Automatica 103 (2019) 461–471

Contents lists available at ScienceDirect

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

Robust MPC with recursive model update✩

∗
Matthias Lorenzen a , , Mark Cannon b , Frank Allgöwer a
a
Institute for Systems Theory and Automatic Control, University of Stuttgart, Germany
b
Department of Engineering Science, University of Oxford, United Kingdom

article info a b s t r a c t

Article history: Robust constrained control of linear systems with parametric uncertainty and additive disturbance is
Received 12 March 2018 addressed. The main contribution is the introduction of a mathematically rigorous and computationally
Received in revised form 20 September 2018 tractable framework for stabilizing model predictive control with online parameter estimation to improve
Accepted 24 January 2019
performance and reduce conservatism. Requirements for closed-loop stability and provable constraint
Available online xxxx
satisfaction are considered separately, resulting in the use of online set-membership system identification
Keywords: combined with homothetic prediction tubes for robust constraint satisfaction, and an H∞ optimal point
Model predictive control estimate of the unknown parameters to achieve a finite closed-loop gain from the disturbance to the
Adaptive model predictive control state. Extensions to time-varying parameters and persistently exciting inputs to guarantee parameter
Receding horizon control convergence are presented. A numerical example illustrates the proven properties and efficacy of the
Control of constrained systems
approach.
Adaptive control
© 2019 Elsevier Ltd. All rights reserved.
Uncertain linear systems
System identification

1. Introduction As MPC includes an explicit model of the process to be con-

trolled, it should be amenable for online parameter adaption and
Model predictive control (MPC) has become one of the main system identification. The difficulty lies in finding a rigorous yet
tools to handle multivariable constrained control problems. Given tractable formulation with provable constraint satisfaction and
a model of the process, the basic idea is to solve an open-loop stability properties. Most system identification methods lack ex-
finite-horizon optimal control problem at each sampling time with plicit error bounds or convergence guarantees when used for
the current state as initial condition. Since the performance and closed-loop systems and in the presence of additive disturbance or
stability of MPC crucially depend on the model accuracy, robust measurement noise. Similarly, given a disturbance and uncertainty
and stochastic MPC has received much attention (Kouvaritakis model, most robust MPC formulations rely on extensive offline
& Cannon, 2016; Rawlings, Mayne, & Diehl, 2017). While these computations and hence are not suitable for a changing uncer-
approaches are suitable for handling unmodeled dynamics and tainty description.
rapidly changing disturbances, they are inherently conservative As a result, the existing literature on the topic can be grouped
for slowly changing or constant parametric uncertainty. To reduce into two main categories, depending on their respective focus. The
the cost of manual tuning in MPC and cope with slowly changing first category subsumes research that focuses on system identifi-
dynamics, e.g. due to changing environment or wear and tear, there cation and online learning applied in MPC. In Marafioti, Bitmead,
is a strong interest in self-tuning predictive control formulations and Hovd (2014), the authors derive a persistence of excitation
but only few solution strategies are available (Mayne, 2014; Qin & (PE) condition, which is introduced as an additional constraint in
Badgwell, 2003). the MPC optimization, thereby ensuring parameter convergence
in a recursive estimation scheme. In Heirung, Foss, and Ydstie
(2015) and Heirung, Ydstie, and Foss (2017), a Kalman Filter is
✩ M. Lorenzen and F. Allgöwer would like to thank the German Research Foun-
employed for parameter estimation and sufficient excitation is
dation (DFG) for financial support of the project within the Cluster of Excellence in
achieved by explicitly incorporating the predicted covariance ma-
Simulation Technology (EXC 310/2) at the University of Stuttgart. The material in
this paper was partially presented at the 20th World Congress of the International trix in the cost. The result is a dual control scheme, which optimizes
Federation of Automatic Control, July 9–14, 2017, Toulouse, France. This paper was identification and regulation objectives simultaneously. Further
recommended for publication in revised form by Associate Editor Franco Blanchini advanced identification methods for nonlinear systems have been
under the direction of Editor Ian R Petersen.
∗ Corresponding author. successfully combined with linear robust MPC and nonlinear MPC,
E-mail addresses: [email protected] (M. Lorenzen),
e.g., Gaussian process regression (Klenske, Zeilinger, Schölkopf,
[email protected] (M. Cannon), [email protected] & Hennig, 2016; Ostafew, Schoellig, & Barfoot, 2015) or particle
(F. Allgöwer). filtering (Bayard & Schumitzky, 2010).

https://doi.org/10.1016/j.automatica.2019.02.023
0005-1098/© 2019 Elsevier Ltd. All rights reserved.
462 M. Lorenzen, M. Cannon and F. Allgöwer / Automatica 103 (2019) 461–471

The second category subsumes research that focuses on the extends to using ellipsoids instead of polytopes; this approach
MPC algorithm, in particular convergence of the state, constraint would suffer increased conservativeness but its complexity would
satisfaction and (recursive) feasibility of the online optimization. scale better.
In Aswani, Gonzalez, Sastry, and Tomlin (2013) and Di Cairano More generally, the approach builds upon set-membership sys-
(2016), a generic online system identification is used within a tem identification, where pioneering results were published in
robust MPC framework. In both contributions, recursive feasibility the 1960s for integral type bounds and, starting in the 1980s, for
and constraint satisfaction are guaranteed by employing robust componentwise bounds (Bai, Cho, & Tempo, 1998), as well as Tube
MPC methods based on an a priori given uncertainty set that is MPC (Langson, Chryssochoos, Raković, & Mayne, 2004; Raković &
not updated online. System identification is used only to update Cheng, 2013), which is still an active research topic.
a nominal prediction model in order to increase the closed-loop Preliminary results have been presented in the conference pa-
per (Lorenzen, Allgöwer, & Cannon, 2017). These results are ex-
performance.
tended and refined here by providing a set-membership update
A small number of publications combine the advantages of both
with a priori bounded complexity and we also discuss time-varying
approaches, namely online model adaption and provable closed-
parameters and persistence of excitation. Most importantly, the
loop properties. In Kim and Sugie (2008), stabilizing adaptive MPC
algorithm has been extended to include a point-estimate and the
for linear SISO systems based on a modified least squares estima-
cost function changed to a certainty-equivalence cost instead of a
tion is introduced but the approach crucially relies on the assump- min–max cost. This allows to prove a finite ℓ2 gain, which is novel
tion of noise and additive disturbances being absent. Results for in the adaptive MPC literature.
nonlinear systems have been presented in Adetola and Guay (2011) The remainder is organized as follows. In Section 2 the main
and Guay, Adetola, and DeHaan (2015) using set-membership sys- assumptions and problem setup are introduced. In Section 3 the
tem identification, yet the algorithm leads to a computationally system identification and constraint reformulation are presen-
demanding, non-convex optimization program to be solved on- ted, followed by a summary of the stabilizing adaptive MPC
line. A computationally tractable combination of set-membership algorithm and its properties. Section 4 presents extensions to time-
system identification and robust constraint tightening has been varying parameters and persistently exciting regressors to guaran-
proposed in Tanaskovic, Fagiano, Smith, and Morari (2014) and tee parameter convergence. A numerical example is presented in
successfully applied to a building control example in Tanaskovic, Section 5 and Section 6 provides some conclusions and directions
Sturzenegger, Smith, and Morari (2017). However, the algorithm for future work.
relies on uncertain FIR models and is hence restricted to stable
Notation. The notation employed is standard. Uppercase letters
systems. Furthermore, the approach thereby requires a long pre-
are used for matrices and lower case for vectors. [A]j and [a]j denote
diction horizon and the parametric uncertainty is described by a
the jth row and entry of the matrix A and vector a, respectively. By
polytope in an high dimensional space, which increases the com- 1 we denote a column vector of ones and by I the identity matrix
putational complexity. Recently, the algorithm has been extended of appropriate size, respectively. Positive (semi)definite matrices A
in Bujarbaruah, Zhang, and Borrelli (2018) with an additional linear are denoted A ≻ 0 (A ⪰ 0). The Euclidean norm of a vector x is
Recursive Least Squares filter to handle chance constraints. denoted ∥x∥ and for A ≻ 0 we define ∥x∥2A = x⊤ Ax. R≥0 is the set
Building upon the results in Tanaskovic et al. (2014) and Veres of the non-negative reals and Nba = {n ∈ N | a ≤ n ≤ b}. We
and Norton (1993), this paper introduces an adaptive MPC al- use xk for the (real, measured) state at time k and xl|k for the state
gorithm for linear systems with model uncertainty and additive predicted l steps ahead at time k. A ⊕ B denotes the Minkowski set
disturbances, however, in a more general state-space formulation. addition and A ⊖ B denotes the Pontryagin set difference. ΠA (x) =
In particular, a computationally tractable framework for MPC with arg miny∈A ∥x − y∥ denotes the Euclidean projection of a vector
online parameter adaption and guaranteed constraint satisfaction x ∈ Rn onto a closed, nonempty set A ⊆ Rn . Bε (x) denotes an open
is presented. Requirements for robust constraint satisfaction and ball of radius ε centered on x.
for closed-loop stability are considered separately. The former re-
quires bounds on possible parameter and future state values, hence 2. Problem formulation and preliminaries
making set-membership estimation, as proposed in Tanaskovic
et al. (2014), combined with a set-membership prediction tube a This paper considers the control of a discrete-time, linear sys-
suitable choice. For the latter, gains from the disturbance to the tem with state xk ∈ Rn , input uk ∈ Rm , additive disturbance wk ∈
W ⊆ Rn , and unknown but constant1 parameter θ = θ ∗ ∈ Rp
prediction error are most relevant, which leads to the additional
use of a suitable point estimate. The state-space formulation makes xk+1 = A(θ )xk + B(θ )uk + wk . (1)
it easy to incorporate extensions that have been developed for the
We make the following assumption on the disturbance and uncer-
non-adaptive case. The emphasis is put on setpoint stabilization,
tainty.
but the proposed combination of system identification and con-
straint tightening constitutes a solid basis for work on Dual MPC
Assumption 1 (Disturbance and Uncertainty). The disturbance set
or tracking problems for systems with uncertain or slowly varying W is a bounded, convex polytope given by
parameters.
The main limitation of the proposed approach is due to using W = {w ∈ Rn | Hw w ≤ 1}
an MPC formulation with terminal constraint and terminal cost. If with Hw ∈ Rqw ×n and 1 ∈ Rqw . The system matrices depend
the uncertain system is unstable, this requires the existence of a affinely on the parameter vector θ ∈ Rp
robustly stabilizing feedback law and a common Lyapunov func- p
tion for the stabilized system, for all model parameters in a given ∑
(A(θ ), B(θ )) = (A0 , B0 ) + (Ai , Bi )[θ]i (2)
prior bounding set. Compared to the classical literature on adaptive
i=1
control, this assumption is restrictive. However, as a requirement
of robust MPC, it is in general unavoidable if state constraints are to and a bounded parameter set given by
be satisfied robustly. Similar assumptions are made in Adetola and Θ = {θ ∈ Rp | Hθ θ ≤ hθ }, (3)
Guay (2011), while Tanaskovic et al. (2014) is restricted to stable
systems. A further limitation comes from the use of polytopes, which contains the true parameter vector θ is known. ∗

which generally cause algorithm complexity to scale badly with

the system dimension. However, the presented approach directly 1 An extension to time-varying parameters in presented in Section 4.1.
 M. Lorenzen, M. Cannon and F. Allgöwer / Automatica 103 (2019) 461–471 463

To capture actuator saturation or restrict the state to desirable properties and still enables a computationally tractable MPC algo-
operating regions, the controller must enforce mixed state and rithm. In order to non-conservatively guarantee recursive feasibil-
input constraints that can be modeled with linear inequalities. ity and constraint satisfaction, the parameter estimation is based
The state and inputs are assumed to be constrained to a bounded upon a set-membership estimation, which at each time step yields
polytopic set Z given by a polytopic parameter bounding set. Furthermore, to ensure a finite
closed-loop gain from the additive disturbance to the system state
Z = {(x, u) ∈ Rn × Rm | Fx + Gu ≤ 1} (4) xk , an additional point estimate for the unknown parameter θ ∗ is
with given matrices F ∈ Rc ×n and G ∈ Rc ×m . employed. We introduce a set-membership system identification
The control objective is to find a stabilizing control law for for the state space description (1), (2) and a projected least mean
system (1) such that the constraints (x, u) ∈ Z are satisfied robustly squares algorithm to update point estimates of the parameter
and the model uncertainty (3) is taken into account and updated vector.
consistently with the state and input history.
Bounding set. Let D(x, u) ∈ Rn×p be defined by
Indirect adaptive MPC. To solve the control problem, a Tube MPC D(x, u) = A1 x + B1 u, A2 x + B2 u, . . . , Ap x + Bp u ,
[ ]
algorithm with online parameter estimation is considered, in which
the requirements for stability and robust constraint satisfaction are Dk = D(xk , uk ), and dk = A0 xk−1 + B0 uk−1 − xk . At time k ≥ 1, given
treated separately. Specifically, for the former, a nominal estimate xk−1 , uk−1 , and xk , the non-falsified parameter set can be described
θ̂k is employed in a certainty equivalence approach. For the latter, by
a bounding set Θk for the unknown parameter θ is updated con- ∆k = {θ ∈ Rp | xk − (A(θ )xk−1 + B(θ )uk−1 ) ∈ W}
sistently with the prior set Θ and the state and input trajectories (6)
{xi , ui }i∈Nk observed up to time k. = {θ ∈ Rp | −Hw Dk−1 θ ≤ 1 + Hw dk }.
0
The bounding set Θk is employed to determine an admissible With this, the membership set of the uncertain parameter θ at time
state tube {Xl|k }l∈NN , i.e., a sequence of sets Xl|k ⊂ Rn with l = k is given by
0
0, . . . , N that provides an outer bound for the predicted states and
Θk = Θk−1 ∩ ∆k (7)
satisfies the constraints (4). Specifically, let
with initial condition Θ0 = Θ .
X0|k ∋ xk , (5a)
Xl+1|k ∋ A(θ )x + B(θ )ul|k (x) + w Lemma 2 (Membership Set). The set Θk is a convex, polytopic set
(5b)
∀x ∈ Xl|k , w ∈ W, θ ∈ Θk , explicitly given in half-space form

x × ul|k (x) ∈ Z ∀x ∈ Xl|k (5c) Θk = {θ ∈ Rp | Hθk θ ≤ hθk } (8)

with some input parametrization ul|k : R → R . n m with Hθk ∈ R qk ×p

and hθk ∈ R . Furthermore, θ
qk ∗
∈ Θk and
A feedback law is then determined by a receding horizon Θk+1 ⊆ Θk .
scheme. At each sampling time k, a finite horizon optimal control
Although not necessary from a theoretical perspective, a mini-
problem is solved, given by
mal representation of Θk should be used to decrease the compu-
VN (xk , θ̂k , Θk ) = min J(xk , θ̂k , uN |k ) tational load of the MPC algorithm. Redundant constraints in (8)
uN |k
can be removed efficiently by solving a series of linear programs,
s. t. (5), XN |k ⊆ Xf cf. Blanchini and Miani (2015, Section 3.3).
with a cost function J to be defined, horizon length N, and terminal Bounded complexity set update. Under the update law (7), the
set Xf . Assuming a minimizer u∗N |k = {u∗l|k }l∈NN −1 exists, the control number of non-redundant half-spaces defining Θk in (8) could
0
law is defined by u(xk ) = u∗0|k . in principle grow without bound, making it necessary to use an
In the following, we introduce a suitable choice for the system approximation of the minimal non-falsified parameter set under
identification and tube parametrization to obtain a computation- which Lemma 2 remains valid. In the literature, several methods
ally tractable optimization problem. This leads to a basic frame- for bounded complexity set updates by means of simpler regions
work to guarantee recursive feasibility and closed-loop constraint have been proposed, including using ellipsoids (Fogel & Huang,
satisfaction in a receding horizon control algorithm. Furthermore, 1982), parallelotopes (Chisci, Garulli, Vicino, & Zappa, 1998) or
we show that a standard quadratic cost with a suitably chosen more general polytopes (Veres, Messaoud, & Norton, 1999). All
gradient-based algorithm for parameter point estimation leads to methods based on polytopic sets limit the complexity by explicitly
a finite gain property from the additive disturbance to the state xk . limiting the number of half-spaces. Similar to using ellipsoids,
the following update rule ensures a constant complexity of the
3. Robust adaptive MPC parameter set, but is more flexible.
Let Hθ ∈ Rq×p be an a priori chosen matrix whose rows
In this section, first a suitable parameter estimation scheme is define q normal directions of the facets of a polytope that bounds
introduced, followed by a computationally tractable parametriza- the estimated parameter set. At each time step k the estimated
tion for an admissible state tube {Xl|k }l∈NN . Thereafter, a cost membership set is given by Θk = {θ ∈ Rp | Hθ θ ≤ hθk } with
0
function and terminal constraint for a stabilizing MPC scheme is hθ0 ∈ Rq satisfying Θ ⊆ Θ0 and hθk+1 being a minimizer of
presented. Finally, the section concludes with a summary of the
min 1⊤ h
MPC algorithm and an analysis of system theoretic properties of h∈Rq ,Λ∈R≥0 w
q×p

the closed loop. [ ]

Hθ
s. t . Λ = Hθ (9)
3.1. Parameter estimation −Hw Dk
[ ]
hθk
Many parameter estimation schemes have been presented in Λ ≤h
1 + Hw dk+1
the literature but our main concern here is to define a parameter
estimation scheme that provides the desirable system theoretic with pw = q + qw .
464 M. Lorenzen, M. Cannon and F. Allgöwer / Automatica 103 (2019) 461–471

Lemma 3. Let hθk be chosen by (9) with hθ0 ∈ Rq such that Θ ⊆ Θ0 , with Dk = D(xk , uk ). The first inequality follows from non-expan
then θ ∗ ∈ Θk and siveness of the projection operator and θ ∗ ∈ Θ . In the second
equality and inequality we use (12) with xk+1 − x̂1|k = x̃1|k + wk ,
Θk ∩ ∆k+1 ⊆ Θk+1 ⊆ Θk . (10)
x̃1|k = Dk (θ ∗ − θ̂k ), and completion of squares. Summing (13) from
Proof. The constraints in (9) are equivalent to Θk ∩ ∆k+1 ⊆ k = 0 to m yields
Θk+1 (Kouvaritakis & Cannon, 2016), which by induction implies 1
m
∑ m
∑ 1
θ ∗ ∈ Θk for all k ≥ 0. Furthermore Λ = [I 0], h = hθk is a feasible ∥θ̂m+1 − θ ∗ ∥2 + ∥x̃1|k ∥2 ≤ ∥wk ∥2 + ∥θ̂0 − θ ∗ ∥2 ,
solution for (9) and since the constraints on the individual rows µ µ
k=0 k=0
of [h, Λ] are decoupled, it necessarily holds that h∗ ≤ hθk , which
which proves the claim.
implies Θk+1 ⊆ Θk .
∑∞
Corollary 6. If supk∈N ∥xk ∥ < ∞, supk∈N ∥uk ∥ < ∞, and k=0
Remark 4. As proposed in Chisci et al. (1998), instead of an update
in each iteration, block processing every nu steps using ∆k , . . . ,
∥wk ∥2 < ∞, then the prediction error converges to 0 asymptotically,
∆k−nu or a moving window could be used to achieve a tighter set i.e., limk→∞ ∥x̃1|k ∥ = 0.
estimate without changing the result.
To decrease the computational load, similar to Veres et al. Remark 7. In order to speed up parameter convergence, in (12)
(1999) one could update only the entries [hθk ]i of facets i that are a projection onto the set Θk could be used without changing
closest to facets in ∆k in an inner product sense. But note that the the result in Lemma 5. Projecting the parameter estimate onto
update algorithm proposed in Veres et al. (1999) does not always a trust region is common practice for robust adaptive control
yield the best reduction in size as only one (and not necessarily the algorithms (Åström & Wittenmark, 2008). Furthermore, note that
best) facet is updated. the update gain µ is non-increasing with an increasing size of Z,
which is in line with common recommendations on stabilizing
Point estimate. Methods based on Kalman filters or Recursive Least indirect adaptive control, cf. Bitmead, Gevers, and Wertz (1990,
Squares have been suggested for parameter estimation. Instead, Chapter 7.4).
in order to provide closed-loop stability guarantees, we propose a
Least Mean Squares (LMS) filter. For FIR models, in Hassibi, Sayed, 3.2. State-tube and constraint satisfaction
and Kailath (1993), the LMS filter has been shown to be an H∞
optimal map from the disturbance sequence to the sequence of
To cope with state predictions under uncertainty an input
prediction errors, reaching the best achievable gain of 1.
parametrization
Given a parameter estimate θ̂k , denote the predicted state by
x̂1|k = A(θ̂k )xk + B(θ̂k )uk and the prediction error by ul|k (x) = Kx + vl|k

x̃1|k = A(θ )xk + B(θ )uk − x̂1|k .

∗ ∗
(11) is introduced with decision variables vN |k = {vl|k }l∈NN −1 , vl|k ∈ Rm
0
and prestabilizing feedback gain K ∈ Rm×n satisfying the following
For given θ̂0 ∈ Θ and parameter update gain µ ∈ R>0 satisfying
assumption.
1
µ
> sup(x,u)∈Z ∥D(x, u)∥2 the estimate θ̂k is defined recursively by

θ̃k = θ̂k−1 + µD(xk−1 , uk−1 )⊤ xk − x̂1|k−1 Assumption 8 (Prestabilization). The feedback gain K is chosen
( )
(12) such that Acl (θ ) = A(θ ) + B(θ )K is stable for all θ ∈ Θ .
θ̂k = ΠΘ θ̃k ,
( )
While this assumption, compared to most adaptive control
where ΠΘ (θ̃ ) = arg minθ ∈Θ ∥θ − θ̃∥ denotes the Euclidean projec- literature, is conservative, it is necessary to prove robust constraint
tion of a point θ̃ ∈ Rp onto the set Θ . satisfaction. Note that, given the initial parameter set Θ , the gain
The following lemma summarizes the relevant properties, K can be determined by standard robust control methods.
which will be used in the following stability analysis of the MPC By restricting the sets Xl|k to be translations and dilations of a
closed loop. given polytope X0 , the MPC optimization can be recast as a com-
putationally tractable, finite dimensional optimization program.
Lemma 5 (Point Estimate). If supk∈N ∥xk ∥ < ∞, supk∈N ∥uk ∥ < ∞, In the context of robust MPC, the following parametrization has
then the parameter estimate θ̂k is bounded, in accordance with the previously been introduced in Langson et al. (2004) and Raković,
prior parameter set, i.e. θ̂k ∈ Θ , and Kouvaritakis, Findeisen, and Cannon (2012).
∑m
∥x̃1|k ∥2 For a given polytope X0 = {x ∈ Rn | Hx x ≤ 1} with vertices
sup k=0
∑m ≤ 1. {x1 , . . . , xv }, MPC decision variables zN |k = {zl|k }l∈NN , zl|k ∈ Rn , and
m∈N,wk ∈W,θ̂0 ∈Θ µ
1
∥θ̂0 − θ ∗ ∥2 + k=0 ∥wk ∥2 αN |k = {αl|k }l∈NN , αl|k ∈ R≥0 , define
0

Proof. Boundedness of θ̂k and θ̂k ∈ Θ follow trivially from the set Xl|k = {zl|k } ⊕ αl|k X0
update (6), (7) and projection. To prove the bound on the prediction = {x ∈ Rn | Hx (x − zl|k ) ≤ αl|k 1} (14)
error consider v
= {zl|k } ⊕ αl|k co{x , x , . . . , x }.
1 2
1 1
∥θ̂k+1 − θ ∥ −
∗ 2
∥θ̂k − θ ∥
∗ 2
The reason for choosing this tube parametrization is the explicit
µ µ
1 1 description of Xl|k in both vertex and half-space form. This can be
≤ ∥θ̃k+1 − θ ∗ ∥2 − ∥θ̂k − θ ∗ ∥2 exploited to reformulate the prediction constraints (5) as linear
µ µ constraints while taking into account the online updated param-
1 2 eter set Θk .
= ∥θ̃k+1 − θ̂k ∥2 + (θ̃k+1 − θ̂k )⊤ (θ̂k − θ ∗ ) (13)
µ µ For a more compact notation in the following proposition, de-
j j j j j j
1 fine dl|k = A0 xl|k + B0 ul|k − zl+1|k and Dl|k = D(xl|k , ul|k ), where
k (x̃1|k + wk )∥ + 2(x̃1|k + wk ) Dk (θ̂k − θ )
2
= ∥µD⊤ ⊤ ∗
j j j
µ xl|k = zl|k + αl|k xj and ul|k = ul|k (xl|k ). Furthermore, for all i ∈ Nu1
≤ (µ∥Dk ∥2 − 1)∥x̃1|k + wk ∥2 − ∥x̃1|k ∥2 + ∥wk ∥2 define [w̄]i = maxw∈W [Hx ]i w and for all i ∈ Nc1 define [f¯ ]i =
maxx∈X0 [F + GK ]i x with F , G from (4).
≤ −∥x̃1|k ∥2 + ∥wk ∥2
 M. Lorenzen, M. Cannon and F. Allgöwer / Automatica 103 (2019) 461–471 465

Proposition 9 (Prediction Tube). Let {Xl|k }l∈NN be parametrized as Lemma 2 and Proposition 9 constitute the basic framework for
0
in (14) with decision variables zN |k , αN |k , and vN |k . a predictive control algorithm with model adaption and robust
Eqs. (5a)–(5c) are satisfied if and only if for all j ∈ Nv1 , l ∈ NN0 −1 constraint satisfaction. In the next section, this is complemented
j u×q with a suitable terminal constraint and a convex, positive definite
there exists Λl|k ∈ R≥0 k such that
objective function to derive a stabilizing, adaptive MPC scheme.
(F + GK )zl|k + Gvl|k + αl|k f¯ ≤ 1 (15a)
3.3. Terminal constraint and objective function
− Hx z0|k − α0|k 1 ≤ −Hx xk (15b)
j j
Λl|k hθk + Hx dl|k − αl+1|k 1 ≤ −w̄ (15c) As is common in stabilizing MPC, to prove stability and recur-
sive feasibility, we assume the existence of a robustly invariant
j j
Hx Dl|k = Λl|k Hθk . (15d) terminal set for the state tube. Note that if the system dynamics are
The value of Proposition 9 is that it allows the constraints on stable and no state constraints are present, the terminal constraint
state tubes to be incorporated as linear constraints into the MPC can be omitted by setting the prestabilizing feedback gain K to
optimization. zero.

Proof. Inequality (5c) is equivalent to Assumption 11 (Terminal Set). There exists a nonempty terminal
set Xf = {(z , α ) ∈ Rn × R≥0 | HT z + hT α ≤ 1} such that (x, Kx) ∈ Z
(F + GK )zl|k + Gvl|k + αl|k (F + GK )x ≤ 1 ∀x ∈ X0 , for all x ∈ {z } ⊕ α X0 , (z , α ) ∈ Xf and for all θ ∈ Θ we have
which is equivalent to (15a) when maximized over x ∈ X0 . (z , α ) ∈ Xf H⇒ ∃(z + , α + ) ∈ Xf s. t.
Inequality (5a) is equivalent to (15b), and (5b) is equivalent
Acl (θ )({z } ⊕ α X0 ) ⊕ W ⊆ {z + } ⊕ α + X0 .
to (15c), (15d) as shown by the following reformulation.
Xl+1|k ⊇ Acl (θ )Xl|k ⊕ B(θ )vl|k ⊕ W ∀θ ∈ Θk The assumption on the terminal set Xf is analogous to the
terminal set in robust MPC with homothetic tubes, cf. Raković and
Hx (Acl (θ )x + B(θ )vl|k + w − zl+1|k ) ≤ αl+1|k 1 Cheng (2013). It can be computed recursively, following standard
⇔
∀x ∈ Xl|k , θ ∈ Θk , w ∈ W algorithms to determine a robust invariant set through considering
the dynamics of (z , α ). For completeness, an explicit algorithm is
Hx (Acl (θ )(zl|k + αl|k xj ) + B(θ )vl|k − zl+1|k )
⇔ given in Appendix A.1.
− αl+1|k 1 ≤ −w̄ ∀j ∈ Nv1 , θ ∈ Θk In the following, we use the point estimate θ̂k to define a
max Hx (Acl (θ )(zl|k + αl|k xj ) + B(θ )vl|k ) certainty equivalence cost, which leads to a finite ℓ2 gain from
{ }
⇔ θ ∈Θk the disturbance to the state of the closed-loop system. Similar to
− Hx zl+1|k − αl+1|k 1 ≤ −w̄ ∀j ∈ Nv1 nominal MPC for linear systems, a quadratic cost on the nominal
{ } predicted state and input based on the parameter estimate θ̂k is
⇔ max Hx Djl|k θ + Hx djl|k − αl+1|k 1 ≤ −w̄ ∀j ∈ Nv1 employed. Let Q ∈ Rn×n and R ∈ Rm×m be positive definite cost
θ ∈Θ
⎧ kj j
⎫ matrices and let P ∈ Rn×n satisfy
⎪ Λl|k hθk + Hx dl|k − αl+1|k 1 ≤ −w̄ ⎪
Acl (θ )⊤ PAcl (θ ) + Q + K ⊤ RK ⪯ P ∀θ ∈ Θ .
⎪ ⎪
⎨ ⎬ (16)
⇔ j j
Hx Dl|k = Λl|k Hθk ∀j ∈ Nv1
⎪
⎩ Λj ∈ Ru×qk
⎪ ⎪
⎪
⎭ The finite horizon MPC cost function is then given by
l|k ≥0 N −1
∑
The first equivalence follows from the equivalence of x ∈ Xl+1|k JN (xk , θ̂k , vN |k ) = ∥x̂l|k ∥2Q + ∥ûl|k ∥2R + ∥x̂N |k ∥2P (17)
with Hx (x − zl+1|k ) ≤ αl+1|k 1. The second follows from the left hand l=0
side being convex in x for given θ so that therefore the inequality
where x̂l|k , ûl|k are defined recursively by
holds for all x ∈ Xl|k if and only if it holds for the vertices of
Xl|k . In the third equivalence the maximization is to be understood x̂l+1|k = A(θ̂k )x̂l|k + B(θ̂k )ûl|k , x0|k = xk ,
row-wise and the last equivalence follows from strong duality in (18)
ûl|k = K x̂l|k + vl|k .
linear programming, cf. Ben-Tal, Ghaoui, and Nemirovski (2009).
Specifically, considering row i ∈ Nu1 in the maximization, which is
finite since Θk is compact, we have that Remark 12. Note that in the proposed cost function it is not
{ }
max [Hx ]i Dl|k θ
j necessary to use ‘‘delayed’’ parameter estimates as proposed in
θ ∈Θk Di Cairano (2016), where the updated parameter is inserted only
at the end of the prediction horizon. Delayed parameter estimates
{ }
= minq max [Hx ]i Djl|k θ + λ⊤
i (hθk − Hθk θ ) allow for a simpler analysis but slow down the convergence of the
λi ∈R≥k0 θ
prediction error.
= minq λ⊤
i hθk
λi ∈R≥k0 Remark 13. Alternatively, as proposed in Lorenzen et al. (2017),
j based only on the set estimate Θk , a min–max cost
s. t. [ Hx i Dl|k
] − λi Hθk = 0.
⊤

ℓ(zl|k , αl|k , vl|k ) = max ∥Qx∥∞ + ∥Rul|k (x)∥∞

{ }
(19)
Since the minimization in the inequality constraint can be re- x∈Xl|k
moved, this concludes the proof.
can be employed, which has the advantage that the point estimate
Remark 10. The equality constraint (15d) can equivalently be θ̂k for the parameter can be omitted. With a suitable terminal cost,
written in vector form as a linear constraint and under an additional, technical assumption, practical stability
⎡ ⎤⎡ ⎤ of the closed loop has been proved. Yet, the disadvantage of the
Hx A1 Hx A1 xj Hx B1 −Iu ⊗ [Hθ⊤ ]1 zl|k approach in Lorenzen et al. (2017) is that the ultimate bound on
⎢Hx A2 Hx A2 xj Hx B2 −Iu ⊗ [Hθ⊤ ]2 ⎥ ⎢ αl|k the state depends on the set W whereas the approach proposed
⎢ ⎥⎢ ⎥
⎥ = 0. in this paper, i.e. using the point estimate θk , allows the ℓ2 norm
⎥
...
⎢ ⎥⎢
ul | k
or the state to be bounded in terms of the ℓ2 norm of the realized
⎣ ⎦⎣ ⎦
j
Hx Ap Hx Ap xj Hx Bp −Iu ⊗ [Hθ⊤ ]p vec((Λl|k )⊤ ) disturbance sequence.
466 M. Lorenzen, M. Cannon and F. Allgöwer / Automatica 103 (2019) 461–471

3.4. MPC algorithm and stability analysis time k0 + 1 and l ∈ NN0 −1 define the candidate input ũl|k0 +1 (x) =
Kx + vl∗+1|k with vN∗ |k = 0 and candidate state tube X̃l|k0 +1 =
0 0
Having derived tractable reformulations for the parameter esti- X∗l+1|k0 . Since X̃N −1|k0 +1 ⊆ Xf and Θk0 ⊆ Θ , by Assumption 11,
mation and optimal control problem, we summarize the adaptive
there exists X̃N |k0 +1 ⊆ Xf satisfying A(θ )x ⊕ B(θ )ũN −1|k0 +1 (x) ⊕
MPC algorithm and provide a brief analysis of the relevant control
W ⊆ X̃N |k0 +1 for all x ∈ X̃N −1|k0 +1 , θ ∈ Θk0 . By construction
theoretic properties.
In order to simplify notation, we denote the decision variables {ũl|k0 +1 , X̃l|k0 +1 }l∈NN −1 satisfy the constraints (5b), (5c), and since
0
in the online optimization program by dN |k = {zN |k , αN |k , vN |k , xk0 +1 = A(θ ∗ )xk0 +B(θ ∗ )uk0 +wk0 ∈ X∗1|k = X̃0|k0 +1 , constraint (5a)
0
j
ΛN |k } with ΛN |k = {Λl|k }j∈Nv ,l∈NN −1 . Given xk and Θk , the set of is satisfied. By Proposition 9 this is equivalent to feasibility of (15)
1
admissible decision variables is
0
and hence D(xk0 +1 , Θk0 ) ̸ = ∅, which implies D(xk0 +1 , Θk0 +1 ) ̸ = ∅
as Θk0 +1 ⊆ Θk0 .
D(xk , Θk ) = {dN |k | (15), (zN |k , αN |k ) ∈ Xf }.
Claim : If xk0 is bounded, the optimal solution v0∗|k and hence uk0
0
The adaptive MPC algorithm can be summarized as follows. and xk0 +1 are well defined and bounded (Rawlings et al., 2017). The
Offline: Choose cost matrices Q and R. Determine a robustly stabi- claim then follows from Lemmas 2 and 5.
lizing feedback gain K , terminal set Xf according to Assumption 11,
Claim is a direct corollary of D(xk , Θk ) ̸ = ∅ and Proposition 9.
and terminal cost matrix P satisfying (16).
To prove the finite ℓ2 gain note that vl|k+1 = vl∗+1|k for l ∈ N0N −2
Online: For each time step k = 0, 1, 2, . . .
and vN −1|k+1 = 0 denotes a feasible input sequence at time k + 1.
(i) Measure the state xk . Let {x̂l|k }l∈NN and {x̂l−1|k+1 }l∈NN denote the corresponding predicted
1 1
(ii) If k > 0 update the membership set Θk and point estimate state trajectories, which evolve according to (18) with initial con-
θ̂k according to (7) or (9), and (12), respectively. ditions xk and xk+1 , respectively, and denote the difference by
(iii) Determine the minimizer of the linearly constrained
δx̂l|k = x̂l−1|k+1 − x̂l|k
quadratic program
= Acl (θ̂k+1 )l−1 (wk + x̃1|k )
d∗N |k = arg minJN (xk , θ̂k , vN |k )
dN |k l−1
(20) ∑
s. t. dN |k ∈ D(xk , Θk ). + Acl (θ̂k+1 )l−1−i D(x̂i|k , ûi|k )(θ̂k+1 − θ̂k ).
i=1
(iv) Apply
By non-expansiveness of the projection operator, we have ∥θ̂k+1 −
uk = Kxk + v0∗|k . (21) θ̂k ∥ ≤ ∥θ̃k+1 − θ̂k ∥, and using (12) together with
1
> sup(x,u)∈Z
µ
The following theorem establishes recursive feasibility of the ∥D(x, u)∥2 this leads to
adaptive MPC algorithm, consistent parameter estimation, robust ( l−1 )
satisfaction of the constraints (4), and finite gain ℓ2 stability of the l−i
∑
∥δx̂l|k ∥ ≤ ∥Acl (θ̂k+1 ) ∥ ∥x̃1|k + wk ∥.
closed loop. Recursive feasibility guarantees a well-defined control i=0
law; if there exists a solution to the optimization program for
a given initial condition, a solution to the optimization program The claim then follows by a standard argument. With Q̄ =
exists for all future states resulting from the application of the Q + K ⊤ RK consider
proposed MPC control law. Consistent parameter estimation im- VN (xk+1 , θ̂k+1 , Θk+1 ) − VN (xk , θ̂k , Θk )
plies that the true parameter is always contained in the estimated
parameter set and the estimated parameter is consistent with the ≤ JN (xk+1 , θ̂k+1 , ṽN |k ) − VN (xk , θ̂k , Θk )
initially given parameter set. N −2
∑
≤ − ∥xk ∥2Q − ∥uk ∥2R + ∥x̂l|k+1 ∥2Q + ∥ûl|k+1 ∥2R
Theorem 14 (Closed-loop Properties). Suppose Assumptions 1 and 11 l=0
are satisfied. If θ ∗ ∈ Θk0 , D(xk0 , Θk0 ) ̸ = ∅, then for all k ≥ k0 (N −1
∑
)
+ ∥x̂N −1|k+1 2P ∥ − ∥x̂l|k 2Q
∥ + ∥ûl|k 2R
∥ + ∥x̂N |k 2P∥
(i) D(xk , Θk ) ̸ = ∅
(ii) θ ∗ ∈ Θk , θ̂k ∈ Θ l=1
(23)
N −1
(iii) xk × uk ∈ Z. ∑
xk 2Q ε ∥x̂l|k ∥2Q + ∥ûl|k ∥2R
uk 2R
( )
≤−∥ ∥ −∥ ∥ +
Moreover, the system (1) in closed loop with the proposed MPC control l=1
law u(xk ) = Kxk + v0∗|k is finite gain ℓ2 stable, i.e., there exist constants N −1 ( )
c0 , c1 , c2 ∈ R>0 such that for all K ∈ N
∑ 1
+ ε∥x̂N |k ∥2P + 1+ ∥δx̂l|k ∥2Q̄ + ∥δx̂N |k ∥2P
K K
ε
∑ ∑ l=1
∥xk ∥2 ≤ c0 ∥x0 ∥2 + c1 ∥θ̂0 − θ ∗ ∥2 + c2 ∥wk ∥2 . (22) ≤ − ∥ ∥ + ε VN (xk , θ̂k , Θk ) + cA ∥x̃1|k + wk ∥2
xk 2Q
k=0 k=0
≤ − c ∥xk ∥2 + cA ∥x̃1|k + wk ∥2 .
Remark 15. Note that, unlike nominal MPC, the controller is not The third inequality follows by Cauchy–Schwarz and Young’s in-
a static state feedback, but a dynamic controller with states θ̂k and equality, which implies ∥x̂l−1|k+1 ∥2Q = ∥x̂l|k + δx̂l|k ∥2Q ≤ (1 +
Θk . The term c0 ∥x0 ∥2 + c1 ∥θ̂0 −θ ∗ ∥2 bounds the possible overshoot ε )∥x̂l|k ∥2Q + (1 + 1ε )∥δ xl|k ∥2Q . Since for each parameter set Θ and
due to the initial condition of the plant and the controller.
θ ∈ Θ , VN (x, θ, Θ ) is a continuous, piecewise quadratic func-
tion in x, cf. Bemporad, Morari, Dua, and Pistikopoulos (2002),
Proof. We prove claim and for k = k0 + 1; since k0 is arbitrary
it can be upper bounded by a quadratic function on the feasible
the results follow by induction. set. In particular for each θ ∈ Θ , there exists cθ , such that
Claim : At time k0 , for l ∈ NN0 −1 , let ul|k0 (x) = Kx + vl∗|k and X∗l+1|k VN (x, θ, Θ ) = JN (x, θ, v∗N (x, θ, Θ )) < cθ ∥x∥2 , where v∗N is the
0 0
be a feasible input and admissible state tube trajectory satisfying optimal solution of (20) with xk = x, θ̂k = θ , Θk = Θ . Since
the MPC constraints (5) and terminal constraint XN |k0 ∈ Xf . For JN (x, ·, v∗N (x, θ, Θ )) is continuous, we can choose an ε > 0 such
 M. Lorenzen, M. Cannon and F. Allgöwer / Automatica 103 (2019) 461–471 467

that JN (x, θ̃ , v∗N (x, θ , Θ )) ≤ cθ ∥x∥2 for each θ̃ ∈ Bε (θ ) and each Thus, analogous to Proposition 9, the admissible prediction tube for
θ ∈ Θ . By compactness of Θ , there exists a finite collection {θ i }i∈I systems with time-varying parameters, including the parameter
with θ i ∈ Θ such that ∪i∈I Bε (θ i ) ⊇ Θ . Hence, JN (xk , θ̂k , v∗N |k ) ≤ set prediction (26), can be reformulated as linear constraints.
ĉθ ∥x∥2 with ĉθ = maxi∈I cθ i . Thus there exist suitable constants
ε, c , cA ∈ R>0 such that the last two inequalities hold. The final Proposition 18 (Prediction Tube). Let {Xl|k }l∈NN be parametrized as
0
result follows by summing over k and using Lemma 5 and again in (14) with decision variables zN |k , αN |k , and vN |k . Without loss of
Young’s inequality for ∥x̃1|k + wk ∥2 . q
generality, assume ∥[Hθ ]i ∥ = 1, ∥[Hθk ]i ∥ = 1 for all i ∈ N1k .
The prediction tube equations
Corollary 16. Suppose ∑∞ Assumptions 1 and 11 are satisfied and
D(x0 , Θ0 ) ̸ = ∅. If 2
< ∞, then limk→∞ xk = 0 for X0|k ∋ xk ,
k=0 ∥wk ∥ (27)
the system (1) in closed loop with the proposed MPC control law Xl+1|k ⊇ Acl (θ )Xl|k ⊕ B(θ )vl|k ⊕ W ∀θ ∈ Θl+k
u(xk ) = Kxk + v0∗|k .
are satisfied if, for all j ∈ Nv1 , l ∈ NN0 −1 , there exists Λl|k ∈ R≥0
j u×qk
such
We have proved robust constraint satisfaction and stability for that
a computationally tractable MPC algorithm with online parameter
adaption. − Hx z0|k − α0|k 1 ≤ −Hx xk (28a)
j j
Λl|k hθl|k + Hx dl|k − αl+1|k 1 ≤ −w̄ (28b)
4. Extensions
j j
Hx Dl|k = Λl|k Hθl|k (28c)
Being based on a modern state-space formulation, the proposed
MPC scheme with parameter adaption can be directly combined with
[ ] [ ]
with recent results in the MPC literature, e.g. output feedback Hθk ′ hθk + ldθ 1 ′
or offset-free tracking. In the following, we present two relevant Hθl|k = ∈ Rqk ×p , hθl|k = ∈ Rqk .
Hθ hθ
modifications of the basic algorithm, one for time-varying pa-
rameters and one for persistently exciting regressors to achieve Proof. The proof follows analogously to the case with non time-
parameter convergence. varying parameters and by noting that {θ ∈ Rp | Hθl|k θ ≤ hθl|k } =
Dl (Θk ) ∩ Θ = Θl|k ⊆ Θl+k .
4.1. Time-varying parameters
Xl+1|k ⊇ Acl (θ )Xl|k ⊕ B(θ )vl|k ⊕ W ∀θ ∈ Θl+k
We extend the admissible prediction tube and constraint re- ⇐ Xl+1|k ⊇ Acl (θ )Xl|k ⊕ B(θ )vl|k ⊕ W ∀θ ∈ Θl|k
formulation to include time-varying parameters θk∗ , i.e., to include
Hx (Acl (θ )(zl|k + αl|k xj ) + B(θ )vl|k − zl+1|k )
time-varying systems ⇔
− αl+1|k 1 ≤ −w̄ ∀j ∈ Nv1 , θ ∈ Θl|k
xk+1 = A(θk∗ )xk + B(θk∗ )uk + wk (24) { }
⇔ max Hx Djl|k θ + Hx djl|k − αl+1|k 1 ≤ −w̄ ∀j ∈ Nv1
under the following assumption. θ ∈Θl|k
⎧ j j
Λl|k hθl|k + Hx dl|k − αl+1|k 1 ≤ −w̄ ⎪
⎫
Assumption 17 (Time-varying Parameters). There exists dθ ∈ R,
⎪
⎪ ⎪
⎨ ⎬
j j
such that the parameter vector θk∗ satisfies θk∗ ∈ Θ for all k ∈ N ⇔ H D
x l|k = Λ H
l|k θl|k ∀j ∈ Nv1
and ⎪
⎩ j
⎪ u×q ′ ⎪
⎪
Λl|k ∈ R≥0 k
⎭
∥θk∗+1 − θk∗ ∥ ≤ dθ .
Similar to Section 3, the parameter set update law (25) and
Assumption 17 restricts the parameter variation between con- admissible state tube description in Proposition 18 constitute a rig-
secutive time steps. Furthermore, as before, an absolute prior orous and computationally tractable framework for model adap-
bound on the parameter vector is assumed, which is necessary in tion and robust constraint satisfaction in receding horizon control
order to be able to guarantee robust (state) constraint satisfaction. algorithms. For systems with time-varying parameters, the MPC al-
Note that different rates of change of the parameters can easily gorithm presented in Section 3.4 remains the same, however with
be incorporated by a weighting matrix or a suitable scaling of the
the update of the membership set Θk in step (ii) replaced by (25)
model.
and the admissible set D(xk , Θk ) in optimization (20) replaced by
To incorporate the dynamics of θk∗ , the parameter set update (7)
needs to be changed. To this end, we introduce a dilation operator Dt (xk , Θk ) = {dN |k | (15a), (28), (zN |k , αN |k ) ∈ Xf }.
Dl , which is defined for l ∈ N, dθ ∈ R>0 , and a polytope Θ = {θ ∈
Rp | Hθ θ ≤ hθ } by The terminal set is equivalent to the non time-varying case, i.e.
based on the entire parameter set Θ and assumed to satisfy
Dl (Θ ) = {θ ∈ Rp | Hθ θ ≤ hθ + ldθ 1} Assumption 11.
Using Dl and assuming (without loss of generality) ∥[Hθk ]i ∥ = 1 for The following proposition summarizes the relevant closed-loop
q
all i ∈ N1k , the online parameter set estimation (7) becomes properties, analogously to Theorem 14.

Θk = D1 (Θk−1 ∩ ∆k ) ∩ Θ . (25) Proposition 19 (Closed-loop Properties). Suppose Assumptions 1, 11,

17 are satisfied. If θk∗ ∈ Θk0 and Dt (xk0 , Θk0 ) ̸ = ∅, then for all k ≥ k0
Similarly, to ensure constraint satisfaction, the equation for admis- 0
we have
sible state tubes (5b) needs to be adapted. Specifically, the open-
loop prediction for the estimated parameter membership set needs (i) Dt (xk , Θk ) ̸ = ∅
to be changed from Θk to (ii) θk∗ ∈ Θk ,
(iii) xk × uk ∈ Z
Θl|k = Dl (Θk ) ∩ Θ . (26)
for the system (24) in closed loop with the proposed MPC control law
Note that Dl is easily implemented for sets that are given in
half space form as only the right hand side needs to be changed. u(xk ) = Kxk + v0∗|k .
468 M. Lorenzen, M. Cannon and F. Allgöwer / Automatica 103 (2019) 461–471
∑q
Proof. As before we prove the claim by induction. equal to i=1 φi φi⊤ . If a regressor is PE, it has been proved that, un-
der suitable assumptions on the disturbance, gradient algorithms
Claim : First note that Θl|k+1 ⊆ Θl+1|k since Θl|k+1 = Dl (D1 (Θk ∩
and the set-membership estimator converge to the true parameter.
∆k+1 ) ∩ Θ ) ∩ Θ ⊆ Dl (D1 (Θk )) ∩ Θ = Dl+1 (Θk ) ∩ Θ = Θl+1|k .
Let {u∗l|k , X∗l+1|k }l∈NN −1 be an admissible input and state tube. For
0 Proposition 22 (Convergence of θ̂k ). Assume wk ≡ 0. The true
l ∈ NN0 −2 consider the candidate solution ṽl|k+1 = vl∗+1|k , X̃l+1|k+1 = parameter θ ∗ is a globally exponentially stable fixed point of the
X∗l+2|k and ṽN −1|k+1 = 0. For l ∈ NN0 −2 we have difference equation (12) if the regressor {Dk } is persistently exciting.

X∗l+2|k ⊇ Acl (θ )X∗l+1|k ⊕ B(θ )vl∗+1|k ⊕ W ∀θ ∈ Θl+1|k Proposition 23 (Convergence of Θk ). Assume {wk } are realizations of
⇔X̃l+1|k+1 ⊇ Acl (θ )X̃l|k+1 ⊕ B(θ )ṽl|k+1 ⊕ W ∀θ ∈ Θl+1|k a sequence of independent random variables {Wk } with support W.
With probability one as k → ∞
⇒X̃l+1|k+1 ⊇ Acl (θ )X̃l|k+1 ⊕ B(θ )ṽl|k+1 ⊕ W ∀θ ∈ Θl|k+1
sup ∥θ1 − θ2 ∥ → 0
where the last implication follows from Θl|k+1 ⊆ Θl+1|k . Further- θ1 ,θ2 ∈Θk
more, by Assumption 11 there exists XN |k+1 satisfying the last set if the regressor {Dk } is persistently exciting.
inclusion with l = N − 1. Since (x, Kx + vl∗|k ) ∈ Z for all x ∈ X∗l|k ,
this proves the claim for k + 1. The proof of Proposition 22 follows classical results, e.g. Åström
and Wittenmark (2008, Theorem 6.4), with, as previously, a minor
Claim : Let θk ∈ Θk and define δθk = θk∗+1 − θk∗ . Let Θk ∩ ∆k+1 =
∗
adaption due to the matrix notation and projection. The proof of
{θ Rp | H̃θk θ ≤ h̃θk }. Then H̃θk θk∗+1 = H̃θk θk∗ + H̃θk δθk ≤ h̃θk + dθ 1, Proposition 23 follows Bai et al. (1998, Theorem 2.1) with again a
hence θk∗+1 ∈ D1 (Θk ∩ ∆k+1 ). Since by assumption θk∗+1 ∈ Θ the minor difference due to the matrix notation and vector valued dis-
result follows. turbance. The assumption on the disturbance could be weakened
to the boundary of W being a subset of the support of Wk .
Claim is again a direct corollary of Dt (xk , Θk ) ̸= ∅ and
The following result enables the application of the previous two
Proposition 18.
propositions within the proposed MPC framework, by providing
The parameter variation leads to an additional tracking error in sufficient conditions on the input such that the regressor D⊤ k is
the LMS filter, invalidating the previously proved ℓ2 gain. Hence, persistently exciting.
the certainty equivalence approach as presented in this paper
should only be employed for ‘slow’ (Narendra & Annaswamy, 2005) Lemma 24 (Persistent Excitation). Suppose Assumption 20 is satisfied
parameter variations dθ and combined with a persistently exciting and wk ≡ 0. The regressor {D⊤ k } is persistently exciting if the stacked
input. Alternatively, this difficulty could be circumvented by using input vector {u⊤
k = [u ⊤
k . . . u ⊤
k−n ]} is persistently exciting.
the min–max cost (19) as presented in Lorenzen et al. (2017). The proof of Lemma 24 can be found in the Appendix. With
Lemma 24 different results to achieve a PE input in MPC can be
4.2. Persistent excitation directly incorporated into the presented setup, e.g. augmenting the
cost function (Heirung et al., 2017), introducing an additional con-
The property of persistent excitation (PE) is central to classi- straint (Marafioti et al., 2014) or a two-step procedure (Tanaskovic
cal adaptive control design (Narendra & Annaswamy, 2005). It is et al., 2014).
well known that a lack of excitation can lead for example to a We conclude this section by remarking that instead of PE, a
drift in the parameter estimates, resulting in bursts or oscillatory min–max cost could be used to obtain optimal Dual Control (Feld-
behavior of the closed-loop system. Similar to the classical self- baum, 1961; Veres, 1995). However, although this leads to optimal
tuning regulator (Åström & Wittenmark, 1973), for the presented inputs for a given cost function, it does not necessarily guarantee
adaptive MPC algorithm a PE condition is not necessary to prove parameter convergence and generally leads to highly intractable,
stability. Yet it ensures convergence of the parameter estimates, non-convex optimization programs.
thereby improving closed-loop performance compared to a robust
MPC approach. In the following, we give a definition of persistent 5. Numerical example
excitation, provide two relevant results on parameter convergence,
and show how the assumption on PE can be verified in the setup In this section, examples are presented to illustrate the advan-
considered in this work. tages of the proposed adaptive MPC scheme. We first demonstrate
the online parameter identification and constraint satisfaction in a
Throughout this section, we make the following linear inde-
setup where stabilization of the origin is considered. Thereafter,
pendence assumption on the parameters, which is necessary for
we extend the problem setup and discuss convergence of the
uniquely estimating the parameters. If Assumption 20 is violated,
parameter estimates with a persistently exciting input as well as
one or more parameters can be equivalently expressed by a linear
time-varying parameters.
combination of other parameters. Note that it can be rewritten as
a matrix rank condition and checked easily. Example 1 Consider the second-order discrete-time linear system
of the form (1) with randomly generated uncertainty matrices
A1 , A2 , B3
∑p
Assumption 20. The equation i=1 λi [Ai Bi ] = 0 has only the
trivial solution λi = 0. 0.5 0.2
[ ] [ ]
0
A0 = , B0 = ,
−0.1 0.6 0.5
Definition 21 (Persistent Excitation). A regressor {Φk } with Φk ∈
0.042 0.015 0.019
[ ] [ ]
Rp×q is persistently exciting if there exist positive constants α , β , A1 =
0
, A2 = , A3 = 02×2 ,
P such that for all k0 0.072 0.03 0.009 0.035
0.040
[ ]
k0 +P −1
{Bi }i=1,2 = 02×1 , B3 = ,
0.054
∑
αI ⪯ Φk Φk⊤ ⪯ β I .
k=k0
Θ = {θ ∈ R3 | ∥θ∥∞ ≤ 1}, θ̂0 = 0, θ ∗ = [0.8 0.2 −0.5]⊤ ,
This definition generalizes the more standard definition of a and an iid disturbance sequence with wk uniformly distributed on
regressor φk ∈ Rp ; note that for Φ = [φ1 . . . φq ] the term Φk Φk⊤ is W = {w ∈ R2 | ∥w∥∞ ≤ 0.1}.
 M. Lorenzen, M. Cannon and F. Allgöwer / Automatica 103 (2019) 461–471 469

Fig. 1. Realized closed-loop trajectory from initial condition x0 = [2 3]⊤ , predicted

state tube at time k = 0, and constraint [xk ]2 ≥ −0.3.

The MPC parameters were set to horizon length N = 10, cost Fig. 2. Closed-loop state and input trajectory with enforced PE input (solid line),
state and input constraints (dashed line).
weights
1.467 0.207
[ ] [ ]
1 0
Q = , R = 1, P =
0 1 0.207 1.731

and prestabilizing feedback gain K = [0.017 − 0.41]. Separate

state and input constraints [xk ]2 ≥ −0.3, |uk | ≤ 1, which should
be satisfied robustly, were applied to the system. An additional
box constraint ∥xk ∥∞ ≤ 3 has been introduced to determine the
parameter update gain µ.
Fig. 1 shows the state constraint [xk ]2 ≥ −0.3, a closed-loop
state trajectory for a single disturbance sequence starting from
initial condition x0 = [2 3]⊤ , and the predicted state tube at time
k = 0. As proven in Theorem 14, the state constraint is satisfied
robustly for all possible predicted states and the state converges to
a neighborhood of the origin. Similarly, the input constraints (not
plotted) are satisfied for all k ∈ N.
To highlight the parameter estimation, the PE condition as de-
scribed in Section 4.2 has been implemented, following Marafioti
et al. (2014), via an additional constraint on the input Fig. 3. Parameter membership set at time steps k ∈ {0, 5, 25, 70, 120, 500}.
P −1
∑
k−l ⪰ α I
uk−l u⊤ (29)
l=0

with P = n + 1 and α = 2. Starting from an initial condition x0 =

[0 0]⊤ , the closed loop exhibits a persistently exciting regressor,
with the typical cyclic state and input (Fig. 2). Due to the state
constraint, the center of the trajectory path is shifted to the positive
orthant, such that the closed-loop state trajectory does not violate
the constraint [xk ]2 ≥ −0.3. As predicted by Proposition 23, the
parameter membership set converges to a singleton (Fig. 3). Given
the realized state and input trajectory, falsified parameters are
removed and the uncertainty set is non-increasing.
Finally, to demonstrate the capability of handling time-varying
systems, in the following, the problem setup has been changed to
a time-varying parameter θk∗ with θ0∗ = θ ∗ and a bound on the
variation of ∥θk∗+1 − θk∗ ∥ ≤ 0.01. In the simulation, the parameter
has been taken to be a periodic deterministic function in time. Each
parameter is increased/decreased linearly by 0√.01 , i.e.
3 Fig. 4. Parameter membership set for the system with time-varying parameters at
time steps k ∈ {0, 100, 200, 300, 400, 500}.
0.01
[θk∗+1 ]i = [θk∗ ]i ± √ ,
3
where the sign is changed upon hitting the boundary of Θ . As of convergence to a singleton as in Fig. 3, the parameter set varies
above, the simulation has been initialized with x0 = [0 0]⊤ and the in position, shape, and size.
additional PE constraint (29). Fig. 4 shows the estimated parameter The simulations were performed in Matlab with Yalmip for set-
set at sampling times k = 1, 100, 200, 300, 400 and 500. Instead ting up the optimization program, which was solved using MOSEK.
470 M. Lorenzen, M. Cannon and F. Allgöwer / Automatica 103 (2019) 461–471

The median solver time (with PE constraint) reported by Yalmip sequence converges in finite time such that Xf = Xif for some i ∈ N
was 0.068 s (0.10 s) with a maximum of 0.095 s (0.19 s) and mini- satisfying Xif = Xfi+1 .
mum of 0.05 s on an Intel Core i7 with 3.4 GHz. Choosing X0 , i.e. the With {θ k }k∈Nvp being the vertices of the set Θ , [h̄kx ]i =
shape of the tube cross sections, to be the minimal robustly forward 1

invariant set under the local control law decreases conservatism maxx∈X0 [Hx ]i Acl (θ )x, and k

and can increase performance. Yet, due to the number of equality (z , α, z + , α + ) |

⎧ ⎫
constraints in the MPC optimization program, a significant increase
⎪
⎪ ⎪
⎪
⎨ (z , α ) ∈ X 0 , (z + , α + ) ∈ X i
⎪ ⎪
⎬
in computation time was observed with increasing complexity of X̃if+1 =
f f
,
Hx Acl (θ )z − z + h̄x α − 1α ≤ −w̄ ⎪
k + k +
[ ]
X0 . Furthermore, note that the scalar input allows the decomposi- ⎪
⎪ ⎪
tion of the PE input constraint into two linear constraints, leading v
⎪ ⎪
∀k ∈ N1p
⎩ ⎭
to two convex QP problems to be solved and compared in each MPC
iteration (Marafioti et al., 2014). the recursion (A.1) can be computed by Xif+1 = Projn+1 (X̃if+1 )
where Projn+1 is the projection onto the first n + 1 coordinates.
6. Conclusions As the projection of polytopes can be computationally demand-
ing, the recursion can be simplified through setting z = z + = 0 and
A computationally tractable model predictive control algorithm determining only a suitable α satisfying Assumption 11.
with recursive parameter update has been presented that provides
guarantees for closed-loop stability and robust constraint satisfac- A.2. Proof of Lemma 24
tion. The requirements for stability and constraint satisfaction are
considered separately. This leads to a set-membership parameter Proof (Lemma 24). Let xk be the solution of (1) with wk ≡ 0.
estimation scheme being employed to derive bounds on the state By Green and Moore (1986, Corollary 2.4) {uk }k being PE implies
k , uk ] }k being PE. Let dk = [Dk ]i ,
i
and input predictions whereas a Least Mean Squares filter is used {[x⊤ ⊤ ⊤

to achieve a finite gain from the disturbance to the state. The

[A1 B1 ]i k0 +P −1
[ ]
online optimization to be solved is a linearly constrained quadratic
∑ [x ] [
Ξi = , and Sk0 = k
k .
]
··· x⊤
k u⊤
program and proven to be recursively feasible. Two numerical uk
[Ap Bp ]i k=k0
examples are provided to demonstrate the effectiveness of the
proposed algorithm. Then
Extensions for time-varying parameters and for PE regressors k0 +P −1 k0 +P −1
n
∑ ∑ ∑ ⊤
are discussed explicitly. As the MPC scheme is formulated in a D⊤ dik dik
k Dk =
modern state-space framework, the proposed setup provides a k=k0 k=k0 i=1
solid framework for adaptive MPC algorithms and can be easily k0 +P −1
n [ ]
combined with further results tailored to specific control objec- ∑ ∑ xk [ ⊤
Ξi k Ξi
]
= xk u⊤
tives, e.g., tracking or output feedback MPC. uk
k=k0 i=1
Compared to the classical adaptive control literature, the as-
n
sumptions made are necessarily more restrictive in order to allow ∑
a robust MPC formulation. The use of less restrictive assumptions = Ξi Sk0 Ξi .
in combination with soft constraints or chance constraints are i=1

currently under investigation. In particular for the time-varying Let σ and σ̄ be the smallest and largest singular value of
case, it would furthermore be of interest to derive bounds on the
[A1 B1 ]1 . . . [A1 B1 ]n
[ ]
estimation error, which could then be used to relax Assumptions 8
··· ,
and 11 to a parameter dependent prestabilizing feedback and
[Ap Bp ]1 . . . [A1 B1 ]n
terminal constraint. Finally, questions on optimal excitation of the
system as discussed in Feldbaum (1961) remain open for future which are nonzero by Assumption 20, and define∑α ′ = σ 2 α and
n
research. β ′ = σ̄ 2 β . Given α I ⪯ Sk0 ⪯ β I, we have α ′ I ⪯
i=1 Ξi Sk0 Ξi ⪯
β ′ I, which proves the claim.
Appendix
References

Adetola, V., & Guay, M. (2011). Robust adaptive MPC for constrained uncertain
A.1. Computation of the terminal region
nonlinear systems. International Journal of Adaptive Control and Signal Processing,
25(2), 155–167.
As shown in Blanchini and Miani (2015) and Raković and Cheng Åström, K. J., & Wittenmark, B. (1973). On self tuning regulators. Automatica, 9(2),
(2013), a terminal set Xf satisfying Assumption 11 can be com- 185–199.
Åström, K. J., & Wittenmark, B. (2008). Adaptive control (2nd ed.). Dover Publications.
puted recursively by the following algorithm. With X0 and f¯ as
Aswani, A., Gonzalez, H., Sastry, S. S., & Tomlin, C. (2013). Provably safe and robust
given above, i.e., X0 = {x ∈ Rn | Hx x ≤ 1} and [f¯ ]i = maxx∈X0 [F + learning-based model predictive control. Automatica, 49(5), 1216–1226.
GK ]i x, let Bai, E.-W., Cho, H., & Tempo, R. (1998). Convergence properties of the membership
set. Automatica, 34(10), 1245–1249.
X0f = {(z , α ) ∈ Rn × R≥0 | (F + GK )z + α f¯ ≤ 1} Bayard, D. S., & Schumitzky, A. (2010). Implicit dual control based on particle
filtering and forward dynamic programming. International Journal of Adaptive
and define Control and Signal Processing, 24(3), 155–177.
Bemporad, A., Morari, M., Dua, V., & Pistikopoulos, E. N. (2002). The explicit linear
⎧ ⎫
⏐ ∃(z + , α + ) ∈ Xi s. t.
⎪
⎨ ⏐ f ⎪
⎬ quadratic regulator for constrained systems. Automatica, 38(1), 3–20.
Xif+1 = (z , α ) ⏐ Acl (θ )({z } ⊕ α X0 ) ⊕ W ∩ X0f . (A.1) Ben-Tal, A., Ghaoui, L. E., & Nemirovski, A. (2009). Princeton series in applied mathe-
⏐
matics, Robust optimization. Princeton University Press.
⊆ {z } ⊕ α X0 ∀θ ∈ Θ
⎪
⎩ ⏐ + +
⎪
⎭
Bitmead, R. R., Gevers, M., & Wertz, V. (1990). Adaptive optimal control, the thinking
Man’s GPC. Prentice Hall.
The sets Xif , i ∈ N are non-increasing with i and the terminal set Blanchini, F., & Miani, S. (2015). Systems & control: Foundations & applications, Set-
is given by the limit for i → ∞. Under the given assumptions, the theoretic methods in control (2nd ed.). Birkhäuser Boston.
 M. Lorenzen, M. Cannon and F. Allgöwer / Automatica 103 (2019) 461–471 471

Bujarbaruah, M., Zhang, X., & Borrelli, F. (2018). Adaptive MPC with chance con- Tanaskovic, M., Fagiano, L., Smith, R., & Morari, M. (2014). Adaptive receding horizon
straints for FIR systems. ArXiv e-prints, abs/1804.09790. control for constrained MIMO systems. Automatica, 50(12), 3019–3029.
Chisci, L., Garulli, A., Vicino, A., & Zappa, G. (1998). Block recursive parallelotopic Tanaskovic, M., Sturzenegger, D., Smith, R., & Morari, M. (2017). Robust adaptive
bounding in set membership identification. Automatica, 34(1), 15–22. model predictive building climate control. In Proceedings of the IFAC world
Di Cairano, S. (2016). Indirect adaptive model predictive control for linear systems congress (pp. 1871–1876). Toulouse, France.
with polytopic uncertainty. In Proceedings of the American control conference Veres, S. M. (1995). Adaptive control by worst-case duality. In C. Bányász (Ed.), 5th
(pp. 3570–3575). Boston, MA. IFAC symposium on adaptive systems in control and signal processing (pp. 73–78).
Feldbaum, A. A. (1961). Dual-control theory I. Automation and Remote Control, 21, Pergamon, Oxford: IFAC Postprint Volume.
874–880, Translated from Avtomatlka i Telemekhanika, 21(9), 1960. Veres, S. M., Messaoud, H., & Norton, J. P. (1999). Limited-complexity model-
Fogel, E., & Huang, Y. (1982). On the value of information in system identification – unfalsifying adaptive tracking-control. International Journal of Control, 72(15),
Bounded noise case. Automatica, 18(2), 229–238. 1417–1426.
Green, M., & Moore, J. B. (1986). Persistence of excitation in linear systems. Systems Veres, S., & Norton, J. (1993). Predictive self-tuning control by parameter bounding
& Control Letters, 7(5), 351–360. and worst-case design. Automatica, 29(4), 911–928.
Guay, M., Adetola, V., & DeHaan, D. (2015). Control, robotics and sensors, Robust and
adaptive model predictive control of nonlinear systems. Institution of Engineering
and Technology.
Hassibi, B., Sayed, A. H., & Kailath, T. (1993). LMS is H ∞ optimal. In Proceedings of the Matthias Lorenzen studied Engineering Cybernetics at
conference on decision and control, Vol. 1 (pp. 74–79). the University of Stuttgart from 2007 to 2013. During his
Heirung, T. A. N., Foss, B., & Ydstie, B. E. (2015). MPC-based dual control with online studies, he spent one year as DAAD fellow at Harvard
experiment design. Journal of Process Control, 32, 64–76. University and four months at Daimler AG for an intern-
Heirung, T. A. N., Ydstie, B. E., & Foss, B. (2017). Dual adaptive model predictive ship in the research department Environment Perception.
control. Automatica, 80, 340–348. In 2013 he joined the Institute for Systems Theory and
Kim, T. H., & Sugie, T. (2008). Adaptive receding horizon predictive control for Automatic Control as research and teaching assistant and
constrained discrete-time linear systems with parameter uncertainties. Inter- is a doctoral student under Prof. Frank Allgöwer in the
national Journal of Control, 81(1), 62–73. Graduate School Simulation Technology within Stuttgart
Klenske, E. D., Zeilinger, M. N., Schölkopf, B., & Hennig, P. (2016). Gaussian process- Research Centre for Simulation Technology. His main re-
based predictive control for periodic error correction. IEEE Transactions on search interests are stochastic and robust Model Predic-
Control Systems Technology, 24(1), 110–121. tive Control.
Kouvaritakis, B., & Cannon, M. (2016). Model predictive control: Classical, robust and
stochastic. Cham, Switzerland: Springer International Publishing.
Mark Cannon received an M.Eng. in Engineering Science
Langson, W., Chryssochoos, I., Raković, S. V., & Mayne, D. (2004). Robust model
in 1993, a D.Phil. in Control Engineering in 1998, both
predictive control using tubes. Automatica, 40(1), 125–133.
from the University of Oxford, and received the M.S. in
Lorenzen, M., Allgöwer, F., & Cannon, M. (2017). Adaptive model predictive control
Mechanical Engineering in 1995 from Massachusetts In-
with robust constraint satisfaction. In Proceedings of the IFAC world congress
stitute of Technology. He is currently an associate profes-
(pp. 3368–3373). Toulouse, France.
sor of Engineering Science, Oxford University and a fellow
Marafioti, G., Bitmead, R. R., & Hovd, M. (2014). Persistently exciting model predic-
of St. John’s College, Oxford.
tive control. International Journal of Adaptive Control and Signal Processing, 28(6),
536–552.
Mayne, D. Q. (2014). Model predictive control: Recent developments and future
promise. Automatica, 50(12), 2967–2986.
Narendra, K. S., & Annaswamy, A. M. (2005). Stable adaptive systems. Dover Publica-
tions. Frank Allgöwer studied Engineering Cybernetics and Ap-
Ostafew, C. J., Schoellig, A. P., & Barfoot, T. D. (2015). Conservative to confident: plied Mathematics in Stuttgart and at the University of
Treating uncertainty robustly within learning-based control. In Proceedings of California, Los Angeles (UCLA), respectively, and received
the IEEE international conference on robotics and automation (pp. 421–427). his Ph.D. degree from the University of Stuttgart in Ger-
Qin, S., & Badgwell, T. A. (2003). A survey of industrial model predictive control many. Since 1999 he is the Director of the Institute for
technology. Control Engineering Practice, 11(7), 733–764. Systems Theory and Automatic Control and professor at
Raković, S. V., & Cheng, Q. (2013). Homothetic tube MPC for constrained linear the University of Stuttgart. His research interests include
difference inclusions. In Proceedings of the 25th Chinese control and decision networked control, cooperative control, predictive con-
conference (pp. 754–761). trol, and nonlinear control with application to a wide
Raković, S. V., Kouvaritakis, B., Findeisen, R., & Cannon, M. (2012). Homothetic tube range of fields including systems biology. For the years
model predictive control. Automatica, 48(8), 1631–1638. 2017–2020 Frank serves as President of the International
Rawlings, J. B., Mayne, D. Q., & Diehl, M. M. (2017). Model predictive control theory Federation of Automatic Control (IFAC) and since 2012 as Vice President of the
and design (2nd ed.). Nob Hill Publishing. German Research Foundation DFG.