Data-Driven Min-Max MPC for Linear Systems

Yifan Xie, Julian Berberich, Frank Allgöwer

Abstract— Designing data-driven controllers in the presence Standard model-based control techniques, in particular
of noise is an important research problem, in particular when standard MPC schemes, rely on a priori knowledge of
guarantees on stability, robustness, and constraint satisfaction system models that are either identified from measured data
are desired. In this paper, we propose a data-driven min-max
arXiv:2309.17307v1 [eess.SY] 29 Sep 2023

model predictive control (MPC) scheme to design state-feedback by system identification methods [8] or derived based on
controllers from noisy data for unknown linear time-invariant first principles. In contrast, data-driven control approaches
(LTI) system. The considered min-max problem minimizes the can design controllers directly from the available data. The
worst-case cost over the set of system matrices consistent with Fundamental Lemma proposed by Willems et al. [20] has
the data. We show that the resulting optimization problem can led to control strategies in a behavioral setting [11], and
be reformulated as a semidefinite program (SDP). By solving the
SDP, we obtain a state-feedback control law that stabilizes the more recently in the context of MPC [2], [5]. For example,
closed-loop system and guarantees input and state constraint [2] proposes a robust data-driven MPC scheme and proves
satisfaction. A numerical example demonstrates the validity of theoretical guarantees in case of noisy data. The behavioral
our theoretical results. framework requires persistently exciting data, enabling the
unique representation of the system from the data in the
I. I NTRODUCTION noise-free scenarios. On the other hand, it has been shown
in the informativity framework that the data need not to be
Model Predictive Control (MPC) is an advanced control
sufficiently informative to uniquely identify the system [19].
technique that can handle nonlinear multi-input multi-output
In this framework, control strategies are proposed to stabilize
systems with constraints [15]. The basic idea of MPC is to
the system based on a representation of the set of system
solve an open-loop optimal control problem at each sampling
matrices consistent with the data [6], [17]–[19]. For example,
time, which uses the current state as the initial condition and
[17] proposes a controller design method directly from noisy
system dynamics to predict future open-loop states.
input-state data. However, the design of MPC schemes,
Ensuring robust constraint satisfaction in the presence of
known for their effectiveness in handling constraints, remains
noise or model uncertainty is challenging. Several MPC
unaddressed in this framework.
methods have been proposed to deal with this issue. Tube-
In this paper, we present a data-driven min-max MPC
based robust MPC [12], [13], where a constraint tightening
scheme which uses noisy data to control linear time-invariant
is included in the MPC optimization problem, ensures that
(LTI) systems with unknown system matrices. Our approach
all possible realizations of the state trajectory lie in an
relies on a representation of the system matrices consistent
uncertainty tube around a nominal system. This approach
with a sequence of noisy input-state data that was employed,
typically assumes that the nominal system is known, whereas
e.g., by [3], [4], [17]. We reformulate the data-driven min-
the noise is unknown and bounded. In [9] and [10], model
max MPC problem with input and state constraints to a
uncertainty is characterized within a set, and additional
semidefinite program (SDP) that yields a state-feedback
measurements are leveraged to reduce model uncertainty and
control law. Further, we show that the proposed data-driven
enhance the performance of the tube-based MPC scheme.
min-max MPC guarantees closed-loop properties including
Furthermore, min-max MPC can effectively address scenar-
recursive feasibility, constraint satisfaction and exponential
ios involving model uncertainty, as discussed in [1], [7],
stability.
and [16]. The goal of min-max MPC is to design control
The remainder of this paper is organized as follows. In
inputs that minimize the worst-case cost w.r.t. disturbances
Section II, we introduce necessary preliminaries. In Sec-
or uncertainties. State-feedback control laws are commonly
tion III, we propose the data-driven min-max MPC scheme
employed in the min-max MPC framework to decrease the
and show that the scheme ensures recursive feasibility and
computational complexity [1], [7], [16].
exponential stability. We apply the developed scheme to a
F. Allgöwer is thankful that his work was funded by Deutsche
numerical example in Section IV. Finally, we conclude the
Forschungsgemeinschaft (DFG, German Research Foundation) under Ger- paper in Section V.
many’s Excellence Strategy - EXC 2075 - 390740016 and under grant
468094890. F. Allgöwer acknowledges the support by the Stuttgart Center
for Simulation Science (SimTech). The authors thank the International Max II. P RELIMINARIES
Planck Research School for Intelligent Systems (IMPRS-IS) for supporting
Yifan Xie. Let I[a,b] denote the set of integers in the interval [a, b]
Yifan Xie, Julian Berberich, Frank Allgöwer are with the Institute and I≥0 denote the set of nonnegative integers. For a matrix
for Systems Theory and Automatic Control, University of Stuttgart,
70550 Stuttgart, Germany. {yifan.xie, julian.berberich, P , we write P ≻ 0 if P is positive definite and P  0
frank.allgower}@ist.uni-stuttgart.de. if P is positive semi-definite. For a vector x and a matrix
 √
P ≻ 0, we write kxkP = x⊤ P x. For matrices A and B of B. Problem setup
 ⊤
compatible dimensions, we abbreviate ABA⊤ to AB ⋆ . In this paper, we employ data-driven min-max MPC for
the unknown system xt+1 = As xt + Bs ut to stabilize the
A. System representation origin, while the input and state satisfy given constraints. As
We consider an unknown discrete-time LTI system explained in Section II-A, the offline measurements (U− , X)
are affected by noise satisfying the instantaneous constraint
xt+1 = As xt + Bs ut + ωt , (1) in Assumption 1. On the other hand, the data collected
online during closed-loop operation are assumed to be noise-
where xt ∈ Rn denotes the state, ut ∈ Rm denotes the input, free. This is assumed for simplicity, to avoid an additional
and ωt ∈ Rn denotes the unknown noise for t ∈ N. The maximization w.r.t. the noise in the min-max MPC problem.
matrices As ∈ Rn×n and Bs ∈ Rn×m are unknown. We
Extending the proposed framework to handling online noise
define a sequence of input, noise and corresponding state is an interesting issue for future research. In Section V, we
measurements, which is denoted by
show with a numerical example that the proposed approach
produces reliable results also in the presence of online noise.
 
U− := u0 u1 . . . uT −1 ,
  We consider ellipsoidal constraints on the input and the
W− := ω0 ω1 . . . ωT −1 ,
  state, i.e.,
X := x0 x1 . . . xT .
kut kSu ≤ 1, kxt kSx ≤ 1, ∀t ∈ N,
Throughout this paper, we assume that data of the form X
and U− are available, whereas W− is unknown. The noise where Su ≻ 0 and Sx  0. The centers of the ellipsoids for
ωt should satisfy the following instantaneous constraint. the input and state constraints are at the origin, but our results
Assumption 1: For all t ∈ N, the noise ωt ∈ R satisfies can be adapted for non-zero centers. In order to stabilize the
n

kωt k22 ≤ ǫ for a known bound ǫ ≥ 0. origin, we define the quadratic stage cost function
We define the set of system matrices (A, B) consistent l(u, x) = kuk2R + kxk2Q ,
with the data xi , ui , xi+1 , i ∈ N by
where R, Q ≻ 0. The following results can be adapted for
Σi := (A, B) : (1) holds for some ωi satisfying kωi k22 ≤ ǫ . non-zero equilibria (us , xs ) 6= (0, 0).


This set includes all system matrices for which there exists III. DATA -D RIVEN M IN -M AX MPC
a noise satisfying Assumption 1 and the system dynamics
In Section III-A, we define a general data-driven min-
(1). We proceed as in [3], [4], [17] to derive a data-driven
max MPC problem with input and state constraints. In
parametrization of the system matrices. The system dynamics
Section III-B, we restrict the optimization to state-feedback
(1) can be rewritten as
control laws, which allows to reformulate the data-driven
ωi = xi+1 − As xi − Bs ui , min-max MPC problem as an SDP. The state-feedback con-
trol law at each time step can be obtained from a receding-
with which the set of system matrices Σi can be equivalently horizon algorithm, which is proposed in Section III-C. Fi-
characterized by the following quadratic matrix inequality nally, we prove recursive feasibility, constraint satisfaction
    and exponential stability for the closed-loop system in Sec-
 I xi+1 ǫI 0  ⊤  tion III-D.
  

Σi = (A, B) : I A B 0 −xi  ⋆ 0 .
0 −I
0 −ui A. Min-max MPC problem
 

The set of system matrices consistent with the sequence of At time t, given an initial state xt , the data-driven min-
input-state measurements (U− , X) is defined by max MPC optimization problem is formulated as follows:
∞
X
T\
−1 ∗
J∞ (xt ) := min max l(ūk (t), x̄k (t)) (3a)
C= Σi . ū(t) (A,B)∈C
k=0
i=0
s.t. x̄k+1 (t) = Ax̄k (t) + B ūk (t), (3b)
We can characterize C by the following quadratic matrix x̄0 (t) = xt , (3c)
inequality [3], [4]
kūk (t)kSu ≤ 1, ∀t ∈ N, (3d)
( ⊤ )
kx̄k (t)kSx ≤ 1, ∀(A, B) ∈ C, t ∈ N.
  
I A B Π(τ ) I A B  0, (3e)
C = (A, B) : , (2)
∀τ = (τ0 , . . . , τT −1 ), τi ≥ 0, i ∈ I[0,T −1] The objective function is a minimization of the worst-case
where cost over all consistent system matrices in C by adapting the
   ⊤ control input ūk (t), ∀k ∈ N. In the optimization problem,
T −1 I xi+1   I xi+1 x̄k (t) and ūk (t) are the predicted state and control input at
X ǫI 0 
Π(τ ) = τi 0
 −xi  0 −xi  . time t+k based on the measurement at time t. The prediction
0 −I
i=0 0 −ui 0 −ui model employs the system matrices consistent with the data
trajectory in constraint (3b). In constraint (3c), we initialize (A, B) ∈ C. As the following theorem shows, this is possible
x̄0 (t) as the state measurement at time t. We consider that based on LMIs.
the input and state should lie in the ellipsoidal constraints in Theorem 1: Suppose that there exist γ > 0, H ∈ Rn×n ,
(3d) and (3e). The state constraint (3e) must be satisfied for L ∈ Rm×n , τ ∈ RT such that the inequalities (8) hold
any states predicted using any system matrices in C. In order
1 x⊤
 
to effectively address this problem and obtain a tractable t
 0, (8a)
solution, we limit our focus to a state-feedback control law xt H
   
of the form ut = Ft xt , where Ft ∈ Rm×n .   0
−H 0

 0 + Π(τ ) H  0 
B. Reformulation as an SDP 0 
 ≺ 0,

  L  (8b)
 0 H L⊤ ⊤

In this subsection, we first prove that the state-feedback −H Φ
gain Ft that minimizes the upper bound on the optimal 0 Φ −γI
cost of the min-max MPC problem (4) can be obtained by τ = (τ0 , . . . , τT −1 ), τi ≥ 0, ∀i ∈ I[0,T −1] , (8c)
solving the SDP (8). Then, we reformulate the input and  
state constraints as linear matrix inequalities (LMIs). MR L
where Φ = and MR⊤ MR = R, MQ ⊤
MQ = Q. Then
We first neglect the input and state constraints (3d) and MQ H
(3e) and obtain the state-feedback gain for the following γ is an upper bound on the optimal cost of (4). Applying
data-driven min-max MPC problem the state-feedback control ut = F xt with F = LH −1 to the
∞
system (1) leads to a cost that is guaranteed to be at most γ.
Proof: As discussed in (5)-(7), the quadratic function
X
∗
J∞ (xt ) := min max l(ūk (t), x̄k (t)) (4a)
ū(t) (A,B)∈C
k=0
V (xt ) = x⊤ t P xt with P ≻ 0 is an upper bound on the

s.t. x̄k+1 (t) = Ax̄k (t) + B ūk (t), (4b) optimal cost of (4). Suppose x⊤ t P xt ≤ γ holds and define
H = γP −1 ≻ 0. Using the Schur complement, x⊤ t P xt ≤ γ
x̄0 (t) = xt . (4c) is equivalent to the inequality (8a).
The method to reformulate the input and state constraints Additionally, V is required to satisfy the inequality (5)
(3d)-(3e) will be proposed later. for any (A, B) ∈ C, k ∈ N. We provide sufficient conditions
Our goal is to derive an upper bound on the worst-case cost for inequality (5) based on LMIs. By substituting ūk (t) =
over the set C and then to find a state-feedback control law F x̄k (t), the inequality (5) holds for all x̄k (t), ūk (t), k ∈ N
to minimize this upper bound. In order to derive the upper and (A, B) ∈ C if
bound of the worst-case cost over all system matrices in C,
(A + BF )⊤ P (A + BF ) − P + F ⊤ RF + Q ≺ 0 (9)
we define a quadratic function V (x) = x⊤ P x for x ∈ Rn ,
where P ≻ 0. Suppose V satisfies the following inequality holds for any (A, B) ∈ C. Multiplying both sides of the
for all states and inputs x̄k (t), ūk (t), k ∈ N predicted by the inequality (9) with H = γP −1 , defining L = F H and
system dynamics (4b) with any (A, B) ∈ C dividing by γ, we obtain
V (x̄k+1 (t)) − V (x̄k (t)) ≤ −l(ūk (t), x̄k (t)). (5) 1
(AH + BL)⊤H −1(AH + BL)− H + (L⊤RL + HQH) ≺ 0.
γ
To ensure that the cost in equation (4a) is finite, we must have (10)
lim x̄k (t) = 0. Therefore, we have lim V (x̄k (t)) = 0.
k→∞ k→∞   Schur complement with H ≻ 0 and defining Φ =
Using the
Summing the inequality (5) from k = 0 to k = T along an MR L
, the inequality (10) is equivalent to
arbitrary trajectory and letting T → ∞, we obtain MQ H
∞
H − γ1 Φ⊤ Φ (AH + BL)⊤
X  
−V (x̄0 (t)) ≤ − l(ūk (t), x̄k (t)). (6) ≻ 0.
k=0
(AH + BL) H

Since x0 (t) = xt and the inequality (6) holds for any Using the Schur complement again, we obtain the equivalent
matrices (A, B) ∈ C, it also holds for the worst-case value, inequalities
i.e., 1 ⊤ −1
∞
X H − (AH + BL)(H − Φ Φ) (AH + BL)⊤≻ 0, (11a)
max l(ūk (t), x̄k (t)) ≤ V (xt ). (7) γ
(A,B)∈C
k=0 1 ⊤
H− Φ Φ ≻ 0. (11b)
This provides an upper bound on the cost (4a). The above γ
method to derive the upper bound on the worst-case cost, The inequality (11a) is equivalent to
i.e., (5)-(7), is inspired by the existing LMI-based min-max  ⊤
H 0
 
MPC approach in [7]. I I
A⊤   ⊤A⊤ ≻ 0.
The goal of our data-driven min-max MPC problem is
   
H 1 ⊤ −1 H
to synthesize a state-feedback control law ut = Ft xt to B⊤ 0 − (H − γ Φ Φ) B⊤
L L
minimize the upper bound V (xt ) satisfying (5) for any (12)
As the set C is characterized by (2), the inequality (12) holds holds for any (A, B) ∈ C, k ∈ N. Since l(ūk (t), x̄k (t)) ≥ 0,
for any (A, B) ∈ C if there exists τ = (τ0 , . . . , τT −1 ), τi ≥ we have
0, i ∈ I[0,T −1] , such that the following inequality holds
x̄k+1 (t)⊤ P x̄k+1 (t) ≤ x̄k (t)⊤ P x̄k (t). (14)
−H 0
 
⊤ holds for any (A, B) ∈ C, k ∈ N. Therefore, if x⊤
t P xt =
H  + Π(τ ) ≺ 0. (13)
   
 H
0 (H − γ1 Φ⊤ Φ)−1 x̄0 (t)⊤ P x̄0 (t) ≤ γ, then we have
L L
(14)
Applying again the Schur complement, (13) together with x⊤ ⊤
t+1 P xt+1 ≤ max x̄1 (t) P x̄1 (t) ≤ γ.
(11b) is equivalent to (A,B)∈C

This argument can be continued for the state xt+k , ∀k ∈

   
  0
−H 0 I[2,∞) by induction, completing the proof.

 0 + Π(τ ) H  
0  ≺ 0.
Based on Lemma 1, the following theorem shows that
 L 
⊤ 1 ⊤ the input and state constraints in (3d) and (3e) can be
 
0 H L −H + γ Φ Φ
reformulated as LMIs.
Using the Schur complement again, we obtain the LMI (8b). Theorem 2: If (8) as well as (15) hold, then the input and
Based on the preceding discussion, if the LMIs (8b) and state constraints (3d) and (3e) hold
(8c) are satisfied, then V satisfies the inequality (5) for any
H L⊤
 
(A, B) ∈ C.  0, (15a)
As we have shown earlier in (7), if the function V satisfies L Su−1
 
the inequality (5) for any (A, B) ∈ C, then V (xt ) is an Sx I
 0. (15b)
upper bound on the cost (4a). Given that (8a) holds, we have I H
thus shown that γ is an upper bound on the optimal cost of Proof: Suppose that a feasible solution of problem (8)
problem (4). is γ, H, L, τ . Defining P = γH −1 and F = LH −1 , by
Remark 1: The proof of Theorem 1 bears similarities to constraint (8a), we have xt ∈ E = {x ∈ Rn : x⊤ P x ≤ γ}.
the results presented in [14], where the authors aim to find a Given that the input is in a state-feedback form, we can write
state-feedback control law that minimizes the upper bound of the input constraint (3d) as
an infinite horizon cost functional. However, it is important max kūk (t)k2Su = max kF x̄k (t)k2Su ≤ 1.
to note some key distinctions. Specifically, in [14], the data k∈I≥0 k∈I≥0
used for designing the LMI-based data-driven controller are By Lemma 1, the set E is an RPI set for the uncertain system
generated by the system without noise. On the other hand, xt+1 = (A + BF )xt with (A, B) ∈ C. Thus, the input
in the present scheme, the given offline data are affected by constraint (3d) can be written as
the noise satisfying instantaneous constraints. Consequently,
we need to employ different technical tools to characterize max kF x̄k (t)k2Su ≤ max kF xk2Su ≤ 1. (16)
k∈I≥0 x∈E
the set C and to derive the LMI constraints (8b)-(8c).
Remark 2: Solving the SDP problem (8) enables us to The inequality (16) holds if
determine an upper bound for the objective of problem (4).  ⊤ 
−F ⊤ Su F
 
x 0 x
The main source of conservatism of this upper bound is due ≥ 0,
1 0 1 1
to the linear state-feedback form of the input. Additionally,
the fact that the upper bound is assumed to be quadratic holds for all x such that
can increase the gap. Reducing conservatism by considering  ⊤   
x −P 0 x
more general state-feedback and cost upper bound functions ≥ 0.
1 0 γ 1
is an interesting issue for future research.
In the following lemma, we show that the sublevel set Using the S-procedure, if there exists λ ≥ 0 such that
of the function V (x) is robust positive invariant (RPI). 
−F ⊤ Su F 0
 
−P 0

This result will be used to reformulate the input and state −λ  0, (17)
0 1 0 γ
constraint in Theorem 2.
Lemma 1: Let xt be the state at time t. Suppose there then the input constraint (3d) must be satisfied. The inequal-
exist γ, H, L, τ such that (8a)-(8c) hold, denote the resulting ity (17) holds iff the following inequalites hold
state-feedback gain by F = LH −1 , and define P = γH −1 . λP − F ⊤ Su F  0, (18a)
The set E = {x ∈ Rn : x⊤ P x ≤ γ} is an RPI set for the
uncertain system xt+1 = (A + BF )xt with (A, B) ∈ C, 1 − λγ ≥ 0. (18b)
i.e., if x⊤ ⊤
t P xt ≤ γ, then xt+k P xt+k ≤ γ, ∀(A, B) ∈ C, k ∈ If there exist γ, P, F, λ such that (18) hold and λ < γ1 , then
I[1,∞) . there must exist λ′ = γ1 such that (18) hold for γ, P, F, λ′ .
Proof: As proved in Theorem 1, the inequality (5)
Therefore, we can choose the multiplier to be λ = γ1 .
holds because the constraints (8b)-(8c) are satisfied. From
Multiplying both sides of (18a) with H, the inequality (18a)
inequality (5), we have that
is equivalent to
x̄k+1 (t)⊤ P x̄k+1 (t) − x̄k (t)⊤ P x̄k (t) ≤ −l(ūk (t)x̄k (t)) H − L⊤ Su L  0. (19)
 satisfied with the feasible solution of the problem (20) at time
Using the Schur complement, the inequality (19) is equiva- t. We need to prove that the inequality (8a) is feasible with
lent to (15a). The state constraint (3e) can be written as this candidate solution for the future states xt+1 . Suppose
max kx̄k (t)k2Sx ≤ max kxk2Sx ≤ 1. γ, L and H are a feasible solution of the problem (20) at time
k∈I≥0 x∈E t. We define P = γH −1 and F = LH −1 . The feasibility
of the problem at time t implies satisfaction of (8a), which
Thus, the state constraint (3e) holds if x⊤ Sx x ≤ 1 holds
implies x⊤ t P xt ≤ γ. Using Lemma 1, this implies that for
for all x such that x⊤ P x ≤ γ. The statement holds if
any (A, B) ∈ C, we have
Sx  γ −1 P . Using the Schur complement, Sx  γ −1 P
is equivalent to (15b). x⊤ ⊤
t (A + BF ) P (A + BF )xt ≤ γ.
C. Receding-horizon algorithm The state at t + 1 is xt+1 = (As + Bs F )xt , where the true
We now combine the optimization problem (8) with the system matrices (As , Bs ) ∈ C. Thus, we have
constraints (15) to present the proposed data-driven min-max
MPC approach. Given an initial state xt , the state-feedback x⊤
t+1 P xt+1 ≤ γ.
gain Ft ∈ Rm×n that minimizes the derived upper bound on Thus, the feasible solution of the optimization problem at
the optimal cost of the min-max MPC problem (3) can be time t is also feasible at time t + 1. This argument can be
obtained by solving the following optimization problem continued for time t + 2, t + 3, . . ., which proves recursive
minimize γ (20a) feasibility.
γ,H,L,τ Part II: As shown in Theorem 2, if the inequalities (15)
s.t. (8) and (15) hold. (20b) hold, then the predicted input and state for the uncertain
system xt+1 = (A + BF )xt with (A, B) ∈ C satisfy the
We denote the optimal solution of the problem (20) by
input and state constraints. Since the true system matrices
γ ⋆ , H ⋆ , L⋆ , τ ⋆ . The corresponding optimal state-feedback
(As , Bs ) lie in the set C, the closed-loop input and state also
gain is given by F = L⋆ (H ⋆ )−1 .
satisfy the input and state constrains.
The data-driven min-max MPC problem (3) is solved in
Part III: Let us denote the optimal solution of problem
a receding horizon manner, see Algorithm 1. At time t, we
(20) at time t by Ht , γt and define Pt = γt Ht−1 . Since Pt
solve the optimization problem (20) and obtain the optimal
is a suboptimal solution at time t + 1, compare Part I of the
state-feedback gain Ft . Only the first computed input ut =
proof, we must have x⊤ ⊤
t+1 Pt+1 xt+1 ≤ xt+1 Pt xt+1 . The state
Ft xt is implemented. At the next sampling time t + 1, we
at time t + 1 is xt+1 = (As + Bs Ft )xt with (As , Bs ) ∈ C.
re-iterate the described procedure.
We choose V (xt ) = x⊤ t+1 Pt+1 xt+1 as the Lyapunov
function candidate. A lower bound on V (xt ) follows from
Algorithm 1: Data-driven min-max MPC. the definition V (xt ) ≥ l(ut , xt ) ≥ kxt k2Q . An upper
1 At time t = 0, measure state x0 ; bound on V (xt ) can be constructed from x⊤ t+1 Pt+1 xt+1 ≤
2 Solve the problem (20); x⊤t+1 Pt xt+1 , which holds for t ∈ N. This implies that
3 Apply the input ut = Ft xt ; V (xt ) ≤ x⊤ t P0 xt .
4 Set t = t + 1 and go back to 2; According to the inequality (5), we must have
x⊤ ⊤
t+1 Pt xt+1 − xt Pt xt ≤ −l(ut , xt ). Choosing c as the
minimum eigenvalue of Q, we have l(ut , xt ) ≥ ckxt k2 .
D. Closed-loop guarantees Therefore, we can get
Now we prove recursive feasibility, constraint satisfaction 2
x⊤ ⊤
t+1 Pt+1 xt+1 − xt Pt xt ≤ −ckxt k .
and exponential stability of the closed-loop system xt+1 =
As xt + Bs ut resulting from the proposed scheme. This shows that xs = 0 is exponentially stable for the closed
Theorem 3: If the optimization problem (20) is feasible loop.
at time t = 0, then The proof of feasibility relies on Lemma 1, which estab-
(i) it is feasible at any time t ∈ N, lishes that the closed-loop state remains contained within the
(ii) the closed-loop trajectory satisfies the constraints, i.e., set E. By choosing the optimal solution as a feasible solution
kut kSu ≤ 1, kxt kSx ≤ 1 for all t ∈ N; at the next time step, we can guarantee the feasibility of
(iii) the desired equilibrium xs = 0 is exponentially stable problem (20). We select the Lyapunov function as V (x) =
for the closed-loop system. x⊤ P x. Using the inequality (5), we have that the decrease
Proof: The proof is composed of three parts. Part rate of the Lyapunov function is less than or equal to the
I, II and III prove feasibility, constraint satisfaction and negative of the stage cost. This deduction allows us to
exponential stability, respectively. establish exponential stability for the closed-loop system.
Part I: Let us assume that the optimization problem (20) Remark 3: Existing data-driven MPC schemes based on
is feasible at sampling time t. The only constraint in the the Fundamental Lemma [2], [5] suffer from the limitation
problem that depends explicitly on the measured data xt is that the computational complexity increases with data length.
the inequality (8a). Constraints (8b)-(8c) and (15) are still In contrast, the proposed approach offers larger flexibility
by trading off computational tractability and conservatism. 0.01
While the computational complexity of Problem (20) also
grows with the data length via the size of τ , it is also 0
possible to use different choices of multipliers which reduce
the computational complexity at the cost of additional con-
-0.01
servatism, see [3, Section V.B]. For example, when imposing
the additional condition τi = τ, ∀i ∈ I[1,T ] for some τ ≥ 0,
-0.02
the size of problem (20) is independent of the data length.

IV. S IMULATION -0.03

In this section, we apply the proposed data-driven min-

-0.04
max MPC scheme to the linearization of the nonlinear -10 -8 -6 -4 -2 0 2
continuous stirred-tank reactor (CSTR) consider
 in [13].
 The 10 -3
0.9831
nonlinear system from [13] is linearized at with a (a)
0.3918
sampling time 0.5 sec. The linearized system is given by
12
   
0.9749 −0.0135 0.041
xt+1 = xt + 10−4 · u . (21) 10
0.0004 0.9888 5.934 t
8
The system matrices are unknown, but an input-state tra-
jectory (U− , X) of length T = 200 with additive noise ωt 6
is available, where the input u ∈ U− is chosen uniformly
4
from the unit interval [−10, 10]. The noise satisfies the
instantaneous constraint in Assumption 1, i.e., ωt ∈ {ω ∈ 2
R2 : kωk22 ≤ 10−6 }. This input-state sequence is used to
0
characterize the set C in the proposed data-driven min-max
MPC scheme. -2
0 50 100 150 200 250 300
In the data-driven min-max MPC problem, the input and
state constraints are given by kut kSu ≤ 1 and kxt kSx ≤ 1,
where (b)
 
1000 0
Su = 0.01, Sx = . Fig. 1: Closed-loop input and state when choosing different
0 500
weighting matrices. (a) Closed-loop state x. (b) Closed-loop
We choose two different weighting matrices of the stage cost input u.
function to compare its influence on the closed-loop, i.e.,
Q = I, R = 1 and Q = I, R = 10−4 . The initial state is
given as x0 = [−0.01, −0.04]⊤. approaches a neighborhood of the origin. We calculate the
Figure 1 illustrates the closed-loop input-state trajectory sum of stage costs over all 300 iterations, resulting in a
resulting from the application of the data-driven min-max value of 0.0411, compared to a cost of 0.0369 for the noise-
MPC scheme. The closed-loop trajectory converges to the free closed-loop. This confirms that the proposed data-driven
origin, and the input and state constraints are satisfied in the min-max MPC scheme also produces reliable results in the
closed-loop operation. When choosing a smaller weighting presence of online noise.
matrix R, i.e., R = 10−4 , the inputs obtained are larger
compared to the case when R = 1. However, the inputs V. C ONCLUSION
remain below the constraint bound of 10. This can be at-
tributed to two factors. First, we constrain the input to a state- In this paper, we presented a data-driven min-max MPC
feedback form. Second, we reformulate the input constraint scheme which uses noisy data to design controller for LTI
using the S-procedure, which can introduce conservatism into systems with unknown system matrices. Our main con-
the constraint. tribution is to reformulate the data-driven min-max MPC
In addition, we implement the data-driven min-max MPC problem as an SDP and to establish that the proposed
scheme in the system (21) with online noise ωt ∈ {ω ∈ R2 : scheme guarantees closed-loop properties including recursive
kωk22 ≤ 10−6 }, i.e., noise affecting the system dynamics feasibility, constraint satisfaction and exponential stability. A
during closed-loop operation. The weight matrices are Q = I numerical example showed the effectiveness of the proposed
and R = 10−4 . All other aspects, such as constraints and method. Compared with the existing min-max MPC scheme
the initial state, remain consistent with the noise-free case. in [1], [7], [16], our scheme does not require any priori model
The results indicate that the input and state constraints are knowledge. Extending the results in this paper to robust data-
still satisfied during the closed-loop operation. However, the driven min-max MPC for systems with online noise is an
closed-loop trajectory does not converge to the origin but interesting future topic.
 R EFERENCES analysis, signal processing, and control. Annual Reviews in Control,
52:42–64, 2021.
[1] A. Bemporad, F. Borrelli, and M. Morari. Min-max control of [12] D. Q. Mayne, M. M. Seron, and S.V. Raković. Robust model predictive
constrained uncertain discrete-time linear systems. IEEE Transactions control of constrained linear systems with bounded disturbances.
on Automatic Control, 48(9):1600–1606, 2003. Automatica, 41(2):219–224, 2005.
[2] J. Berberich, J. Köhler, M. A. Müller, and F. Allgöwer. Data-driven [13] D.Q. Mayne, E.C. Kerrigan, E.J. Van Wyk, and P. Falugi. Tube-based
model predictive control with stability and robustness guarantees. robust nonlinear model predictive control. International journal of
IEEE Transactions on Automatic Control, 66(4):1702–1717, 2021. robust and nonlinear control, 21(11):1341–1353, 2011.
[3] J. Berberich, C. W. Scherer, and F. Allgöwer. Combining prior [14] H. H. Nguyen, M. Friedel, and R. Findeisen. LMI-based data-driven
knowledge and data for robust controller design. IEEE Transactions robust model predictive control. arXiv.2303.04777, 2023.
on Automatic Control, 68(8):4618–4633, 2023. [15] J. B. Rawlings, D. Q. Mayne, and M. Diehl. Model predictive control:
[4] A. Bisoffi, C. De Persis, and P. Tesi. Trade-offs in learning controllers theory, computation, and design, volume 2. Nob Hill Publishing
from noisy data. Systems & Control Letters, 154:104985, 2021. Madison, 2017.
[5] J. Coulson, J. Lygeros, and F. Dörfler. Data-enabled predictive control:
[16] P.O.M. Scokaert and D.Q. Mayne. Min-max feedback model pre-
In the shallows of the DeePC. In Proc. European Control Conference
dictive control for constrained linear systems. IEEE Transactions on
(ECC), pages 307–312, 2019.
Automatic Control, 43(8):1136–1142, 1998.
[6] C. De Persis and P. Tesi. Formulas for data-driven control: Stabi-
lization, optimality, and robustness. IEEE Transactions on Automatic [17] H. J. van Waarde, M. K. Camlibel, and M. Mesbahi. From noisy data
Control, 65(3):909–924, 2020. to feedback controllers: non-conservative design via a matrix S-lemma.
[7] M. V. Kothare, V. Balakrishnan, and M. Morari. Robust constrained IEEE Transactions on Automatic Control, 2020.
model predictive control using linear matrix inequalities. Automatica, [18] H. J. van Waarde, J. Eising, M. K. Camlibel, and H. L. Trentel-
32(10):1361–1379, 1996. man. The informativity approach to data-driven analysis and control.
[8] L. Ljung. System Identification: Theory for the user. Prentice Hall arXiv:2302.10488, 2023.
PTR, 1999. [19] H. J. van Waarde, J. Eising, H. L. Trentelman, and M. K. Camlibel.
[9] M. Lorenzen, M. Cannon, and F. Allgöwer. Robust MPC with Data informativity: a new perspective on data-driven analysis and
recursive model update. Automatica, 103:461–471, 2019. control. IEEE Transactions on Automatic Control, 65(11):4753–4768,
[10] X. Lu, M. Cannon, and D. Koksal-Rivet. Robust adaptive model pre- 2020.
dictive control: performance and parameter estimation. International [20] J. C. Willems, P. Rapisarda, I. Markovsky, and B. De Moor. A note
Journal of Robust and Nonlinear Control, 31(18):8703–8724, 2021. on persistency of excitation. Systems Control Letters, 54(4):325–329,
[11] I. Markovsky and F. Dörfler. Behavioral systems theory in data-driven 2005.
This figure "input.jpg" is available in "jpg" format from:

http://arxiv.org/ps/2309.17307v1
This figure "state.jpg" is available in "jpg" format from:

http://arxiv.org/ps/2309.17307v1