Automatica 105 (2019) 237–245

Contents lists available at ScienceDirect

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

Brief paper

Adaptive model predictive control for linear time varying MIMO

systems✩
∗
Marko Tanaskovic a , , Lorenzo Fagiano b , Vojislav Gligorovski c
a
Singidunum University, Belgrade, Serbia
b
Dipartimento di Elettronica, Informazione e Bioingegneria, Politecnico di Milano, Milano, Italy
c
Laboratory of the physics of biological systems, Swiss Federal Institute of Technology Lausanne (EPFL), Switzerland

article info a b s t r a c t

Article history: A robust, adaptive Model Predictive Control (MPC) approach for asymptotically stable, constrained
Received 21 December 2017 linear time-varying (LTV) systems with multiple inputs and outputs is proposed. The approach consists
Received in revised form 21 February 2019 of two-steps, carried out on-line with a receding horizon strategy. In the first one, a real-time Set
Accepted 28 February 2019
Membership identification algorithm exploits the measured input–output data and the available prior
Available online xxxx
knowledge to build and refine a set of admissible models of the plant (Feasible Parameter Set, FPS). This
Keywords: set is guaranteed to contain also the true system dynamics under the considered working assumptions.
Adaptive control In the second step, a robust finite-horizon optimal control problem is formulated and solved. The
Learning control variation of system dynamics is taken into account by inflating the FPS over the prediction horizon,
LTV systems according to worst-case bounds, assumed a priori, on the parameters’ rate of change. The resulting
Set membership identification
optimal control sequence guarantees that the outputs of all possible plants inside the FPS satisfy the
Model predictive control
operational constraints, also considering all possible future parameter changes. The main theoretical
Constrained control
properties of the proposed approach are demonstrated and the method is showcased in numerical
simulations, highlighting the fundamental improvement over previous approaches not designed for
LTV systems.
© 2019 Elsevier Ltd. All rights reserved.

1. Introduction and motivation and Foss (2017), Hewing, Liniger, and Zeilinger (2018), Kim and
Sugie (2008), Limon, Calliess, and Maciejowski (2017), Loren-
MPC is a successful technique in industrial applications, thanks zen, Allgöwer, and Cannon (2017), Marafioti, Bitmead, and Hovd
to the capability to explicitly account for constraints on inputs (2014), Ostafew, Schoellig, and Barfoot (2014), Tanaskovic, Fa-
and outputs and to incorporate information on future distur- giano, Smith, and Morari (2014), Wang, Sun, and Deng (2014) and
bance and reference signals (Qin & Badgwell, 2003). A research Zhou, Cairano, and Danielson (2017). These works are motivated
direction that is currently receiving growing interest is that of by the difficulty to derive models based on physical principles
dual, adaptive and learning-based MPC, i.e. approaches where the for complex processes, the increasing real-time availability of
model derivation/identification step is considered together with measured data, and the want to derive MPC approaches that can
the control computation, and possibly carried out on-line. There automatically adapt to uncertain and time-varying dynamics.
are several contributions that differ in terms of system dynam- Set Membership (SM) techniques are being adopted by several
ics (linear or nonlinear), uncertainty characterization (stochas- researchers for the model identification phase, since they provide,
tic or unknown-but-bounded), and model identification scheme in addition to a nominal model of the plant, a quantification of the
(off-line or on-line/adaptive), see e.g. Adetola, DeHan, and Guay associated uncertainty, which can be exploited for MPC design.
(2011), Aswani, Gonzalez, Sastry, and Tomlin (2013), Canale, Fa- Examples of contributions exploiting SM techniques are Canale
giano, and Signorile (2014), Di Cairano (2016), Heirung, Ydstie, et al. (2014), Limon et al. (2017), Lorenzen et al. (2017) and
Tanaskovic et al. (2014). In Tanaskovic et al. (2014), we pro-
posed the use of SM identification to derive an adaptive MPC
✩ The material in this paper was not presented at any conference. This paper
approach for uncertain linear time-invariant systems, subject to
was recommended for publication in revised form by Associate Editor Martin both process disturbance and measurement noise. This approach
Guay under the direction of Editor Miroslav Krstic.
∗ Corresponding author. guarantees robust constraint satisfaction on the outputs also dur-
E-mail addresses: mtanaskovic@singidunum.ac.rs (M. Tanaskovic),
ing model adaptation, which is a major challenge in adaptive
lorenzo.fagiano@polimi.it (L. Fagiano), vojislav.gligorovski@epfl.ch control under constraints. However, it does not provide the same
(V. Gligorovski). guarantees in presence of time-varying dynamics: indeed the

https://doi.org/10.1016/j.automatica.2019.03.030
0005-1098/© 2019 Elsevier Ltd. All rights reserved.
238 M. Tanaskovic, L. Fagiano and V. Gligorovski / Automatica 105 (2019) 237–245

Fig. 1. Impulse response of the LTV plant with ξ = 1 (stars) and ξ = 0.3 Fig. 2. Change of damping parameter ξ over time. At k = 42 the value of ξ is
(circles), and considered set Ω (thick black lines). 0.97, 3% smaller than the starting value.

method can easily fail in this case, as we show here through a amplitude 0.05. Thus, the total bound on the additive output
motivating example. The main contribution of this paper is to disturbance d is 0.15. The system dynamics are not known exactly
remove this deficiency, through a modified adaptive technique a priori, but noise-corrupted measurements of the output are
that can cope with LTV systems. As in Tanaskovic et al. (2014), a available, with bounded noise v : |v| ≤ 0.1. As prior information,
two-step procedure is employed. In the first step, a set of models we assume that the FIR coefficients belong to a polytopic set Ω ,
consistent with measured data and prior assumptions is built bounded by the thick black lines in Fig. 1, and that the bounds
and refined (Feasible Parameter Set, FPS). This set is guaranteed on the additive disturbance and measurement noise are equal to
to contain also the true system dynamics, under the considered 0.2 and 0.15, respectively. Note that this prior information is not
working assumptions. Differently from Tanaskovic et al. (2014), tight, i.e. the signal bounds and the polytopic set are not known
we now take into account the time-varying nature of the plant, exactly. The control problem is to track a given reference output,
both in the future predictions and in how the past data is ex- while satisfying the output constraints y ∈ [−6, 6]. The change
ploited to build the FPS. In the second step, a robust finite-horizon of ξ over time in our simulation test is shown in Fig. 2.
optimal control problem is formulated and solved, where we also In Tanaskovic et al. (2014), we proposed an adaptive MPC
predict all possible future changes of the model set. The proce- technique based on SM identification for uncertain LTI plants, able
dure is implemented on-line with a receding horizon strategy. We to robustly enforce output constraints also during adaptation. One
prove that the resulting feedback control law guarantees recur- of the core steps in this approach is the real-time refinement of
sive feasibility and robust constraint satisfaction, and illustrate the FPS, i.e. the set of all models compatible with the collected
its performance in the motivating example, showing that the measurements and prior assumptions. When applied to the LTV
new approach removes efficiently a fundamental limitation of the case, the approach of Tanaskovic et al. (2014) looses its theoretical
previous one. To the best of our knowledge, to date there are guarantees and in practice it can easily fail, as shown in Fig. 3,
no other adaptive techniques in the literature providing similar where at time step k = 42 the FPS becomes empty due to
theoretical guarantees in presence of output constraints, system the inconsistency between the collected data, produced by the
uncertainty, process disturbance, and measurement noise. change of system dynamics. This problem can occur with rather
small changes of the plant (3% reduction of ξ in this example),
2. Motivating example and it does not depend on the parameters’ rate of change: also
with very slow variations there is eventually an instant when
Consider an LTV, single-input, single-output mass–spring- an inconsistency among the past data can arise. Finally, note
damper system with natural frequency equal to 1, gain equal to again that the assumed initial set Ω containing the unknown
4, and eigenvalues in the left half plane with damping ratio ξ . system parameters, i.e. the starting FPS, covers all possible FIR
The value of ξ can change over time inside the interval [0.3, 1]. coefficients of the true system (see Fig. 1). Thus, the problem
The amplitude of the input u is constrained in the compact set is neither due to the parameters being outside the assumed
[−1.5, 1.5] and its rate of change ∆u in [−0.5, 0.5]. By discretiz- initial bounds. Rather, it is caused by a fundamental limitation
ing the dynamics with the trapezoid method and sampling time of the adaptive MPC approach of Tanaskovic et al. (2014) when
Ts = 0.7 s and truncating the system’s infinite impulse response applied to LTV plants. On the one hand, this effect could be
(IIR) at m = 24 samples, one obtains the Finite Impulse Response used for condition monitoring and fault detection schemes, since
(FIR) coefficients reported as an example in Fig. 1 for the values emptiness of the FPS indicates a change in the plant dynamics.
ξ = 0.3 and ξ = 1. On the other hand, we present here an approach that removes
The truncated part of the impulse response is embedded into such a limitation, restoring the recursive feasibility and constraint
an additive disturbance signal d, acting on the output y. Such satisfaction guarantees achieved in Tanaskovic et al. (2014).
a disturbance is bounded, because of the bounded input and
the fact that, for any fixed value of ξ , the corresponding LTI 3. Problem statement
system is asymptotically stable, i.e. its impulse response decays
exponentially to zero. In the described settings, the exact bound We consider a discrete-time LTV system with nu inputs and
on this additive contribution is 0.1. We further assume that an ex- ny outputs. The system at each time step t ∈ Z is known to
ogenous additive disturbance affects the output, with maximum be asymptotically stable, but the exact dynamics and the way
 M. Tanaskovic, L. Fagiano and V. Gligorovski / Automatica 105 (2019) 237–245 239

Assumption 1 (Prior Assumption on Disturbance and Noise). d and

v are bounded as:
|dj (t)| ≤ ϵdj
, ∀t ∈ Z, ∀j = 1, . . . , ny , (5)
|vj (t)| ≤ ϵvj
where ϵdj and ϵvj are positive scalars.
We further assume that the rate of change of the plant param-
.
eters, indicated as ∆H(t) = H(t) − H(t − 1), is limited.

Assumption 2 (Bounds on Parameter Rate of Change).

∆H(t) ∈ M ∈ Rny ×m : Kj Mj ≤ lj , j = 1, . . . , ny , ∀t ∈ Z,
{ }
(6)
n∆ ×m n∆
where Kj ∈ R and lj ∈ R j , j = 1, . . . , ny are chosen
j

matrices and vectors defining n∆j linear inequalities, forming a

bounded set.

Fig. 3. Plant output (thick black line), reference (dashed), uncertainty bounds Assumption 3 (Bounds on Parameter Values). The plant model
(dotted) and constraints (dash-dotted) when the adaptive MPC algorithm parameters belong to the following set at all times: H(t) ∈
from Tanaskovic et al. (2014) is used. At t = 42, the output exits the uncertainty Ω , ∀t ∈ Z, with
range defined by the FPS, which becomes empty at the next step causing the
. {
Ω = H ∈ Rny ×m : Aj0 Hj ≤ bj0 , j = 1, . . . , ny ,
}
control algorithm to fail. (7)
where the inequalities in (7) should be interpreted as element-
wise, and each matrix Aj0 ∈ Rrj0 ×m and vector bj0 ∈ Rrj0 define a
they change over time are not known precisely. We denote the
polytope with rj0 faces.
vector of control inputs by u(t) = [u1 (t), . . . , unu (t)]T , where
ui (t) ∈ R, i = 1, . . . , nu are the individual plant inputs and T The control objective is to track a given output reference and
stands for the matrix transpose operator. Similarly, we denote the reject disturbances over a possibly very long time horizon T (T ≫
vector of plant outputs by y(t) = [y1 (t), . . . , yny (t)]T . We consider m), while enforcing affine input and output constraints:
a family of models for this system, whose equations read: T
∑
yj (t) = HjT (t) ϕ (t) + dj (t), j = 1, . . . , ny . (1) min (y(t) − ydes (t))T Q (y(t) − ydes (t))
u(0),...,u(T ) (8a)
t =0
In (1), the vector d(t) = [d1 (t), . . . , dny (t)] accounts for exoge-
T
+ u(t)T Su(t) + ∆u(t)T R∆u(t)
nous additive disturbances and the effects of neglected dynamics Subject to, ∀t ∈ [0, T ]
on the outputs (e.g. due to truncation of the IIR of the system, see
Cu u(t) ≤ gu
Remark 3.1). ϕ (t) ∈ Rm is a regressor vector with m elements, (8b)
C∆u ∆u(t) ≤ g∆u
that evolves over time according to the following linear model:
Cy y(t) ≤ gy
ϕ (t + 1) = F ϕ (t) + Gu(t), (2) where ydes (t) ∈ Rny is the desired output reference, Q , S and
where matrices F ∈ R and G ∈ R m×m m×nu
depend on the chosen R are positive semi-definite weighting matrices of suitable di-
model parameterization. For example, when nu = 1 and a Finite mensions selected by the control designer, and ∆u(t) = u(t) −
u(t − 1) is the rate of change of the control input. The element-
Impulse Response (FIR) plant model is used, F and G have the
wise inequalities in (8b) define a number nu , n∆u and ny of linear
following structure:
constraints on the inputs, input rates, and outputs, respectively.
0 0 ... 0 0 1 We assume that the set defining the constraints on ∆u(t) contains
⎡ ⎤ ⎡ ⎤
⎢1 0 ... 0 0⎥ ⎢0⎥ the origin, and that the constraint set on u(t) is compact.
⎣ ..
F =⎢ .. .. .. ⎦ , G = ⎣ .. ⎦ .
.. ⎥ (3)
⎢ ⎥
. . . . . . Remark 3.1. The considered settings and assumptions hold in a
0 0 ... 1 0 0 large variety of practical problems. In fact, under the assumption
For the case nu > 1, F and G can be obtained by block diago- of asymptotic stability, one can embed the effects of truncating an
nalizing the matrices in (3). Moreover, suitable F and G matrices IIR model in the additive disturbance acting on the output of the
can be derived for Laguerre (Wahlberg, 1991), Kautz (Wahlberg, resulting FIR model. In this way, the considered model class can
1994) or generalized basis functions (Van den Hof, Heuberger, & describe the dynamic behavior of any asymptotically stable linear
Bokor, 1995) parameterizations. system. Boundedness of the modeling error due to truncation is
Each of the vectors Hj (t) ∈ Rm in (1) contains the model guaranteed by the bounds on the input and the decay rate of the
parameters that describe the influence of ϕ to the plant output impulse response. Regarding Assumptions 2 and 3, these are rea-
. sonable in practice, when physical insight and prior information
j[ at time step t. ]Defining the matrix H(t) ∈ Rny ×m as H(t) =
T on the plant are available. For an example of how to construct
H1 (t), . . . , Hny (t) , we have:
the set Ω for a realistic problem of building climate control, the
y(t) = H(t)ϕ (t) + d(t). (4) interested reader is referred to Tanaskovic, Sturzenegger, Smith,
and Morari (2017). A typical approach in case of FIR structure
The output measurement available for feedback control is cor- entails computing or assuming maximum and minimum values
rupted by noise. In particular, the vector of measured plant out- of each FIR coefficient and of its rate of change, in order to
puts ỹ(t) is given by ỹ(t) = y(t) + v (t), where v (t) = [v1 (t), build the polytopes in (6)–(7). Note that the set Ω need not
. . . , vny (t)]T and vj (t), j = 1, . . . , ny are the individual measure- be tight, and it can be set to be very large, such that a large
ment noise terms that affect each of the plant outputs. number of models are covered. In fact, this set is instrumental
240 M. Tanaskovic, L. Fagiano and V. Gligorovski / Automatica 105 (2019) 237–245

to guarantee recursive feasibility and does not affect much the and ỹ(k), as:
performance. The collected input–output data will provide much ⎧ ny ×m
⎫
tighter parameter bounds, which are adapted in real-time, than ⎪H ∈ R
⎪ : ⎪
−ϵdj − ϵvj + (t − k)ϑ j (k) ≤ HjT ϕ (k) − ỹj (k),⎬
⎨ ⎪
the fixed set Ω . Similarly, the bounds on the parameters’ change Sk (t) = . (14)
between two time steps and on the signals d(t) and v (t) need not H T ϕ (k) − ỹj (k) ≤ ϵdj + ϵvj + (t − k)ϑ j (k), ⎪
⎩ j
⎪
j = 1, . . . ny
⎪ ⎪
⎭
be known precisely: an over-approximation does not impair our
theoretical guarantees. However, assuming too large bounds can This set is guaranteed to contain the matrix of true system param-
lead to excessively cautious control and performance degrada- eters at time step t, H(t). Based on (14) and Assumptions 1–3, we
tion. In principle these bounds can be estimated on-line as well: can now define the FPS at time step t, denoted by F (t), as the one
this is a subject of future research. containing all parameter matrices H(t) consistent with the prior
assumptions and the output measurements collected up to time
4. Adaptive control algorithm and its properties step t:
( )
. ⋂
The optimization problem (8) is generally intractable. As a F (t) = Ω ∩ Sk (t) . (15)
feasible approximate solution, we propose the use of a receding k=1,...,t
horizon control policy that relies on two steps: (1) a recursive
F (t) is defined by polytopic constraints on the rows of the
set membership identification that tracks the FPSs, and (2) a
model parameter matrix H(t). Thus, it can be uniquely described
model predictive controller that exploits the model set to robustly
by a set of matrices and vectors:
enforce constraints while optimizing the plant behavior.
F (t) = H ∈ Rny ×m : Aj (t)Hj ≤ bj (t) ,
{ }
We now describe in detail these two main steps. (16)
rj (t)×m
where each of the matrices and vectors Aj (t) ∈ R , bj (t) ∈
4.1. Recursive set membership identification algorithm Rrj (t) , j = 1, . . . , ny defines rj (t) linear inequalities. Tools to com-
pute automatically Aj (t) and bj (t) are available, see e.g. Herceg,
Under the working assumptions, each new measurement col- Kvasnica, Jones, and Morari (2013). Note that, as time progresses,
lected from the plant at time step t delimits a set to which the the bounds contributed by older data (i.e. the sets Sk (t) with
parameter matrix H(t) is guaranteed to belong: k ≪ t) grow larger and larger, i.e. they naturally become loose.
{
⏐H ϕ (t) − ỹj (t)⏐ ≤ ϵd + ϵv ,
⏐ T ⏐ } To use the defined FPS F (t) for on-line control computation, a
H ∈ Rny ×m : j
St (t) = j j
(9) recursive update approach is needed. To this end, we note that
j = 1, . . . , ny the matrix Aj (t) can be derived from Aj (t − 1), j = 1, . . . , ny
where Si (j) denotes the set that is defined by the regressor and by appending two rows formed by the regressor vector at time
output measurement vectors at time step i, i.e. ϕ (i) and ỹ(i), step t, ϕ (t) and that the vector bj (t) can be formed from bj (t −
and that is guaranteed to contain the model parameter matrix 1), j = 1, . . . , ny , by first adding the terms that should account
H(j) at time step j. Geometrically, the set St (t) is formed by 2ny for the possible change of the plant model with respect to the
linear inequalities defined by the regressor ϕ (t) and the output previous time step, and then by appending two new elements
measurements ỹj (t), j = 1, . . . , ny collected at time t. In addition, that define the constraints related to the newly collected output
we note that the relation between the model parameter matrix measurement ỹj (t), j = 1, . . . , ny :
at time step t, H(t), and the regressor and plant output vectors at
bj (t − 1) + ∆bj (t − 1)
⎡ ⎤ ⎡ ⎤
Aj (t − 1)
time step t − 1, i.e. ϕ (t − 1) and y(t − 1), can be expressed by the
Aj (t) = ⎣ −ϕ T (t) ⎦ , bj (t) = ⎣ −ỹj (t) + ϵdj + ϵvj ⎦ . (17)
following equation:
ϕ T (t) ỹj (t) + ϵdj + ϵvj
y(t − 1) = H(t)ϕ (t − 1) + d(t − 1) + ϑ (t − 1), (10)
In (17) the vectors ∆bj (t − 1) ∈ Rrj (t −1) , j = 1, . . . , ny contain the
where ϑ (t − 1) ∈ R , ϑ (t − 1) = [ϑ1 (t − 1), . . . , ϑny (t − 1)]T ,
ny
bounds on the output perturbation induced by all the possible
and ϑj (t − 1) ∈ R, j = 1, . . . , ny are the contributions of the changes of the model dynamics from one time step to the next:
unmodeled dynamics to the individual plant outputs. These terms ]T
∆bj (t − 1) = 0rjo , −ϑ j (0), ϑ j (0), . . . , −ϑ j (t − 1), ϑ j (t − 1)
[
are present because the parameter matrix H(t) is used in (10)
instead of H(t − 1) in order to relate the regressor ϕ (t − 1) to
(18)
the output y(t − 1):
. rj0
with 0rjo ∈ R denoting a vector of rjo zeros.
ϑ (t − 1) = (H(t − 1) − H(t)) ϕ (t − 1). (11)
Using the recursive equation (17) would result, in general, in
From Assumption 2, it follows that the signal ϑ (t − 1) is bounded a growth of the dimensions rj (t), j = 1, . . . , ny by two with
by: each new output measurement. In this way, storing the matri-
ces Aj (t) and vectors bj (t) over time would become intractable.
ϑ j (t − 1) ≤ ϑj (t − 1) ≤ ϑ j (t − 1), j = 1, . . . , ny , (12) Therefore, in order to have a tractable recursive identification
algorithm, we keep track of the constraints generated by the
where the bounds ϑ j (t − 1) ∈ R and ϑ j (t − 1) ∈ R, j = 1, . . . , ny most recent M measurements, where M is an even number and
are the solution to the following two linear programs (LPs): a design parameter. In this way the dimensions of the matrices
. Aj (t) and the vectors bj (t) remain bounded over time, such that
ϑ j (t − 1) = minm ϕ T (t − 1)x rj (t) ≤ r0j + M , ∀j = 1, . . . , ny , ∀t. The parameter M should be
. x∈R
ϑ j (t − 1) = max ϕ T (t − 1)x (13)
selected such that a good trade-off between conservativeness and
x
computational complexity is reached.
Subject to:
Based on the elements introduced so far, Algorithm 1 summa-
Kj x ≤ l j .
rizes the proposed recursive SM identification algorithm.
Based on these definitions, the set Sk (t) is formed on the basis of From the derived FPS, we compute a nominal model of the
the regressor and output measured at time step k ≤ t, i.e. ϕ (k) plant at each time step. This is given by a matrix Hc (t) ∈ Rny ×m ,
 M. Tanaskovic, L. Fagiano and V. Gligorovski / Automatica 105 (2019) 237–245 241

Algorithm 1 Recursive update the Feasible Parameter Set Satisfaction of input constraints can be enforced by the follow-
ing set of inequalities:
(1) At t = 0, for j = 1, . . . , ny , set Aj (0) = Aj0 , bj (0) = bj0 ;
(2) At t > 0, calculate the regressor vector ϕ (t) according to (2) Cu u(k|t) ≤ gu
∀k ∈ [t , t + N − 1]. (23)
and take the measurement vector ỹ(t); C∆u ∆u(k|t) ≤ g∆u
(3) For j = 1, . . . , ny , calculate ϑ j (t) and ϑ j (t) as in (13); To define the output constraints considered in our MPC de-
(4) For j = 1, . . . , ny update Aj (t), bj (t) as in (17); sign, we first introduce the notion of predicted FPSs, denoted
(5) For j = 1, . . . , ny , if rj (t) > rj0 + M, remove rows rj0 + 1 by F (k|t), k ∈ [t + 1, t + N ]. These sets are computed by
(and if needed rj0 + 2) from Aj (t) and bj (t), such that after propagating the FPS F (t) in the future, considering the bounds
removal it holds rj (t) ≤ rj0 + M; on the rate of change of the parameters. This is done by using the
recursive identification Algorithm 1 at each predicted time step,
(6) Set t = t + 1, go to (2).
but without considering any future output measurements, which
are unknown at the current time:
F (k|t) = H ∈ Rny ×m : Aj (k|t)Hj ≤ bj (k|t) .
{ }
(24)
Hc = [Hc ,1 , . . . Hc ,ny ]T , where the vectors Hc ,j (t) ∈ Rm , j = The predicted matrices Aj (k|t) and the vectors bj (k|t), for k ∈
1, . . . , ny are calculated by solving an LP that aims to find the [t + 1, t + N − 1] and j = 1, . . . , ny are initialized as Aj (t |t) =
point inside the FPS closest to the nominal model in the previous Aj (t), bj (t |t) = bj (t) and computed by the recursion:
time step (i.e. Hc (t − 1)):
if rj (k|t) ≤ M ′
⎧
ny ⎪ A(k|t)
∑ ⎪
⎪
min ∥Hc ,j (t − 1) − Hc ,j (t)∥1 ⎪
⎪
Hc ,j (t),j=1,...,ny
⎪⎡
⎪ ⎤
j=1 (19)
⎪
⎪ aj1 (k|t)
..
⎪
⎪
Subject to:
⎪
⎨⎢
⎪ ⎢ .
⎥
⎥
Aj (t)Hcj (t) ≤ bj (t), ∀j = 1, . . . , ny . A(k + 1|t) = ⎢ (25)
⎢ ⎥
⎪⎢ ajrj0 (k|t) ⎥
otherwise ,
⎪
⎪ ⎥
The matrix Hc (0) can be initialized as an arbitrary nonzero ele- ⎢ajrj0 +3 (k|t)⎥
⎪
⎪ ⎢ ⎥
ment inside the set Ω .
⎪
..
⎪
⎪ ⎢ ⎥
⎪
.
⎪
⎪ ⎢ ⎥
⎪
⎪ ⎣ ⎦
⎩
4.2. Finite horizon optimal control problem ajrj (t) (k|t)

Let u(k|t), k ∈ [t , t + N − 1], N ≥ m, be candidate future ⎧ ⎡

0rjo
⎤
control moves, where the notation k|t indicates the prediction at
⎪
⎪ ( )
⎪ r (t) − rj0 ⎥
⎢−ϑ j t − j
⎪ ⎢
step k ≥ t given the information at the current step t. For brevity, ⎪
.
⎪ ⎥
2
⎪
we collect these decision variables in vector U = [u(t |t)T . . . u(t +
⎪
⎪ ⎢ ⎥
⎪ ⎢ ( )⎥
T T
⎪ r j (t) − rj0
N − 1|t) ] . We also define the vectors of future input increments ⎢ ϑj t −
⎪
⎪ ⎢ ⎥
b(k | t) + ⎥ if rj (k|t) ≤ M ′
⎪ ⎥
∆u(k|t), k ∈ [t , t + N − 1] as: 2
⎪
⎪ ⎢
⎪
..
⎪
⎪ ⎢ ⎥
.
{ ⎪
⎪ ⎢ ⎥
u(t |t) − u(t − 1) if k = t ⎪ ⎢ ⎥
∆u(k|t) =
⎪
⎪ ⎢ ⎥
if t + 1 ≤ k ≤ t + N − 1.
⎪
u(k|t) − u(k − 1|t)
⎪
⎪
⎪
⎣ −ϑ j (t) ⎦
ϑ j (t)
⎪
⎪
Moreover, we define the future regressor vectors ϕ (k|t) ∈ Rm ,
⎪
⎪
⎪
⎪
k ∈ [t + 1, t + N ] as:
⎪
⎪
⎪
⎪⎡ ⎤
⎪
⎪
F ϕ (t) + Gu(t |t)
{
if k = t + 1
⎪
bj1 (k|t)
⎪
ϕ (k|t) =
⎪
(20) ⎪⎢ ⎥
F ϕ (k − 1|t) + Gu(k − 1|t) if t + 2 ≤ k ≤ t + N .
⎪
.
⎪
⎨⎢ ⎥
.
⎢ ⎥
b(k+1|t) = ⎢
⎢ . ⎥
Finally, we define the current prediction error d̂(t) ∈ Rny as: ⎪ ⎥
⎪
⎪⎢ ⎥
.
⎪
⎪⎢ ⎥
d̂(t) = ỹ(t) − Hc (t)ϕ (t). (21)
⎪⎢
⎪⎢
⎪ bjrj0 (k|t) ⎥
⎥
⎪
⎪⎢ ( ) ⎥
⎢bjr +3 (k|t) − ϑ k − rj (t) − rj0 ⎥
⎪
Then, we consider the following cost function:
⎪
⎪⎢ ⎥
⎪
j
.
⎪ j0
⎪⎢ 2 ⎥ otherwise,
⎥
J(U , ỹ(t), ϕ (t)) =
⎪
⎪⎢
⎪
⎪⎢ ( ) ⎥
⎪ rj (t) − rj0 ⎥
ϑ
t +N −1
⎪⎢
⎪ b (k|t) + k −
⎢ jrj0 +4 j
∑ ( )T ⎪⎢ ⎥
2
⎪
ŷ(k + 1|t) − ydes (k + 1|t) Q (ŷ(k + 1|t) (22)
⎪
⎪ ⎥
⎪⎢ ⎥
..
⎪
⎪⎢ ⎥
k=t ⎪
.
⎪⎢
⎪ ⎥
− ydes (k + 1|t)) + u(k|t)T Su(k|t) + ∆u(k|t)T R∆u(k|t),
⎪
⎪⎢ ⎥
⎪
⎪⎢ ⎥
⎪⎢ ⎥
ϑ
⎪
where ŷ(k + 1|t) = Hc (t)ϕ (k + 1|t) + d̂(t). In (22), ỹ(t) and b (k | t) − (t)
⎪⎢ ⎥
jr (t)− 1
⎪ j
⎪
⎪⎢ j ⎥
ϕ (t) are known parameters and ydes (k|t), k ∈ [t + 1, t + N ],
⎪
⎪⎣ ⎦
⎪
bjrj (t) (k|t) + ϑ j (t)
⎩
are the predicted values of the desired output. Note that, if the
nominal model of the plant Hc (t) were equal to the real plant,
(26)
which would not change in the considered time horizon, the
measurement noise v (t) were zero, and the output disturbance where aji (k|t) and bji (k|t) denote the ith row of the matrix Aj (k|t)
d(t) were constant, for N = T , minimizing the cost function and the vector bj (k|t) respectively, rj (k|t) = rj (t) + 2(k − t)
(22) would be equivalent to minimizing the cost function of the represents the predicted dimension of the matrices Aj (k) and the
control objective (8). vectors bj (k) that would be obtained by using Algorithm 1 if no
242 M. Tanaskovic, L. Fagiano and V. Gligorovski / Automatica 105 (2019) 237–245

rows would be removed (i.e. if the dimension of the matrices and Lemma 4.2. Let Assumptions 1–3 hold. Then, when Algorithm 1 is
vectors would be allowed to grow without limit in the future). used, at each time step t, it holds that F (k|t + 1) ⊆ F (k|t), k ∈
Finally, the terminal predicted FPS, F (t + N |t), is equal to the set [t + 2, t + N ].
Ω , to which the model parameters are guaranteed to belong at
all times. We now state the main result related to recursive feasibility of
the finite horizon optimal control problem and robust constraint
Remark 4.1. The choice F (t + N |t) = Ω in principle introduces satisfaction.
additional conservativeness, since the set F (t + N |t) could be
calculated from the set F (t + N − 1|t) in the same way as for the Theorem 4.1. Let Assumptions 1–3 hold, and assume that the
sets F (k|t), k ∈ [t +1, t +N −1], and in general such a set would be problem (29) is feasible at time t = 0. Then the problem (29) is
tighter than the set Ω . However, this approach enables recursive recursively feasible and the closed-loop system obtained by applying
feasibility and robust constraint satisfaction (see Theorem 4.1 the proposed adaptive algorithm is guaranteed to satisfy input and
later on). Moreover, the impact on the performance is rather output constraints ∀t ≥ 0.
small if a long enough prediction horizon is used, as we show
in the numerical example of Section 5. Remark 4.2. With respect to our previous contribution
(Tanaskovic et al., 2014), the key technical additions that allow us
Robust satisfaction of the output constraints is guaranteed by to guarantee recursive feasibility (hence robust constraint satis-
enforcing them for all the parameters inside the predicted FPSs faction) also in the LTV case are: (1) the inflation of the FPS both
F (k|t), k ∈ [t + 1, t + N ] and for all disturbance realizations: with respect to past data (see (14)) and future time (see (25)–
(26); and (2) the use of robust output constraints at the end of the
Cy H ϕ (k|t) + d ≤ gy , ∀H ∈ F (k|t), ∀k ∈ [t + 1, t + N ], (27)
prediction horizon with respect to the whole set Ω . Both these
where d = [d1 , . . . , dno ]T , and dl ∈ R, l = 1, . . . , no are given as: measures generally increase the conservativeness of the adaptive
MPC algorithm, this is the price to pay to obtain guaranteed
ny
∑ properties. As mentioned (see Remark 4.1), such conservativeness
dl = |clj |ϵdj , can be mitigated by increasing the prediction horizon N. In this
j=1 way, the presence of the terminal constraint does not have a large
where clj stands for the element of the lth row and jth column impact on the performance at the beginning of the prediction
of the matrix Cy . Constraints (27) can be reformulated into a set horizon, which is the part that impacts the close-loop behavior.
of linear equality and inequality constraints by using Lemma 3.2 Finally, one can show that if the nominal model parameters do
from Tanaskovic et al. (2014). not change over a given time period, then during such an interval
To guarantee recursive feasibility, we finally introduce an ad- the approach exhibits integral action. The result is omitted since it
ditional generalized terminal equality constraint, as done e.g. is a minor modification of Lemma 4.1 in Tanaskovic et al. (2014).
in Fagiano and Teel (2012):
Remark 4.3. It is interesting to compare our approach with
ϕ (t + N |t) = F ϕ (t + N |t) + Gu(t + N − 1|t). (28) MPC techniques for Linear Parameter-Varying (LPV) systems. In
This means that we require the terminal regressor to correspond LPV-MPC, the system is known a priori, meaning that for a given
to a steady state for the considered model structure. Combined value of the scheduling parameter the dynamics are known. The
with (27), this guarantees robust output constraint satisfaction on scheduling parameter might be measured or not. In the first (most
an infinite horizon. frequent) case (see e.g. Casavola, Famularo, & Franze, 2002), the
For fixed values of N, Q , S and R, we can now define the finite past and current system dynamics are known, and the predictions
horizon optimal control problem (FHOCP) to be solved at each need to account for changes in the scheduling parameter, which
time step t: results in a change of the system dynamics. In the second case,
the past and current dynamics are not exactly known and one
minJ(U , ỹ(t), ϕ (t)) would need to estimate the scheduling parameter. Differently
U ,Λ
(29) from either setup, in our work we assume that the time-varying
Subject to: (23), (27), (28),
dynamics are not known. Our setup is equivalent to an LPV one
which (by using Lemma 3.2 from Tanaskovic et al. (2014)) can with unmeasured scheduling parameters, where all the coeffi-
be converted into a quadratic program (QP). The number of deci- cients of the system matrices are time-varying parameters that
sion variables and constraints of this QP depends on the chosen need to be estimated from data. Moreover, we employ a robust
prediction horizon N and the dimension of matrices and vectors approach, meaning that we estimate on-line not only a nominal
that define the FPS F (t). Its computational complexity can be value of the parameters, but also all possible system parameters
decreased by reducing the tuning parameter M, which bounds the consistent with the measured data, and we robustify the design
dimension of matrices Aj (t) and the vectors bj (t), j = 1, . . . , ny , at against such uncertainty.
the cost of higher conservativeness.
5. Simulation results — motivating example revisited
4.3. Properties of the proposed adaptive control algorithm
To illustrate its effectiveness, we apply the new approach to
The described approach guarantees recursive feasibility and the example of Section 2. We take the set Ω as in Fig. 1. We
robust satisfaction of both input and output constraints. We first use a FIR model structure with m = 24 and assume the bounds
state two results that are instrumental to prove the main one. All on ∆H(t) shown in Fig. 4. The actual values of ∆H(t) during the
the proofs are included in the Appendix. simulation are presented in Fig. 4 as well, showing that it is not
required that the assumed bounds are tight or known exactly. We
Lemma 4.1. Let Assumptions 1–3 hold. Then, the feasible parameter set the MPC prediction horizon to 30. Fig. 5 shows the obtained
set F (t) obtained by using the recursive Algorithm 1 is guaranteed results when the proposed algorithm is used. As it can be seen,
to contain the true model parameter matrix at each time step, the new approach copes with the change of the actual plant
i.e. F (t) ̸ = ∅ and H(t) ∈ F (t), ∀t ≥ 0. parameters, satisfying robustly the constraints, and the actual
 M. Tanaskovic, L. Fagiano and V. Gligorovski / Automatica 105 (2019) 237–245 243

assumptions. The recursive update of such a set accounts for

the time-varying nature of the system through assumed (not
necessarily tight) bounds on the parameters’ rate of change. The
MPC design guarantees recursive feasibility and robust satisfac-
tion of output constraints. A simulation study shows that the
new algorithm removes a fundamental limitation of a previously
proposed adaptive MPC approach.

Appendix. Main proofs

Proof of Lemma 4.1. We use induction to prove the claim of

the lemma. At time step t = 0, from the step (1) of Algorithm 1,
it holds that F (0) = Ω and from Assumption 3, it then follows
that F (0) ̸ = ∅ and that H(0) ∈ F (0). Let us now, for the sake of
the inductive argument, assume that at some time step t ≥ 0, it
Fig. 4. Bounds on the rate of change of the impulse response coefficients (solid holds that H(t) ∈ F (t). We shall show, that it then follows that
line) and actual values of ∆H(t) during the simulation (circles). H(t + 1) ∈ F (t + 1). To this end, we define matrices A′j (t) ∈
Rrj0 ×m , A′′j (t) ∈ R(rj (t)−rj0 )×m and vectors b′j (t) ∈ Rrj0 , b′′j (t) ∈
Rrj (t)−rj0 , j = 1, . . . , ny , as:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
bj1 bjrj0 +1 aj1
. .. ⎢ . ⎥
b′ (t) = ⎣ .. ⎦ , b′′ (t) = ⎣ ⎦ , Aj (t) = ⎣ .. ⎦ ,
⎢ ⎥ ⎢ ⎥ ′
.
bjrj0 bjrj (t) ajrj0
⎡ ( )⎤
r (t)−r
−ϕ T t − j 2 j0
⎢ ( )⎥
⎢ϕ t − rj (t) − rj0 ⎥
⎡ ⎤
ajrj0 +1
⎢ T ⎥
⎢ . ⎥ 2
A′′j (t) = ⎣ .. ⎦ = ⎢ ⎥.
⎢ ⎥
⎢ .
..
⎥
ajrj (t)
⎢ ⎥
⎢ ⎥
T
⎣ −ϕ (t) ⎦
ϕ T (t)
[ ′ ] [ ′ ]
Aj (t) bj (t)
Note that Aj (t) = ′′ and bj (t) = ′′ . From Assump-
Aj (t) bj (t)
Fig. 5. Plant output (thick black line), reference (dashed line), uncertainty tion 3, it holds that:
bounds (dotted lines) and constraints (dash-dotted lines) with the proposed
adaptive MPC algorithm. A′j (t)Hj (t + 1) ≤ b′j (t), j = 1, . . . , ny . (A.1)

In addition, we note that from the inductive assumptions, it holds

that A′′j (t)Hj (t) ≤ b′′j (t), j = 1, . . . , ny . Therefore, it then also holds
plant output is always inside the estimated uncertainty bounds, that:
meaning that the true plant parameters are inside the FPS despite
their time varying nature. Very good tracking performance is A′′j (t)Hj (t + 1) ≤ b′′j (t) + ej (t)
achieved, too. Regarding this aspect, we also evaluated the effect ( )
where ej (t) ∈ Rrj (t)−rj0 , ej (t) = A′′j (t) Hj (t + 1) − Hj (t) , j =
of the robust terminal constraint on the conservativeness of the
1, . . . , ny . From the definition of A′′j (t) (note that this matrix
control algorithm (see Remark 4.2), by repeating the simulation
is exclusively formed from the past regressor vectors), and the
without such a constraint. In cases when the optimal control
definition of ϑ (t) and ϑ (t) in (13), we note that the vectors
problem was not feasible, we applied the last calculated control
ej (t), j = 1, . . . , ny are bounded such that it holds ej (t) ≤ ej ,
input. During the simulation there were actually 4 time steps
where:
in which infeasibility occurred. On the other hand, the obtained [ ( )
control performance was extremely similar, with just a 3% re- rj (t) − rj0
ej = −ϑ t− ,
duction of the mean square tracking error with respect to the 2
one obtained in presence of the terminal constraint. Another sim- ( ) ]T
rj (t) − rj0
ulation study, considering a multivariable three-tank system, is ϑ t− , . . . , −ϑ (t), ϑ (t)
available in Tanaskovic, Fagiano, and Gligorovski (2018), showing 2
the advantages with respect to an adaptive MPC approach that Therefore, it holds that:
relies on a certainty equivalence model identification strategy.
A′′j (t)Hj (t + 1) ≤ b′′j (t + 1) + ej , j = 1, . . . , ny . (A.2)
6. Conclusion
Moreover, from Assumption 1, it holds that the following two
We described an adaptive MPC algorithm for asymptotically inequalities have to be satisfied:
stable, constrained LTV systems with multiple inputs and out-
−ϕ (t + 1)Hj (t + 1) ≤ −ỹj (t + 1) + ϵdj + ϵvj
puts. The technique relies on a novel recursive SM identifica- , j = 1, . . . , ny .
−ϕ (t + 1)Hj (t + 1) ≤ ỹj (t + 1) + ϵdj + ϵvj
tion approach to keep track of the set of all possible model
parameters that are consistent with data and with the prior (A.3)
244 M. Tanaskovic, L. Fagiano and V. Gligorovski / Automatica 105 (2019) 237–245

Based on (A.1), (A.2) and (A.3), it holds that: holds that F (k|t + 1) ⊆ F (k|t), k ∈ [t + 2, t + N ], which completes
† † the proof. ■
Aj (t + 1)Hj (t + 1) ≤ bj (t + 1), j = 1, . . . , ny ,
where Proof of Theorem 4.1. We first show that the FHOCP (29) is
′ b′j (t)
⎡ ⎤ recursively feasible. To this end, we use induction. The problem
Aj (t)
⎡ ⎤
′′ (29) is feasible for t = 0 by assumption. Let us assume that
† ⎢ A′′j (t) ⎥ † bj (t + 1) + ej
⎦ , bj (t + 1) = ⎢
⎢ ⎥
Aj (t + 1) = ⎣ ⎣−ỹj (t + 1) + ϵdj + ϵvj ⎦
⎥ the problem (29) is feasible at a generic time step t and let the
−ϕ (t + 1) optimal control sequence be U ∗ (t) = [u∗ (t |t), . . . , u∗ (t + N − 1|t)],
ϕ (t + 1) ỹj (t + 1) + ϵdj + ϵvj
and its corresponding sequence of predicted regressor vectors be
are the matrices that would be obtained after running the step ϕ ∗ (k|t), k = t + 1, . . . , t + N. Then, a possible feasible control
(4) of Algorithm 1 at time t + 1 (i.e. before removing any rows sequence at t + 1 is U(t + 1) = [u∗ (t + 1|t), . . . , u∗ (t + N −
from the matrices and vectors in order to{keep their dimensions 1|t), u∗ (t + N − 1|t)]. This sequence satisfies constraints (23) and
† (28). In addition, we note that the predicted regressor vectors
bounded). Therefore, the set F † (t + 1) = H ∈ Rny ×m : Aj (t + 1)
} ϕ (k|t + 1), k = t + 2, . . . , t + N + 1 that correspond to the
†
Hj ≤ bj (t + 1) is a nonempty set that is guaranteed to contain input sequence U(t + 1), by construction satisfy the equalities
H(t + 1), i.e. H(t + 1) ∈ F † (t + 1). Set F † (t + 1) represents the ϕ (k|t + 1) = ϕ ∗ (k|t), for k ∈ [t + 2, t + N ] and that from (28)
updated feasible parameter set before possible removal of any it follows that ϕ (t + N + 1|t + 1) = ϕ ∗ (t + N |t). Moreover, we
inequalities in order to bound the complexity of its description. note that from Lemma 4.2, it holds that F (k|t + 1) ⊆ F (k|t), ∀k ∈
The set F (t + 1) is obtained by either taking the set F † (t + 1) as [t + 1, t + N ]. In addition, we note that F (t + N + 1|t + 1) =
it is (i.e. when rj (t) ≤ M + rj0 , ∀j = 1, . . . , ny ), or by removing F (t + N |t) = Ω . Based on this, the sequence of inputs U(t + 1)
several inequalities that constitute it (see step 5) of Algorithm 1. satisfies the output constraints (27), and hence the FHOCP (29)
Therefore it holds that F † (t + 1) ⊆ F (t + 1), and hence it holds has a feasible solution. Repeating this argumentation for all t > 0,
that H(t + 1) ∈ F (t + 1), which means that F (t + 1) ̸ = ∅. By it can be concluded that the FHOCP (29) remains feasible ∀t > 0.
invoking the argument of mathematical induction, it then holds From this and Lemma 4.1, the other claim of the theorem follows
that H(t) ∈ F (t), ∀t ≥ 0, which completes the proof. ■ directly. ■

Proof of Lemma 4.2. We first note that, from the definition of References
F (t + 1|t) (see (24), (25) and (26)), and the way Algorithm 1
Adetola, V., DeHan, D., & Guay, M. (2011). Robust adaptive MPC for constrained
works, it holds that: uncertain nonlinear systems. International Journal of Adaptive Control and
⎡ ⎤
Aj (t + 1|t)
[ ]
bj (t + 1|t) Signal Processing, 25, 155–168.
Aswani, A., Gonzalez, H., Sastry, S., & Tomlin, C. (2013). Provably safe and robust
Aj (t + 1) = −ϕ (t + 1) , bj (t + 1) = ⎣−ỹj (t + 1) + ϵdj + ϵvj ⎦ . learning-based model predictive control. Automatica, 49, 1216–1226.
ϕ (t + 1) ỹj (t + 1) + ϵdj + ϵvj Canale, M., Fagiano, L., & Signorile, M. C. (2014). Nonlinear model predictive
control from data: a set membership approach. International Journal of Robust
Matrices Aj (k|t + 1) and vectors bj (k|t + 1), j = 1, . . . , ny are then, and Nonlinear Control, 24(1), 123–139.
by construction, formed from the matrices Aj (t + 1) and bj (t + 1). Casavola, A., Famularo, D., & Franze, G. (2002). A feedback min-max mpc
Therefore we have that, for j = 1, . . . , ny and k ∈ [t + 2, t + N ], algorithm for lpv systems subject to bounded rates of change of parameters.
it holds that: IEEE Transactions on Automatic Control, 47(7), 1147–1153.
Di Cairano, S. (2016). Indirect adaptive model predictive control for linear
Aj (k|t)
[ ]
systems with polytopic uncertainty. In American control conference (ACC),
Aj (k|t + 1) = −ϕ (t + 1) , 2016 (pp. 3570–3575). Boston, MA, USA.
ϕ (t + 1) Fagiano, L., & Teel, A. (2012). Generalized terminal state constraint for model
⎡ ⎤ predictive control. Automatica, 49, 2622–2631.
bj (k|t) Heirung, T. A. N., Ydstie, E., & Foss, B. (2017). Dual adaptive model predictive
bj (k|t + 1) = ⎣−ỹj (t + 1) + ϵdj + ϵvj ⎦ control. Automatica, 80, 340–348.
ỹj (t + 1) + ϵdj + ϵvj . Herceg, M., Kvasnica, M., Jones, C. N., & Morari, M. (2013). Multi-parametric
toolbox 3.0. In Proc. of the European Control conference (pp. 502–510). Zürich,
As it holds that Switzerland.
Hewing, L., Liniger, A., & Zeilinger, M. N. (2018). Cautious nmpc with gaussian
F (k|t) = H ∈ Rny ×m : Aj (k|t)Hj ≤ bj (k|t) ,
{ }
process dynamics for autonomous miniature race cars. In European control
conference (pp. 4029–4036). Limassol, Cyprus.
and Kim, T. -K., & Sugie, T. (2008). Adaptive receding horizon predictive control
for constrained discrete-time linear systems with parameter uncertainties.
F (k|t + 1) = H ∈ Rny ×m : Aj (k|t + 1)Hj ≤ bj (k|t + 1) ,
{ }
International Journal of Control, 81, 62–73.
∀k ∈ [t + 2, t + N − 1], Limon, D., Calliess, J., & Maciejowski, J. M. (2017). Learning-based nonlinear
model predictive control. In IFAC world congress 2017 (pp. 8035–8042).
it holds that each of the sets F (k|t + 1), k ∈ [t + 1, t + N − 1] is Toulouse, France.
formed by the same inequalities as the set F (k|t) and that it has Lorenzen, M., Allgöwer, F., & Cannon, M. (2017). Adaptive model predictive
control with robust constraint satisfaction. In IFAC world congress 2017
two additional inequalities defined by the regressor vector and
(pp. 3313–3318). Toulouse, France.
output measurement at time step t + 1. Therefore, it holds that Marafioti, G., Bitmead, R., & Hovd, M. (2014). Persistently exciting model
F (k|t + 1) ⊆ F (k|t), k ∈ [t + 2, t + N − 1]. In addition, we note predictive control. Adaptive Control and Signal Processing, 28, 536–552.
that F (t + N |t) = Ω and that for j = 1, . . . , ny , it holds that: Ostafew, C. J., Schoellig, A. P., & Barfoot, T. D. (2014). Learning-based nonlinear
[ ] [ ] model predictive control to improve vision-based mobile robot path-tracking
Aj0 b in challenging outdoor environments. In 2014 IEEE international conference
Aj (t + N |t + 1) =
A′j
, bj (t + N |t + 1) = bj0′ , on robotics and automation (ICRA) (pp. 4029–4036). Hong Kong.
j
Qin, S. J., & Badgwell, T. A. (2003). A survey of industrial model predictive control
where the matrices A′j and the vectors b′j , j = 1, . . . , ny are technology. Control Engineering Practice, 11, 733–764.
obtained by using the rules for generating the predicted matrices Tanaskovic, M., Fagiano, L., & Gligorovski, V. (2018). Adaptive model predictive
control for constrained time varying systems. In European control conference
Aj (k|t) and vectors bj (k|t) in (25) and (26). Therefore, from the (pp. 1698–1703). Limassol, Cyprus.
definition of F (t + N |t + 1) (see e.g. (24)) and the definition of Tanaskovic, M., Fagiano, L., Smith, R., & Morari, M. (2014). Adaptive receding
the set Ω in (7), it holds that F (t + N |t + 1) ⊆ F (t + N |t). Hence, it horizon control for constrained MIMO systems. Automatica, 50, 3019–3029.
 M. Tanaskovic, L. Fagiano and V. Gligorovski / Automatica 105 (2019) 237–245 245

Tanaskovic, Marko, Sturzenegger, David, Smith, Roy, & Morari, Manfred Lorenzo Fagiano received the M.Sc. degree in Auto-
(2017). Robust adaptive model predictive building climate control. motive Engineering in 2004 and the Ph.D. degree in
IFAC-PapersOnLine, 50(1), 1871–1876, 20th IFAC World Congress. Information and Systems Engineering in 2009 from
Van den Hof, P. M. J., Heuberger, P. S. C., & Bokor, J. (1995). System identification Politecnico di Torino, Italy. In 2005 he was with Fiat
with generalized orthonormal basis functions. Automatica, 31, 1821–1834. Research Centre, Torino, Italy. From 2010 to 2013 he
Wahlberg, B. (1991). System identification using Laguerre models. IEEE held positions as Marie Curie Fellow at the Politecnico
Transactions on Automatic Control, 36, 551–562. di Torino, the University of California at Santa Barbara,
Wahlberg, B. (1994). System identification using Kautz models. IEEE Transactions and ETH Zurich. From 2013 to 2016 he was scientist
on Automatic Control, 39, 1276–1282. and senior scientist at ABB Switzerland, Corporate Re-
Wang, X., Sun, Y., & Deng, K. (2014). Adaptive model predictive control of search. Lorenzo Fagiano is currently associate professor
uncertain constrained systems. In Proceedings of the 2014 American control of automation and control engineering at the Politec-
conference (pp. 2857–2862). Portland, OR. nico di Milano, Italy. His research interests include constrained estimation and
Zhou, J., Cairano, S. Di, & Danielson, C. (2017). Indirect adaptive MPC for output control, set membership methods, and applications to industrial, robotic and
tracking of uncertain linear polytopic systems. In Proceedings of the 2017 energy systems. He serves as associate editor of the IEEE Transactions of Control
American control conference (pp. 3054–3059). Seattle, WA. Systems Technology since 2015. He is recipient of the 2011 IEEE Transactions on
Systems Technology Outstanding Paper Award, of the 2010 ENI award "Debut in
Research" prize, and of two Marie Curie individual fellowships. He is co-founder
of two start-up companies, in the fields of airborne wind energy and multicopter
drones for civil applications.
Marko Tanaskovic was born in Valjevo, Serbia, in 1986.
He received the B.Sc. degree with honors from Univer-
sity of Belgrade, Serbia in 2009, the M.Sc. degree with Vojislav Gligorovski was born in Belgrade, Serbia, in
honors from ETH Zurich in 2011 and the Ph.D. degree 1995. He received B.Sc. degrees, one from School of
from ETH Zurich in 2015, all in Electric Engineering. Biology, University of Belgrade, and one from School of
During 2011 he worked for ABB Switzerland, Corporate Electrical Engineering, University of Belgrade, both in
Research, in the area of electronic component mod- 2018. From 2018, he is a Ph.D. student in Computa-
eling. During 2016 he worked for maxon motor as tional and Quantitative biology at École polytechnique
motor control development engineer. From September fédérale de Lausanne, working as a member of the
2016 he is assistant professor at Singidunum Univer- Laboratory of the Physics of Biological Systems. His
sity in Belgrade, Serbia. His research interests include research interests include application of control theory
adaptive and learning control schemes, model predictive control and sensor- to molecular systems biology, computational biology
less motor control. He has been the recipient of the scholarship for best and synthetic biology. He has been the recipient of the
Serbian students studying abroad given by the Government of Serbia for several scholarship for best students given by the Government of Serbia for several
years. years.