties are constant states (Marafioti, Bitmead, & Hovd, 2014) or control inputs (Vi-
cente & Trodden, 2019) to the MPC optimization problem. With
✩ This work was supported by the Natural Sciences and Engineering Research
this strategy, the convergence of parameter estimates can be
ensured, but the system state can only be stabilized in a small
Council of Canada (NSERC). The material in this paper was not presented at any
conference. This paper was recommended for publication in revised form by
region around the origin due to the presence of constraints from
Associate Editor Alessandro Abate under the direction of Editor Ian R. Petersen. the PE condition.
∗ Corresponding author. In the literature, another category of research on adaptive
E-mail addresses: [email protected] (K. Zhang), [email protected] (Y. Shi). MPC is to combine set-membership identification with robust

https://doi.org/10.1016/j.automatica.2020.108974
0005-1098/© 2020 Elsevier Ltd. All rights reserved.
2 K. Zhang and Y. Shi / Automatica 117 (2020) 108974

MPC. In Adetola, DeHaan, and Guay (2009), an ellipsoidal un- reduced conservatism. In addition, to provide a trade-off between
certainty set is constructed based on the recursive least square the computational complexity and conservatism, a specialization
(RLS) technique, then a stabilizing min–max MPC scheme is of the proposed adaptive method is also given with reduced
developed for constrained continuous-time nonlinear systems. computational complexity and comparable control performance.
The discrete-time version of this approach is presented in Adetola A numerical example and comparison study are given to illustrate
and Guay (2011). The polytope based set-membership identifi- the benefits of the proposed method.
cation is considered in Tanaskovic, Fagiano, Smith, and Morari The remainder of this paper is organized as follows: Section 2
(2014), where an adaptive output feedback MPC approach is demonstrates the problem formulation. In Section 3, the esti-
designed for constrained stable finite impulse response (FIR) mation of the unknown parameter and the uncertainty set are
systems. This method has been extended to handle chance con- discussed. An adaptive MPC algorithm is presented in Section 4,
straints (Bujarbaruah, Zhang, & Borrelli, 2018a) and time-varying followed by the analysis of closed-loop stability and recursive
systems (Tanaskovic, Fagiano, & Gligorovski, 2018). A combina- feasibility. Simulation and comparison studies are illustrated in
tion of the set-membership identification and homothetic tube Section 5. Finally Section 6 concludes this work.
MPC is proposed in Lorenzen, Allgöwer, and Cannon (2017),
Lorenzen, Cannon, and Allgöwer (2019), in which the worst- 2. Problem formulation
case realization of the uncertainty is considered based on a
set-based state prediction and uncertainty estimation. Recently,
2.1. Notation
incorporating machine learning techniques with robust MPC has
also attracted much attention. The true model is described by
Let R, Rn and Rm×n denote the sets of real numbers, column
a nominal model plus a learned model, e.g., the Gaussian pro-
real vectors with n components and real matrices consisting of
cess (Ostafew, Schoellig, Barfoot, & Collier, 2016) and the neural
n columns and m rows, respectively. The notation N denotes the
network (Nagabandi et al., 2017; Zhang, Bujarbaruah, & Bor-
set of non-negative integers, and Nba = {x ∈ N|a ≤ x ≤ b}. Given
relli, 2019). Although aforementioned works on machine learning
a vector x ∈ Rn , the Euclidean norm and infinity norm of x are
based methods have showed empirical success, how to theoreti-
cally guarantee the closed-loop stability and recursive feasibility denoted by ∥x∥ and ∥x∥∞ , respectively. We define ∥x∥Q = xT Qx.
with desired estimation performance is still a major challenge. The Pontryagin difference of sets X ⊆ Rn and Y ⊆ Rn is denoted
The application of adaptive MPC to repetitive or iterative pro- by X ⊖ Y = {z ∈ Rn : z + y ∈ X ; ∀y ∈ Y }, and the Minkowski sum
cesses can be found in Bujarbaruah, Zhang, Rosolia, and Borrelli is X ⊕ Y = {x + y|x ∈ X , y ∈ Y }. The column operation col(· · ·)
(2018b) and Rosolia and Borrelli (2017). is defined as col(x1 , x2 , . . . , xn ) = [xT1 , xT2 , . . . , xTn ]T . We use In to
In this work, we propose a computationally tractable adaptive denote an identity matrix of size n. For an unknown vector θ ,
MPC algorithm for a class of constrained linear systems subject the notations θ̂ and θ ∗ represent its estimation and real value,
to parametric uncertainties. Similar to Adetola et al. (2009), the respectively. Then the estimation error is defined as θ̃ = θ ∗ − θ̂ .
proposed method uses an RLS based estimator to identify the
unknown system parameters. Note that the estimated uncer- 2.2. Problem setup
tainty set in Adetola et al. (2009) is employed to update the
min–max optimization problem for robust constraint satisfaction, Consider a discrete-time linear time-invariant (LTI) system
which is non-convex and computationally complicated. Alterna- with an unknown parameter θ ∈ Rnθ
tively, the proposed work employs the tube MPC technique, (e.g.,
Fleming, Kouvaritakis, & Cannon, 2015; Hanema, Lazar, & Tóth, xk+1 = A(θ )xk + B(θ )uk , (1)
2017; Langson, Chryssochoos, Raković, & Mayne, 2004; Raković, subject to a mixed constraint
Kouvaritakis, Findeisen, & Cannon, 2012), to handle the uncer-
tainty, which has a comparable computational complexity to M = {(xk , uk )|Fxk + Guk ≤ 1}, (2)
standard MPC. Recently, there are some novel adaptive MPC nx nu
where xk ∈ R and uk ∈ R are the system state and input,
strategies (Lorenzen et al., 2017, 2019) combining the homothetic
respectively. The matrices A(θ )∑ and B(θ ) are the real affine
∑nθ func-
tube MPC technique, (e.g., Raković et al., 2012), with the set- nθ
tions of θ , i.e., A(θ ) = A0 + i=1 Ai θi , B(θ ) = B0 + i=1 Bi θi .
membership identification, where the sequence of state tubes
θ = col(θ1 , θ2 , . . . , θnθ ) is the vector of unknown parameters,
{Xl|k } is developed with the form Xl|k = zl|k + σl|k X0 to guarantee
which is assumed to be uniquely identifiable (Saccomani, Audoly,
the robust constraint satisfaction. Here, zl|k is the nominal system
& D’Angiò, 2003). It is assumed that the parameter θ is bounded
state, X0 is a given set and σl|k is a scalar to be optimized via
by a given set Θ0 = {θ|∥θ∥ ≤ r0 } which contains the real
the MPC optimization problem. It can be seen that the tube
parameter θ ∗ .
cross sections are shaped by the set X0 , translated and scaled by
In this paper, the goal is to design a state feedback control law
the MPC optimization problem. The set X0 is calculated offline
according to the initial knowledge of the uncertainty set, which for the perturbed and constrained system in (1) while ensuring
may be conservative under recursive updates of the uncertainty the desirable closed-loop performance and robust constraint sat-
set. Inspired by the tube MPC approach in Fleming et al. (2015), isfaction by means of adaptive MPC. In particular, we consider the
we construct the homothetic tubes in this work, where both following parameterization of the control input
the size and shape of the tube cross sections are optimized via uk = Kxk + vk , (3)
the MPC optimization problem. Consequently, it will promisingly
lead to control performance improvement by using the proposed where vk ∈ Rnu is the decision variable of the MPC optimization
method. The main contribution of this work is to extend the problem; K ∈ Rnu ×nx is a prestabilizing state feedback gain such
robust MPC framework in Fleming et al. (2015) to allow for online that φ (θ ) = A(θ ) + B(θ )K is quadratically stable for all θ ∈ Θ0 .
model adaptation, while guaranteeing the closed-loop stability
and recursive feasibility. Compared with the methods in Lorenzen Definition 1 (Blanchini & Miani, 2008). A polyhedral set Z is
et al. (2017, 2019), the proposed approach introduces additional robustly positively invariant (RPI) for the system in (1) with
decision variables in the MPC optimization problem to optimize respect to the constraint (2) and the feedback control law uk =
both the shape and size of the tube cross sections, resulting in the Kxk iff (xk , Kxk ) ∈ M and φ (θ )xk ∈ Z for all xk ∈ Z and θ ∈ Θ0 .
 K. Zhang and Y. Shi / Automatica 117 (2020) 108974 3

Suppose that Z is an RPI set for the system in (1) with respect col(I , −Ke , (−Ke )2 , . . . , (−Ke )Np −1 ). Since Ke is Schur stable, it is
to the constraint (2) and the control law uk = Kxk , if Z contains possible to find Np , lp ∈ N∞ , t d ∈ N∞ ts , ρ0 > 0 and ρ1 > 0
every RPI set, then Z is the maximal RPI (MRPI) set for the system ∑lp −1 0 t
such that ρ1 INp Nx > j=0 (w k+ j wT
k+j ) > ρ0 INp Nx for all k ∈ Ntds .
in (1). As shown in Pluymers, Rossiter, Suykens, and De Moor Therefore, the sequence wk satisfies the PE condition during a
(2015), if the MRPI set Z for the system in (1) exists, it is certain period when the system is stable. In addition, it can be de-
unique. An example of calculating the MPRI set Z can be found rived from (9) that θ̂k+1 ≈ θ̂k , Γk+1 ≈ λΓk and the corresponding
in Pluymers et al. (2015). Θ̂k ≈ Θ̂k+1 when wk is sufficiently small. Since wk is decreasing
when the system in (1) is stable, Θ̂k will converge to a fixed set
3. Uncertainty estimation in finite time.

In this section, we introduce an online parameter estima- 3.2. Uncertainty set estimation
tion scheme based on the RLS technique with guaranteed non-
increasing estimation errors. Thereafter, in order to reduce con-
To bound the unknown parameters, we introduce the follow-
servatism in robust MPC, the approximation of feasible solu-
ing ellipsoidal uncertainty set
tion set (FSS) of the unknown parameters is presented. Finally,
we conclude this section by analyzing the performance of the Θ̂k = {θ|∥θ − θ̂k ∥Γk ≤ Vk }. (10)
proposed estimation scheme.
where Vk > 0 is the bound of the estimation error. According
to (9b), we define the propagation of Vk as Vk+1 = λVk with
3.1. Parameter estimation
V0 = Λ̄(Γ0 )r02 , where Λ̄(Γ0 ) is the maximal eigenvalue of Γ0 .
∑n Let Θk denote the FSS of unknown parameters. Since unknown
Let g(xk , uk )θ = i=1 (Ai xk + Bi uk )θi , then we can formulate
θ
parameters are uniquely identifiable and stay in the a priori
a regressor model yk = g(xk , uk )θ ∗ with yk = xk+1 − A0 xk −
known set Θ0 , Θk must be the subset of Θ0 . Therefore, for all
B0 uk to estimate θ ∗ by using the standard RLS method. But the
k ≥ 1, Θk is computed as follows
convergence of this solution relies on the PE of g(xk , uk ), which
cannot be guaranteed if xk = 0 and uk = 0. Similar to Adetola Θk = Θk−1 ∩ Θ̂k . (11)
and Guay (2011), we introduce the following filter wk for the
regressor g(xk , uk ) to improve the convergence performance, By choosing suitable θ̂0 , Γ0 and V0 , Θ̂0 can be equivalent to Θ0 .
The following lemma shows the performance of uncertainty set
wk+1 = g(xk , uk ) − Ke wk , (4) estimation.
where w0 = 0 and Ke is a Schur stable gain matrix. Let x̂k denote
the system state estimated at time k − 1, based on (1) and (4), a Lemma 2. Let Θk denote the estimated uncertainty set updated by
state estimator at time k is designed as follows: following (4)–(11) at each time instant. Suppose that θ ∗ ∈ Θ0 , then
we have θ ∗ ∈ Θk for all k ≥ 0.
x̂k+1 = A0 xk + B0 uk + g(xk , uk )θ̂k+1
(5)
+ Ke x̃k + Ke wk (θ̂k − θ̂k+1 ), Proof. To prove this lemma, we firstly show that θ ∗ ∈ Θ̂k for
all k ≥ 0. Let V (θ̃k ) = θ̃kT Γk θ̃k , then it follows from Johnstone
where x̃k = xk − x̂k is the state estimation error. Then subtracting
and Anderson (1982) that V (θ̃k ) is non-increasing and V (θ̃k ) ≤
(1) from (5) yields
λV (θ̃k−1 ). When k = 0, the condition V (θ̃0 ) = θ̃0T Γ0 θ̃0 ≤ V0 holds
x̃k+1 = g(xk , uk )θ̃k+1 − Ke x̃k − Ke wk (θ̂k − θ̂k+1 ). (6) by using ∥θ̃0 ∥ ≤ r0 . When k > 0, we still have Vk ≥ V (θ̃k ) since
Vk = λk V0 and V (θ̃k ) ≤ λk V (θ̃0 ). Therefore, one gets V (θ̃k ) ≤ Vk for
In order to establish an implicit regression model for θ̂ , we all k ≥ 0. Then according to (10), it can be derived that θ ∗ ∈ Θ̂k
introduce an auxiliary variable ηk in the following for all k ≥ 0. Suppose that θ ∗ ∈ Θk . At next time instant, we have
ηk = x̃k − wk θ̃k . (7) θ ∗ ∈ Θ̂k+1 , which implies that θ ∗ ∈ Θk ∩ Θ̂k+1 = Θk+1 . Hence, it
can be concluded that θ ∗ ∈ Θk for all k ≥ 0 if θ ∗ ∈ Θ0 . ■
Then by substituting (4)–(6) into (7), one gets
Generally, the tightened state constraints are widely employed
ηk+1 = −Ke ηk . (8) in robust MPC to guarantee recursive feasibility and closed-loop
stability. These constraints are designed based the given bounds
Based on this implicit regression model, we develop the following
parameter estimator by using the standard RLS algorithm (John- of uncertainties. Hence, having an accurate description on the
uncertainty is crucial to obtain the desirable closed-loop per-
stone & Anderson, 1982)
formance. By incorporating the proposed parameter estimator,
θ̂k+1 = θ̂k + Γk−+11 wkT (x̃k − ηk ), (9a) it is possible to use the estimated parameters and uncertainty
sets at each time instant to obtain more accurate predictions
Γk+1 = λΓk + w wk , T
k (9b)
and less conservative tightened state constraints in robust MPC,
where Γ0 = β Inθ ; β is the positive scalar, and λ ∈ (0, 1) is the and thus improving the control performance. In the following
forgetting factor. Then it follows from Johnstone and Anderson section, a computationally tractable integration of tube MPC and
(1982) that the non-increasing estimation error is guaranteed, the proposed estimator is presented.
and the convergence of parameter estimates θ̂k can be achieved
if the sequence wk is persistently exciting. 4. Adaptive model predictive control
By using the proposed estimation mechanism (9), the con-
vergence of the estimation error θ̃k relies on the persistently In this section, we present a computationally tractable adap-
exciting sequence of wk instead of g(xk , uk ). Suppose that the tive MPC algorithm based on the homothetic tube MPC technique.
system is stable when k ≥ ts , ts ∈ N∞ 0 , and wts ̸ = 0. Ac- Let xl|k denote the predicted real system state l steps ahead
cording to (4), we have wk+1 = −Ke wk for all k ≥ ts . Let from time k and xl|k = zl|k + el|k , where zl|k and el|k are the
wk = {wk , wk+1 , . . . , wk+Np −1 } with Np ∈ N∞ 0 . Then it can predicted nominal system state and the error state, respectively.
be derived that wk wTk = Ke wk wkT KeT for k ≥ ts , where Ke = Our objective is to design a sequence of state tubes {Xl|k } for
4 K. Zhang and Y. Shi / Automatica 117 (2020) 108974

robust constraint satisfaction, i.e., the following conditions hold Proposition 3 shows a sequence of tightened sets for the
for some ul|k : nominal system state. By considering tube parameters {αl|k } as
extra decision variables of the MPC optimization problem, we can
xk ∈ X0|k (12a) obtain the optimal tube cross sections online.
A(θ )x + B(θ )ul|k ∈ Xl+1|k , ∀x ∈ Xl|k , θ ∈ Θk+1 (12b) According to the proposed parameter estimator, we can obtain
the new estimation of the real system with non-increasing esti-
(x, ul|k ) ∈ M, ∀x ∈ Xl|k (12c)
mation error at each time instant. Hence, a time-varying nominal
Instead of designing the state tube Xl|k directly, in this work we system is used to improve the accuracy of prediction. However,
construct the tube cross section Sl|k for the error state el|k . There- the system is considered to be invariant during the prediction. In
fore, the state tube can be established indirectly as Xl|k = zl|k ⊕Sl|k . order to improve the control performance, a time-varying termi-
In the following, we present how to design the homothetic tubes nal set is constructed based on the new estimation of uncertainty,
according to the estimation of uncertainties. which is presented in the following.

4.1. Error tube and constraint satisfaction 4.2. Construction of terminal sets

As mentioned in Section 3.1, we predict θ̂k+1 and Θk+1 at Based on Proposition 3, we define the following dynamics of
time k based on the state estimation error x̃k . Hence the system zl|k and αl|k for l ∈ N∞
N at time k
matrices A(θ̂k+1 ) and B(θ̂k+1 ) are considered in the following for
predicting the nominal system state at time k: αl+1|k = max
n
{Hkj αl|k + V ∆φkj +1 zl|k }, (16a)
j∈N0c
zl+1|k = Ak+1 zl|k + Bk+1 ul|k (13)
zl+1|k = φk+1 zl|k , (16b)
where Ak+1 = A(θ̂k+1 ) and Bk+1 = B(θ̂k+1 ); N is the prediction
horizon and l ∈ NN0 −1 . where the maximization is taken for each element in the vector.
Then subtracting (1) from (13) results in Let Zk denote the polytopic RPI set for the system xk+1 = (A(θ ) +
B(θ )K )xk with respect to the uncertainty set Θk+1 . Since θ̂k+1 ∈
el+1|k = xl+1|k − zl+1|k Θk+1 , Zk is also RPI for the system in (16b).
(14)
= φ ∗ el|k + ∆φk+1 zl|k + ∆Bk+1 vl|k , j j j j j n
Define Zl+1|k as Zl+1|k = φk+1 Zl|k with Z0|k = Zk for all j ∈ N0c ,
where φ ∗ = A(θ ∗ ) + B(θ ∗ )K , φk+1 = Ak+1 + Bk+1 K , ∆φk+1 = φ ∗ − j j j
then we have Zl+1|k ⊆ Zl|k ⊆ Zk since φk+1 is Schur stable for all
φk+1 and ∆Bk+1 = B(θ ∗ ) − Bk+1 . Since Θk is compact and convex, n
j ∈ N0c . Inspired by Proposition 3 in Fleming et al. (2015), the
we can find a polytope to over approximate Θk by following following proposition is given to construct the invariant set for
the algorithm in Sapatnekar, Vaidya, and Kang (1993). Let Θ̄k the system in (16a).
denote the polytopic over approximation of Θk , and Pol(·) is
the polytopic approximation operator from the algorithm in Sap-
Proposition 4. Define
atnekar et al. (1993). Hence Θ̄k can be directly calculated as
Θ̄k = Pol(Θk ). Due to the recursive set intersection in (11), j
f¯l|k = maxz ∈Z j {(F + GK )z },
we calculate Θ̄k indirectly to reduce the computational load, j
l|k
j
i.e., Θ̄k = Pol(Θ̂k ) ∩ Θ̄k−1 with Θ̄0 = Pol(Θ0 ). Suppose that Θ̄k c̄l|k = maxz ∈Z j {V (φk+1 − φk+1 )z }, (17)
l|k
j
can be equivalently represented by a convex hull Co(θ̂k ) where j ∈ j
ḡl|k = maxz j {V φkj +1 (z1 − z2 )}.
n 1 ,z2 ∈Zl|k
N0c and nc is an integer denoting the number of extreme points
in the convex hull. Hence, a set for the system pair (A(θ ), B(θ )) The set Ak = {α|∥α∥∞ ≤ γk , α ≥ 0} is invariant for the system in
associated with Θk can be approximated by using a convex hull (16a) while the constraint H α + (F + GK )z ≤ 1 is satisfied if the
j j j j j j
Co(Ak , Bk ) where Ak = A(θ̂k ) and Bk = B(θ̂k ). following condition holds
Inspired by the previous work (Fleming et al., 2015), we con-
sider a polytopic tube with the form Sl|k = {el|k |Vel|k ≤ αl|k } for the γ̄l|k ≥ γk ≥ γ l|k (18)
error el|k to handle multiplicative uncertainties, where V ∈ Rnv ×nx
where
is a matrix describing the shape of Sl|k ; αk ∈ Rnv ×1 is the tube
j j
parameter to be optimized. The following proposition shows a max n ∥c̄l|k ∥∞ +∥ḡl|k ∥∞ 1−max n ∥f¯l|k ∥∞
j∈N c j∈N c
sufficient condition for the robust satisfaction of constraint (2). γ l|k = 0
j , γ̄l|k = 0
∥H ∥∞
.
1−max nc ∥Hk+1 ∥∞
j∈N
0

Proposition 3. Let Sl|k = {el|k |Vel|k ≤ αl|k }. Suppose that el|k ∈ Sl|k , In addition, there exists a γk satisfying the condition (18) if l is
then el+1|k ∈ Sl+1|k . In addition, the constraint (2) is satisfied at each sufficiently large.
time instant if the following conditions hold:
Proof. This proposition can be proved by following the proof of
H αl|k + (F + GK )zl|k + Gvl|k , l ∈ NN0 −1
{
1≥ (15a) Proposition 3 in Fleming et al. (2015). ■
H αl|k + (F + GK )zl|k , l ∈ N∞
N
As shown in Fleming et al. (2015), the invariant set Ak for the
αl+1|k ≥ Hkj +1 αl|k + V (∆φkj +1 zl|k + ∆Bjk+1 vl|k ) j
system in (16a) is nonempty if ∥Hk ∥∞ < 1 for all k ≥ 0. This
n
0 , j ∈ N0
l ∈ N∞ c
(15b) condition can be satisfied by choosing the appropriate V such
j j j j j j
that the set {x|Vx ≤ 1} is a λ-contractive set for the system
where φ k = Ak + , φ
Bk K ∆ k+1 = φk+1 −φ
k+1 and ∆Bk+1 = zk+1 = φ (θ )zk , ∀θ ∈ Θ0 . An example of computing the matrix
j j
Bk+1 − Bk+1 ; H and Hk+1 are non-negative matrices satisfying the V can be found in Blanchini and Miani (2008).
j j
conditions HV = F + GK and Hk+1 V = φ
V k+1 .
Lemma 5. Given uncertainty sets Θk and Θk+1 , assume that the
Proof. Consider the uncertain input matrix B(θ ) in the system sets Ak and Ak+1 are not empty. Let Mk denote the minimum l that
(1), this proof is completed by following the proof of Proposition the condition (18) is satisfied. Then we have Mk ≥ Mk+1 if the
2 in Fleming et al. (2015). ■ condition Θk+1 ⊆ Θk holds.
 K. Zhang and Y. Shi / Automatica 117 (2020) 108974 5

Proof. According to (15) and (17), if Θk+1 ⊆ Θk , it can be derived 4.3. Construction of the cost function
that
Let vk = col(v0|k , v1|k , v2|k , . . . , vN −1|k ). Define E and T as
γ̄l|k+1 ≥ γ̄l|k , γ l|k+1 ≤ γ l|k . (19) shift matrices such that v0|k = Evk and vk = T vk+1 , then the
Since the condition (18) holds for all l ≥ Mk , one gets γ̄Mk |k ≥
prediction of zl|k[can be written ] as ξl+1|k = Ψk+1 ξl|k , where ξl|k =
φk+1 Bk+1 E
[ ]
zl|k
γ M |k . Hence, γ̄Mk |k+1 ≥ γ M |k+1 by following (19). In addition, , Ψk+1 = . Similarly, the real system state
k k vk 0 T
from the condition Zl+1|k ⊆ Zl|k , we have γ̄l+1|k ≥ γ̄l|k and
xl|k can be predicted [by using the following
[ ∗ dynamics ξ̄l+1|k =
γ l+1|k ≤ γ l|k . Therefore, there must exist a non-empty set xl|k
]
φ B∗ E
]
Ψ ξ̄l|k , where ξ̄l|k =
∗
and Ψ = ∗
. In this work,
M vk 0 T
Mk+1 = {m ∈ N0 k |γ̄Mk −m|k+1 ≥ γ }. the objective
Mk −m|k+1 ∑∞ T is to minimize a cost function with a quadratic form
i=0 (xi|k Qxi|k + ui|k Rui|k ), where Q > 0 and R > 0 are penalty
T
J̄k =
According to Proposition 4, Mk+1 can be chosen as Mk+1 = Mk −m.
matrices for the state and input, respectively. Note that the cost
Therefore, we have Mk+1 ≤ Mk . ■
function J̄k can be equivalently represented by J̄k = ξ0T|k W ∗ ξ0|k
where W ∗ is the solution of a Lyapunov equation
Remark 6. From Proposition 4, it can be seen that extra Mk
steps are required to steer αl|k into the terminal set Ak . Hence, (Ψ ∗ )T W ∗ (Ψ ∗ ) − W ∗ + Q̄ = 0 (22)
the prediction horizon is extended from N to N + Mk . Based [ T T
]
Q + K RK K RE
on Lemmas 2 and 5, it can be derived that the sequence {Mk } with Q̄ = . Since φ ∗ is unknown, we cannot
E T RK E T RE
is non-increasing. Hence, when k increases, the computational
find the matrix W ∗ exactly. Alternatively, we consider an over
complexity of MPC optimization problem is non-increasing. approximation of J̄k based on the uncertainty set updated at each
time instant.
To find the terminal set for the nominal state zN |k , we have the
following assumption:
Lemma 10. Define a new cost function Jk as Jk = ξ0T|k Wk+1 ξ0|k ,
where Wk+1 is a positive definite matrix, then Jk ≥ J̄k if the following
Assumption 7. Let Zk and Zk+1 denote the MRPI sets with
condition
respect to the uncertainty set Θk+1 and Θk+2 , respectively. Then ]T
φ (θ ) B(θ )E φ (θ ) B(θ )E
[ [ ]
the following condition holds
Wk+1 ≥ Wk+1 + Q̄ (23)
0 T 0 T
φ (θ )x ∈ Zk+1 , ∀(x, θ ) ∈ Zk × Θk+2 (20)
holds for all θ ∈ Θk+1 .
if Θk+2 ⊆ Θk+1 .
Proof. From Lemma 2, we have θ ∗ ∈ Θk+1 . Then following
Remark 8. To compute the set Zk+1 satisfying the condition (20), (23) yields Wk+1 ≥ (Ψ ∗ )T Wk+1 (Ψ ∗ ) + Q̄ . By substituting Q̄ =
we can compute the RPI set Z̄k+1 by following Algorithm 1 W ∗ − (Ψ ∗ )T W (Ψ ∗ ) into the above equation, we have Wk+1 − W ∗ ≥
in Pluymers et al. (2015) without considering (20). Then starting (Ψ ∗ )T (Wk+1 − W ∗ )(Ψ ∗ ) ≥ 0. In addition, Jk − J̄k = ξ0T|k Wk+1 ξ0|k −
with Z̄k+1 , Zk+1 can be computed by solving the linear program- ξ̄0T|k W ∗ ξ̄0|k . Since ξ̄0|k = ξ0|k and Wk+1 − W ∗ ≥ 0, it can be
ming problem with the additional constraint (20). In addition, concluded that Jk ≥ J̄k for all θ ∈ Θk+1 . ■
given Zk , Θk+1 and Θk+2 with Θk+2 ⊆ Θk+1 , there always exists
one Zk+1 such that (20) holds. A simple example is to set Zk+1 = Assumption 11. Let Wk+1 denote the weighting matrix at time
Zk directly. k, if Θk+1 ⊆ Θk , then the following condition holds for all k ≥ 0

Assumption 9. Let Mk , Mk+1 , Ak and Ak+1 are the horizons and ξ0T|k Wk+1 ξ0|k ≤ ξ0T|k Wk ξ0|k . (24)
invariant sets satisfying Proposition 4 with respect to uncertainty
sets Θk+1 and Θk+2 , respectively. Given Mk and Ak , if the condition Remark 12. Following (4)–(11), it can be guaranteed that Θk+1 ⊆
Θk+2 ⊆ Θk+1 holds, there exist Mk+1 and Ak+1 such that Mk ≤ Θk for all k ≥ 0. Given Wk ,by imposing (24) as an additional
Mk+1 and Ak ⊆ Ak+1 . constraint for the LMI problem used for computing Wk+1 , we
can find a Wk+1 satisfying the condition (24). An example of
According to Proposition 4, the feasible solution set of γk formulating the LMI problem can be found in Kouvaritakis and
in (18) becomes larger when l increases. Therefore, the larger Cannon (2015) for details.
invariant set Ak can be found by choosing the larger horizon Mk .
In addition, it follows from (11) that Θk+2 ⊆ Θk+1 for all k ≥ 0. 4.4. Adaptive MPC algorithm
Let γk = γ̄Mk |k , γk+1 = γ̄Mk+1 |k+1 and Mk+1 = Mk , then we have
Ak ⊆ Ak+1 since γk+1 ≥ γk . Therefore, given Mk and Ak , we According to the developed terminal sets and cost function,
can always find Mk+1 and Ak+1 such that Assumption 9 holds. the adaptive MPC algorithm is based on the following MPC opti-
As a result, the computational complexity of MPC optimization mization problem:
problem is still non-increasing under this assumption. P : min Jk = ξ0T|k Wk+1 ξ0|k
Suppose that the RPI set Zk has the polyhedral form Zk = vk ,{αl|k }

{x|Vk x ≤ 1}, then the terminal constraints for the systems in (16) s.t. z0|k = xk
are summarized as follows: (3), (13), (15a), (15b), (21a), (21b)
Vk zN |k + Dk αN |k ≤ 1, (21a) At time instant k, we update the estimation of the unknown
0 ≤ αN +Mk |k ≤ γk 1, (21b) parameters and the uncertainty set based on new measurements,
then reformulate the optimization problem P. Note that the refor-
where Dk is a non-negative matrix satisfying Dk V = Vk . mulation of P with respect to the new estimation is not necessary
6 K. Zhang and Y. Shi / Automatica 117 (2020) 108974

Algorithm 1: The Adaptive MPC algorithm. Since Xl−1|k+1 = Xl∗|k and Θk+2 ⊆ Θk+1 , following
Proposition 4 and Assumption 9, it is concluded that
Input: Given initial conditions x0 , Θ0 and weighting matrices
∥αl|k+1 ∥∞ ≤ γk ≤ γk+1 . Hence, the constraint (21b) is
Q , R, determine the prestabilizing feedback gain K and MRPI
satisfied.
set Z0 . Compute the terminal set A0 and the horizon M0
according to Proposition 4. Calculate the weighting matrix W0 Case (2): Suppose that the estimation termination criterion in
satisfying (23). Algorithm 1 is satisfied. Then we have θ̂k+2 = θ̂k+1 , Zk+1 =
1: for each time instant k = 0, 1, 2, · · · do Zk , Wk+2 = Wk+1 , γk+1 = γk and Mk+1 = Mk . The recursive
2: if ∥x̃k ∥ ≥ ϵx or Vk ≥ ϵr then feasibility can be proved by constructing the following candidate
Calculate θ̂k+1 and Θk+1 by using (4)–(11).
j
3: sequence v̄k+1 , {α1∗|k , α2∗|k , . . . , αN∗ +M |k , max{Hk+1 αN +Mk |k +
k n
4: Compute Mk , Ak , Wk+1 and Zk with respect to Θk+1 such j∈N0c
j
that Assumptions 7, 9 and 11 hold. V ∆φ k+1 zN +Mk |k
}}.
5: else In summary, there is a feasible solution for the optimal control
6: Let θ̂k+1 = θ̂k , Θk+1 = Θk , Zk = Zk−1 , Wk+1 = Wk , Mk = problem P at time k + 1 if it is feasible at time k. Therefore P is
Mk−1 and Ak = Ak−1 . proved to be recursively feasible. ■
7: end if
8: Reformulate and solve the optimization problem P based on Theorem 14. Suppose that Assumptions 7, 9 and 11 hold, then the
θ̂k+1 and Θk+1 to obtain v∗k . , system in (1) in closed-loop is asymptotically stable by applying the
9: Calculate the control input as uk = Kxk + v0∗|k , and then adaptive MPC Algorithm 1.
implement uk to the system.
10: end for Proof. To prove this theorem, in the following, we show that the
optimal cost Jk∗ is a Lyapunov function for the system in (1) in
closed-loop with Algorithm 1.
Case (1): Suppose that the estimation termination criterion
if the estimation error is sufficiently small. To reduce redundant
in Algorithm 1 is not satisfied. Let z0|k+1 = xk+1 , ξ0|k+1 =
estimating actions, we introduce a termination criterion for the
col(z0|k+1 , v̄k+1 ), ξ0|k = col(z0∗|k , v∗k ) and J̄k+1 = ξ0T|k+1 Wk+2 ξ0|k+1 ,
proposed estimator. Let ϵx > 0 and ϵr > 0 denote the tol-
based on Lemma 10, we have
erances for the state estimation error and the error bound of
parameter estimation, then the proposed adaptive MPC algorithm ξ0T|k+1 Wk+1 ξ0|k+1 − Jk∗
is summarized in Algorithm 1. = ξ0T|k (Ψ ∗ )T Wk+1 Ψ ∗ ξ0|k − ξ0T|k Wk+1 ξ0|k
≤ −ξ0T|k Q̄ ξ0|k
Theorem 13. Suppose that Assumptions 7, 9 and 11 hold, and there
is a feasible solution to the optimal control problem P when k = 0.
= −z0T|k Qz0|k − uT0|k Ru0|k
Then P is recursively feasible by following Algorithm 1. Since Q and R are positive definite and z0|k = xk , it can be
derived that ξ̄0T|k+1 Wk+1 ξ̄0|k+1 − Jk∗ ≤ −xTk Qxk − uT0|k Ru0|k . In
Proof. Suppose that P is feasible at time k. Let v∗k and α∗k =
addition, from Assumption 7, we have J̄k+1 = ξ̄0T|k+1 Wk+2 ξ̄0|k+1 ≤
{αl∗|k } N +Mk denote the optimal solution of the MPC problem at
l∈N0 ξ̄0T|k+1 Wk+1 ξ̄0|k+1 , which yields Jk∗+1 − Jk∗ ≤ J̄k+1 − Jk∗ ≤ −xTk Qxk −
time k. {zl∗|k , Sl∗|k = {el|k |Vk el|k ≤ αl∗|k }, Xl∗|k = zl∗|k ⊕ Sl∗|k }
l∈N0
N +Mk
uTk Ruk ≤ 0, ∀xk ̸ = 0 and uk ̸ = 0. Since Wk is positive definite, Jk∗
are the corresponding nominal states, error tubes and state tubes, is a Lyapunov function for the system in (1).
respectively. Define a candidate input sequence at time k + 1 as Case (2): Suppose that the estimation termination criterion
v̄k+1 = {v1∗|k , v2∗|k , . . . , vN∗ −1|k , 0}. in Algorithm 1 is satisfied. Then we have θ̂k+2 = θ̂k+1 , Zk+1 =
Two cases are investigated to prove this theorem. Zk , Wk+2 = Wk+1 , γk+1 = γk and Mk+1 = Mk . By repeating the
Case (1): Suppose that the estimation termination criterion in above procedure, we can prove that Jk∗ is a Lyapunov function.
Algorithm 1 is not satisfied. Based on z0|l+1 and v̄l|k+1 , we firstly In summary, the optimal cost function Jk∗ is a Lyapunov func-
construct the following sequence ᾱk+1 = {αl|k+1 } N +Mk+1 −1 such tion for the system in (1) in closed-loop with Algorithm 1. Hence,
l∈N0
j the closed-loop system is asymptotically stable. ■
that Xl|k+1 = Xl∗+1|k . Let αN +Mk+1 |k+1 = max{Hk+2 αN +Mk+1 −1|k+1 +
n
j∈N0c
j Remark 15. Note that, unlike the robust method in Fleming
V ∆φ }, we show that {v̄k+1 , ᾱk+1 } is a feasible
k+2 zN +Mk+1 −1|k+1 et al. (2015), the propagation of homothetic tube Sl|k (15) in our
solution for P in the following.
proposed method depends on the estimation θ̂k+1 and Θ̂k+1 . In
N +M −1
• For l ∈ N0 k+1 , since Xl|k+1 = Xl∗+1|k , we have {zl∗+1|k , addition, the nominal system in (13), the terminal conditions in
Sl∗+1|k } satisfying the condition zl∗+1|k ⊕ Sl∗+1|k = zl|k+1 ⊕ Sl|k+1 , (21) and the weighting matrix Wk+1 are also updated based on the
which verifies that the candidate sequence estimation of uncertainty at each time instant. By following (4)–
{zl|k+1 , αl|k+1 , v̄l|k+1 } N +Mk+1 −1 satisfies the constraints (15a) (11), the non-increasing properties on the proposed estimation
l∈N0
scheme are guaranteed. Therefore, the resulting adaptive MPC
and (15b).
scheme can reduce conservatism compared with the original
• When l = N, it follows form (21a) that Vk zN∗ |k + Dk VeN |k ≤ 1.
robust MPC method. The numerical simulations will elaborate
By using Dk V = Vk , we have Vk (zN∗ |k + eN |k ) = Vk xN |k ≤ 1,
this argument.
implying that XN∗|k ⊆ Zk . As aforementioned, XN −1|k+1 =
XN∗|k , then XN −1|k+1 ⊆ Zk . Since Zk is an RPI set, v̄N −1|k+1 = 0
Remark 16. As shown in Algorithm 1, when updating the param-
and Θk+2 ⊆ Θk+1 , it yields that XN |k+1 ⊆ φk+2 Zk ⊆ Zk+1 by
following Assumption 7. Hence, we have zN |k+1 + eN |k+1 ∈ eter estimate θ̂k and uncertainty set Θk , we need to re-compute
Zk+1 for all admissible eN |k+1 . As a result, the constraint Mk , Zk+1 and Wk+1 , which is relatively computationally expen-
(21a) is satisfied. sive. For some problems which have the strict requirement on
• When l = N + Mk+1 , taking the infinity norm of (16a) we the computational load, a solution to reduce the computational
j j
have ∥αl|k+1 ∥∞ ≤ max{∥Hk+2 ∥∞ ∥αl−1|k+1 ∥∞ + ∥c̄l−1|k+1 ∥∞ }. complexity is to choose the relatively large ϵx and ϵr . An al-
j∈N0c
n ternative is to omit the update of terminal conditions and cost
 K. Zhang and Y. Shi / Automatica 117 (2020) 108974 7

Table 1
The comparison of system performance.
Algorithm 1 Remark 15 RMPC1 RMPC2
J̄p 9.2023 9.2524 9.2524 9.3747

function by setting Mk = M0 , Zk = Z0 and Wk = W0 for all

k ≥ 0. Due to the fact that Θk+1 ⊆ Θk ⊆ Θ0 , this strategy
can significantly reduce the computational load with guaranteed
closed-loop stability and recursive feasibility, but results in a rel-
atively conservative control performance. Note that the recursive
updates of system model and uncertainty set are considered in
the tube propagation, and thus, this simplified method still has
less conservative closed-loop performance compared with the Fig. 1. The time evolution of the system state x1 .
robust MPC method. The numerical simulation will demonstrate
this argument.

5. Simulation results

In this section, a numerical example is presented to show the

effectiveness of proposed adaptive MPC algorithms. The numer-
ical test is conducted in Matlab, where the MPC optimization
problem is formulated and solved by using Yalmip (Löfberg, 2004).
We consider the following example for testing:
0.42 −0.28 −0.12 −0.08
[ ] [ ]
A0 = , A1 = , A2 = −A1 ,
0.02 0. 6 −0.12 −0.17
]T ]T
B0 = 0.3 −0.4 , B1 = 0.04 −0.08 , B2 = −1.5B1 .
[ [

Θ0 = {θ ∈ R2 |∥θ∥ ≤ 1}, {x|∥x∥∞ ≤ 17} and {u|∥u∥∞ ≤ 4}. The

weighting matrices are chosen as Q = I2 and R = 1. By follow- Fig. 2. The time evolution of the system state x2 .
ing Kouvaritakis and Cannon (2015), the prestabilizing feedback
gain is chosen as K = [−0.4187 1.1562]. Set the prediction
horizon N = 10, then the horizon and terminal region are derived
as M0 = 3 and γ0 = 0.4266. The parameters used in Algorithm
1 are given in the following ϵr = 0.001, ϵx = 0.001, λ = 0.5 and
Γ0 = 0.15I2 .
The robust MPC method in Fleming et al. (2015) (RMPC1)
and Lorenzen et al. (2019) (RMPC2) are introduced for the pur-
pose of comparison. The initial point is set as x0 = [8, 8]T . The
real system parameter θ ∗ = [−0.2, 0.5]T is given to evaluate
the proposed parameter estimator. Figs. 1–3 show the trajectories
of system state and control input obtained by applying different
control methods. From these figures, it can be seen that the
recursive feasibility can be guaranteed by using these methods
while the proposed method can accelerate the convergence of
system state. To further compare the control performances of Fig. 3. Trajectories of control input u.
different MPC formulations, we introduce the following index
∑Tstp T
k=0 (xk Qxk + uk Ruk )/Tstp , where Tstp denotes the simulation
T
J̄p =
time. The corresponding results are illustrated in Table 1, imply-
ing that the proposed method can achieve the less conservative
performance. The polytopic approximation of uncertainty sets
obtained at time k = 0, 3, 7, 20 are depicted in Fig. 4. It can
be seen that the estimate of uncertainty set is non-increasing,
and finally converges to a fixed set, which verifies the proposed
results.

6. Conclusion

In this paper, we have investigated adaptive MPC for con-

strained linear systems subject to multiplicative uncertainties.
An online parameter estimator has been designed based on the
RLS technique for simultaneous parameter identification and un-
certainty set estimation. By integrating the proposed estimator
with homothetic prediction tubes, the resulting tube-based adap-
Fig. 4. The estimated uncertainty set Θ obtained at k = 0, 3, 7, 20.
tive MPC scheme is recursively feasible with recursive model
8 K. Zhang and Y. Shi / Automatica 117 (2020) 108974

updates, while giving rise to enhanced performance compared Nagabandi, A., Yang, G., Asmar, T., Kahn, G., Levine, S., & Fearing, R. S. (2017).
with the robust tube MPC method. The simplified version of Neural network dynamics models for control of under-actuated legged
millirobots. ArXiv preprint arXiv:1711.05253.
the proposed adaptive MPC method was also given to provide Ostafew, C. J., Schoellig, A. P., Barfoot, T. D., & Collier, J. (2016). Learning-based
a trade-off between conservatism and computational complex- nonlinear model predictive control to improve vision-based mobile robot
ity. We have proven that the closed-loop system is asymptoti- path tracking. Journal of Field Robotics, 33(1), 133–152.
cally stable. Numerical simulations and comparison studies have Pluymers, B., Rossiter, J. A., Suykens, J. A. K., & De Moor, B. (2015). The efficient
computation of polyhedral invariant sets for linear systems with polytopic
been given to demonstrate the efficacy and advantages of the uncertainty. In Proceedings of the 2005 American Control Conference (pp.
proposed adaptive MPC method. On the other hand, the main 804–809). Portland, OR, USA.
limitation of proposed adaptive MPC approach comes from the Raković, S. V., Kouvaritakis, B., Findeisen, R., & Cannon, M. (2012). Homothetic
polytopic over approximation of the uncertainty set employed tube model predictive control. Automatica, 48(8), 1631–1638.
Rosolia, U., & Borrelli, F. (2017). Learning model predictive control for iterative
in the construction of homothetic tubes, leading to an undesired tasks. A data-driven control framework. IEEE Transactions on Automatic
increase in conservatism and computational complexity. Further- Control, 63(7), 1883–1896.
more, this work considered the constant parametric uncertainties Saccomani, M. P., Audoly, S., & D’Angiò, L. (2003). Parameter identifiability of
nonlinear systems: the role of initial conditions. Automatica, 39(4), 619–632.
only, which potentially poses certain limitations to practical ap-
Sapatnekar, S. S., Vaidya, P. M., & Kang, S. M. (1993). Feasible region approx-
plications. The future research will focus on how to efficiently use imation using convex polytopes. In Proceedings of 1993 IEEE International
polytopes for bounding the FSS of unknown parameters with a Symposium on Circuits and Systems (pp. 1786–1789). Chicago, IL, USA.
tight overestimation. Incorporating such an idea into the design Tanaskovic, M., Fagiano, L., & Gligorovski, V. (2018). Adaptive model predic-
tive control for constrained time variying systems. In Proceedings of 2018
of adaptive MPC algorithms for handling time-varying multiplica- European Control Conference (pp. 1698–1703). Limassol, Cyprus.
tive and additive disturbances is also an interesting direction for Tanaskovic, M., Fagiano, L., Smith, R., & Morari, M. (2014). Adaptive reced-
future research. ing horizon control for constrained MIMO systems. Automatica, 50(12),
3019–3029.
References Vicente, B. A. H., & Trodden, P. A. (2019). Stabilizing predictive control with
persistence of excitation for constrained linear systems. Systems & Control
Adetola, V., DeHaan, D., & Guay, M. (2009). Adaptive model predictive control Letters, 126, 58–66.
for constrained nonlinear systems. Systems & Control Letters, 58(5), 320–326. Yu, S., Reble, M., Chen, H., & Allgöwer, F. (2014). Inherent robustness properties
Adetola, V., & Guay, M. (2011). Robust adaptive MPC for constrained uncer- of quasi-infinite horizon nonlinear model predictive control. Automatica,
tain nonlinear systems. International Journal of Adaptive Control and Signal 50(9), 2269–2280.
Processing, 25(2), 155–167. Zhang, X., Bujarbaruah, M., & Borrelli, F. (2019). Safe and near-optimal policy
Blanchini, F., & Miani, S. (2008). Set-Theoretic Methods in Control. Springer. learning for model predictive control using primal-dual neural networks. In
Bujarbaruah, M., Zhang, X., & Borrelli, F. (2018a). Adaptive MPC with chance Proceedings of 2019 American Control Conference (pp. 354 – 359). Philadelphia,
constraints for FIR systems. In Proceedings of 2018 Annual American Control PA, USA.
Conference (pp. 2312–2317). Milwaukee, WI, USA.
Bujarbaruah, M., Zhang, X., Rosolia, U., & Borrelli, F. (2018b). Adaptive MPC for
Kunwu Zhang received the B.Eng. degree in electrical
iterative tasks. In Proceedings of the 2018 IEEE Conference on Decision and
engineering and automation from Hubei University of
Control (pp. 6322–6327). Miami Beach, FL, USA.
Science and Technology, Xianning, China, in 2013, and
Fleming, J., Kouvaritakis, B., & Cannon, M. (2015). Robust tube MPC for linear
the M.A.Sc. degree in mechanical engineering from the
systems with multiplicative uncertainty. IEEE Transactions on Automatic
University of Victoria, Victoria, BC, Canada, in 2016,
Control, 60(4), 1087–1092.
where he is currently working toward the Ph.D. degree
Fukushima, H., Kim, T. H., & Sugie, T. (2007). Adaptive model predictive control
for a class of constrained linear systems based on the comparison model. in the Department of Mechanical Engineering.
Automatica, 43(2), 301–308. His current research interests include adaptive con-
Hanema, J., Lazar, M., & Tóth, R. (2017). Stabilizing tube-based model predictive trol, model predictive control, optimization and robotic
control: Terminal set and cost construction for LPV systems. Automatica, 85, systems.
137–144.
Joe Qin, S., & Badgwell, T. A. (2003). A survey of industrial model predictive
control technology. Control Engineering Practice, 11(7), 733–764.
Johnstone, R. M., & Anderson, B. D. O. (1982). Exponential convergence of Yang Shi received the Ph.D. degree in electrical and
recursive least squares with exponential forgetting factor—adaptive control. computer engineering from the University of Alberta,
Systems & Control Letters, 2(2), 69–76. Edmonton, AB, Canada, in 2005.
Kouvaritakis, B., & Cannon, M. (2015). Model Predictive Control: Classical, Robust From 2005 to 2009, he was an Assistant Professor
and Stochastic. Springer. and Associate Professor in the Department of Mechani-
Langson, W., Chryssochoos, I., Raković, S. V., & Mayne, D. Q. (2004). Robust model cal Engineering, University of Saskatchewan, Saskatoon,
predictive control using tubes. Automatica, 40(1), 125–133. SK, Canada. In 2009, he joined the University of Victo-
Li, H., & Shi, Y. (2014). Robust distributed model predictive control of constrained ria, and now he is a Professor in the Department of
continuous-time nonlinear systems: A robustness constraint approach. IEEE Mechanical Engineering, University of Victoria, Victo-
Transactions on Automatic Control, 59(6), 1673–1678. ria, BC, Canada. His current research interests include
Li, H., & Shi, Y. (2016). Robust Receding Horizon Control for Networked and networked and distributed systems, model predictive
Distributed Nonlinear Systems. Springer. control (MPC), cyber–physical systems (CPS), robotics and mechatronics, navi-
Liu, C., Li, H., Gao, J., & Xu, D. (2018). Robust self-triggered min–max model gation and control of autonomous systems (AUV and UAV), and energy system
predictive control for discrete-time nonlinear systems. Automatica, 89, applications.
333–339. Dr. Shi received the University of Saskatchewan Student Union Teaching
Löfberg, J. (2004). YALMIP: A toolbox for modeling and optimization in MATLAB. Excellence Award in 2007. At the University of Victoria, he received the Faculty
In Proceedings of 2004 IEEE International Symposium on Computer Aided of Engineering Teaching Excellence Award in 2012, and the Craigdarroch Silver
Control System Design (pp. 284 – 289). Taipei, Taiwan. Medal for Excellence in Research in 2015. He received the JSPS Invitation
Lorenzen, M., Allgöwer, F., & Cannon, M. (2017). Adaptive model predictive Fellowship (short-term) was a Visiting Professor with the University of Tokyo
control with robust constraint satisfaction. In Proceedings of the 20th World during Nov.–Dec. 2013. His co-authored paper was awarded the 2017 IEEE Trans-
Congress of the International Federation of Automatic Control (pp. 3313–3318).
actions on Fuzzy Systems Outstanding Paper Award. He received the Humboldt
Toulouse, France.
Research Fellowship for Experienced Researchers in 2018. He is a member of
Lorenzen, M., Cannon, M., & Allgöwer, F. (2019). Robust MPC with recursive
the IEEE IES Administrative Committee during 2017–2019, and he is currently
model update. Automatica, 103, 461–471.
the Chair of IEEE IES Technical Committee on Industrial Cyber–Physical Systems.
Marafioti, G., Bitmead, R. R., & Hovd, M. (2014). Persistently exciting model pre-
Currently, he is Co-Editor-in-Chief for IEEE Transactions on Industrial Electronics;
dictive control. International Journal of Adaptive Control and Signal Processing,
He also serves as Associate Editor for Automatica, IEEE Transactions on Control
28(6), 536–552.
Mayne, D. Q. (2014). Model predictive control: Recent developments and future Systems Technology, IEEE/ASME Transactions on Mechatronics, IEEE Transactions on
promise. Automatica, 50(12), 2967–2986. Cybernetics, etc. He is a Fellow of IEEE, ASME, Engineering Institute of Canada
Mayne, D. Q., & Michalska, H. (1993). Adaptive receding horizon control for (EIC), and Canadian Society for Mechanical Engineering (CSME), and a registered
constrained nonlinear systems. In Proceedings of the 32nd IEEE Conference on Professional Engineer in British Columbia, Canada.
Decision and Control (pp. 1286–1291). San Antonio, TX, USA.