2020 American Control Conference

Denver, CO, USA, July 1-3, 2020

Robust data-driven state-feedback design

Julian Berberich1 , Anne Koch1 , Carsten W. Scherer2 , and Frank Allgöwer1

Abstract— We consider the problem of designing robust state- or an extension of [9] to certain classes of nonlinear sys-
feedback controllers for discrete-time linear time-invariant tems [11]. Moreover, the recent work [12] derives a simple
systems, based directly on measured data. The proposed design data-dependent closed-loop parametrization of LTI systems
procedures require no model knowledge, but only a single open-
loop data trajectory, which may be affected by noise. First, a under state-feedback. This parametrization is used to solve
data-driven characterization of the uncertain class of closed- various control problems from data, including stabilization
loop matrices under state-feedback is derived. By considering and linear-quadratic regulation. However, no meaningful
this parametrization in the robust control framework, we design guarantees were given in the presence of noisy data.
data-driven state-feedback gains with guarantees on stability It is the goal of this paper to provide non-conservative end-
and performance, containing, e.g., the H∞ -control problem as a
special case. Further, we show how the proposed framework can to-end guarantees for data-driven control. To be more precise,
be extended to take partial model knowledge into account. The we employ a single noisy input-state trajectory of finite
validity of the proposed approach is illustrated via a numerical length to design controllers which guarantee closed-loop
example. stability and performance for all systems which are consistent
I. I NTRODUCTION with the measured data and the assumed noise bound. This
is achieved by extending the approach of [12] to account for
Recently, the design of controllers directly from mea- noise and applying robust control techniques to the resulting
sured data has received increasing interest [1], [2]. While uncertain system class. Another recent paper [13] considers
established methods, e.g., those based on reinforcement data-driven analysis and control with not persistently exciting
learning, rarely address closed-loop guarantees, there has data. In particular, it is shown for noise-free data that certain
been a renewed effort to provide such guarantees using control problems can be solved from data, even if the system
novel statistical estimation techniques [3], [4], [5], [6]. cannot be uniquely identified, thus illustrating advantages of
Potential alternatives are, e.g., robust control with prior direct data-driven control. Similarly, the results of this paper
set membership identification [7], which is however well- do not require persistence of excitation explicitly. Moreover,
known to be computationally demanding, and unfalsification- our results lead to simple design procedures for direct data-
based approaches [8], which typically require infinitely long driven control with desirable closed-loop guarantees, and are
data for closed-loop guarantees. In general, providing non- thus a promising alternative to identification-based control.
conservative end-to-end guarantees for the closed loop using The paper is structured as follows. After stating the prob-
noisy data of finite length is an open problem, even if the lem formulation in Section II, we use noisy data to describe
data is generated by a linear time-invariant (LTI) system. the uncertain closed loop under state-feedback, and we apply
A promising approach towards this goal relies on behav- known robust control methods to design controllers with sta-
ioral systems theory. In [9], it was proven that the vector bility and performance guarantees in Section III. Moreover,
space of all input-output trajectories of an LTI system is we extend the proposed, purely data-driven approach to sys-
spanned by time-shifts of a single measured trajectory, given tems with mixed data-driven and model-based components.
that the respective input signal is persistently exciting. Thus, In Section IV, we apply the robust state-feedback design
a single data trajectory can be used to characterize an LTI techniques successfully to an unstable example system. The
system, without any prior identification steps. Recently, there paper is concluded in Section V.
have been various contributions which consider this result
in the context of data-driven system analysis and control, II. P RELIMINARIES
including dissipativity verification from measured data [10] We denote the n×n identity matrix by In , where the index
is omitted if the dimension is clear from the context. Further,
This work was funded by Deutsche Forschungsgemeinschaft (DFG,
German Research Foundation) under Germany’s Excellence Strategy - EXC A⊥ denotes a matrix containing a basis of the kernel of A.
2075 - 390740016. The authors thank the International Max Planck Research We write `2 for the space of square-summable sequences. In
School for Intelligent Systems (IMPRS-IS) for supporting Julian Berberich a linear matrix inequality (LMI), ∗ represents blocks, which
and Anne Koch.
1 J. Berberich, A. Koch and F. Allgöwer are with the Institute for Systems can be inferred from symmetry. Moreover, we define, for
−2
Theory and Automatic Control, University of Stuttgart, 70569 Stuttgart, elements {xk }i+L+N
k=i of a sequence x, the Hankel matrix
Germany. E-mail: {julian.berberich, anne.koch,  
frank.allgower}@ist.uni-stuttgart.de xi xi+1 . . . xi+N −1
2 C. W. Scherer is with the Institute of Mathematical  xi+1 xi+2 . . . xi+N 
Methods in the Engineering Sciences, Numerical Analysis N
Xi,L :=  . .
 
.. .. ..
and Geometrical Modeling, Department of Mathematics,  .. . . . 
University of Stuttgart, 70569 Stuttgart, Germany. E-mail:
[email protected] xi+L−1 xi+L ... xi+L+N −2

978-1-5386-8266-1/$31.00 ©2020 AACC 1532

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:16:15 UTC from IEEE Xplore. Restrictions apply.
 N
That is, the matrix Xi,L starts with the element xi and has L Assumption 3. The matrix Ŵ is an element of
rows and N columns. As a shorthand notation, we abbreviate  >   
N -windows of x, starting at i = 0 and i = 1, by
n
mw ×N W Qw Sw W o
W= W ∈R > 0 ,
I Sw Rw I
N
 
X = X0,1 = x0 x1 . . . xN −1 ,
N
  for some known matrices Qw ∈ Rmw ×mw , Sw ∈ Rmw ×N ,
X+ = X1,1 = x1 x2 . . . xN ,
Rw ∈ RN ×N with Rw  0.
respectively. In the present paper, we consider LTI systems
Through Assumption 3 it is assumed that the unknown
of the form
  disturbance realization, which affects the measured data, lies
 xk
in some known set which is described by a quadratic matrix
  
xk+1 Atr Bw Btr 
= wk , (1) inequality. Implicitly, Ŵ ∈ W implies a quadratic bound
zk C Dw D
uk on the sequence {ŵk }N −1
k=0 and encompasses many practical
where xk ∈ Rn is the state, wk ∈ Rmw is the disturbance, bounds as special cases. For instance, if the maximal singular
uk ∈ Rm is the control input, and zk ∈ Rpz is the value of Ŵ is bounded as σmax (Ŵ ) ≤ w̄, then Ŵ ∈
performance output. We design state-feedback controllers W holds with Qw = −I, Sw = 0, Rw = w̄2 I. More
uk = Kxk to control the system (1). Our design procedures generally, a description of the form Ŵ ∈ W provides
are purely data-driven and do not require knowledge of the a flexible framework to model general noise signals, in
true system matrices Atr , Btr . We do, however, assume that particular when multiple quadratic matrix inequalities are
the matrices Bw , C, Dw , D are known. For our purposes, combined. It is an interesting aspect for future research to
Bw is essentially a parameter to model the influence of the derive suitable matrices Qw , Sw , Rw for different, practically
disturbance, whereas C, Dw , D constitute a user choice for relevant scenarios such as norm bounds on the sequence
−1
performance. In this paper, we use the following definition {ŵk }N
k=0 .
of persistence of excitation of the data under consideration.
III. DATA - DRIVEN STATE - FEEDBACK
−1
Definition 1. The sequence {xk , uk }N
k=0 is called persis- In this section, we consider the design of state-feedback
X gains, based directly on measured data which is perturbed
tently exciting if the matrix has full row rank.
U by a disturbance satisfying Assumption 3. First, we derive
According to [9], controllability and a certain rank prop- a data-driven characterization of the uncertain closed loop,
erty of the input are sufficient for persistence of excitation. using a single open-loop data trajectory. Thereafter, we apply
 
known robust control methods to this parametrization in
Theorem 2 ([9, Corollary 2]). If Atr , Btr Bw is
order to design state-feedback controllers which guarantee
controllable and the matrix
 N −n  stability and performance for all closed-loop matrices that
W0,n+1 are consistent with the measured data. Finally, we extend
N −n
U0,n+1 the proposed framework to systems with mixed data-driven
has full row rank, then {xk , uk }N −1 and model-based components.
k=0 is persistently exciting.

Definition 1 differs from the notion of persistence of exci- A. Uncertain closed-loop parametrization
tation considered in [9], which concerns only the input data, In the following, we extend [12] by characterizing the
and is preferred in the present paper for convenience. In [12], closed-loop dynamics of (1) under state-feedback, using
it is shown how a single, persistently exciting open-loop noisy measurements. Let {xk , uk }N k=0 be a measured tra-
trajectory can be employed to recover the system matrices jectory of (1), corresponding to an unknown disturbance
of an LTI system. Furthermore, a linear parametrization of realization Ŵ . We define ΣX,U as the set of all pairs (A, B)
the closed loop under state-feedback is derived, depending that are consistent with the data {xk , uk }N
k=0 for some noise
also only on a single open-loop data trajectory. It is the instance W ∈ W, i.e.,
contribution of the present paper to extend the framework
of [12] in order to provide robust stability and performance ΣX,U = {(A, B) | X+ = AX + BU + Bw W, W ∈ W}.
guarantees in the presence of noise. In contrast to [12], Using fixed data matrices X and U , ΣX,U parametrizes the
persistence of excitation will generally not be required for unknown system matrices A and B via W. By assump-
our results. tion, the true disturbance realization Ŵ satisfies X+ =
Throughout this paper, we consider the following scenario: Atr X + Btr U + Bw Ŵ and Ŵ ∈ W; therefore, the true
From simulation or an experiment, a single open-loop input- pair (Atr , Btr ) is an element of ΣX,U . Furthermore, for
state sequence {xk , uk }Nk=0 is obtained as a trajectory of (1) some state-feedback gain K, we define the set of closed-
−1
for some unknown disturbance {ŵk }N k=0 . This trajectory loop matrices that are consistent with the data as
is used directly for robust controller design, without prior
system identification. The only available information on the ΣK
X,U = {AK | AK = A + BK, (A, B) ∈ ΣX,U }.
disturbance realization is the following bound on the matrix
In the following, we show that an exact parametrization of
ΣK
 
Ŵ = ŵ0 ŵ1 . . . ŵN −1 . X,U can be constructed directly from open-loop data. To

1533
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:16:15 UTC from IEEE Xplore. Restrictions apply.
 this end, for some matrix G ∈ RN ×n , we define AG as the Theorem 4 provides an exact parametrization of the un-
set of matrices AG ∈ Rn×n such that certain closed loop under a fixed state-feedback K, using a
single open-loop trajectory of the unknown system. In partic-
AG = (X+ − Bw W )G, (2) ular, no closed-loop measurements and no model knowledge
for some W ∈ W satisfying are required to construct the set AG , which parametrizes the
 ⊥ uncertain closed loop. This set relies on fixed data matrices
X X and U , which are obtained offline, and is parametrized
(X+ − Bw W ) = 0. (3)
U via the disturbance W ∈ W satisfying (3). The equation (3)
ensures that the matrices in AG contain only those W ∈ W
Theorem 4. If G ∈ RN ×n and K ∈ Rm×n satisfy
    for which there exist matrices A, B satisfying the system
X I dynamics.
G= , (4)
U K In general, the condition (4) only requires that X has full
row rank, but not necessarily that the data are persistently
then ΣK
X,U = AG . exciting. Nevertheless, if {xk , uk }N −1
k=0 is persistently excit-
Proof. First, we note that the constraint (3) is equivalent to ing, then, for any state-feedback K, (4) can be solved for
the implication G, i.e., any possible closed-loop matrix can be constructed.
  Equivalently, the set of all AG ∈ AG with G ∈ RN ×n
X satisfying XG = I is equal to the set of all possible closed-
Ṽ = 0 ⇒ (X+ − Bw W )Ṽ = 0,
U loop matrices under state-feedback.
for any matrix Ṽ with N rows. By the Fredholm alternative, −1
Corollary 5. If {xk , uk }N
k=0 is persistently exciting, then it
this is in turn equivalent to the existence of a solution V to holds that
the system of linear equations
{AG ∈ AG | G ∈ RN ×n , XG = I}
 
X (9)
V = X+ − Bw W. (5) ={AK ∈ ΣK
X,U | K ∈ R
m×n
}.
U
Proof. This follows directly from Theorem 4.
Proof of ΣK K
X,U ⊆ AG : Let AK ∈ ΣX,U , i.e., there exist
matrices A, B as well as W ∈ W such that Corollary 5 suggests that the set AG can be employed to
design controllers with robustness guarantees for all closed-
AK = A + BK, (6) loop matrices in ΣK X,U , by optimizing over the parameter G
X+ = AX + BU + Bw W. (7) instead of the gain K. If the data are not persistently exciting,
then (9) holds with ”⊆” instead of ”=”, since AG contains
Then, it follows that only closed-loop matrices resulting from feedback gains K
(6) 
 
 I (4) 
 
 X of the form K = U G (compare (4)). In this case, Theorem 4
AK = A + BK = A B = A B G can still be used for robust controller design since, for any
K U
(7)
fixed K = U G, AG captures the full closed-loop uncertainty
= (X+ − Bw W )G. induced by the noise, i.e., AG = ΣK X,U . However, the
conservatism of robust controller design based on Theorem 4
It remains to show that W satisfies (3) or, equivalently, there
increases if the data are not persistently exciting since there
existsV such that (5) holds. It follows directly from (7) that
may exist a controller K which, e.g., renders all matrices in
V = A B solves (5), which thus proves AK ∈ AG .
ΣK X,U stable, but for which there exists no G satisfying (4).
Proof of AG ⊆ ΣK X,U : Let AG ∈ AG , i.e., there exists W ∈
The disturbance W parametrizing AG is not only restricted
W such that (2) and (3) hold. We need to show the existence
by W ∈ W but also via the affine constraint (3) and
of matrices A, B as well as W̃ ∈ W such that
therefore, the construction of AG requires the computation
>
of the kernel of X > U > , which may be undesirable

A + BK = (X+ − Bw W )G,
X+ = AX + BU + Bw W̃ . from a numerical viewpoint. In Sections III-B and III-C, we
employ a superset of AG to derive simple robust controller
If we choose W̃ = W , these equations are equivalent to design procedures for closed-loop stability and performance,


 X I

  respectively.
A B = (X+ − Bw W ) I G .
U K B. Robust state-feedback for stability
Using (4), this is in turn equivalent to In this section, we apply known robust control methods to
  render all matrices in AG stable. To facilitate the design, we
  X    
A B I G = (X+ − Bw W ) I G . (8) consider
U
AsG = {AG | AG = (X+ − Bw W )G, W ∈ W}, (10)
Since AG ∈ AG , there exists a solution V to (5). Hence,
the choice A B = V satisfies (8), which implies AG ∈ which is a superset of the uncertain closed loop AG , i.e.,
ΣKX,U . AG ⊆ AsG . The difference between AsG and AG is that

1534
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:16:15 UTC from IEEE Xplore. Restrictions apply.
 the latter considers only those disturbances W ∈ W, which Nevertheless, persistence of excitation is required for
satisfy the n constraints defined by (3). Hence, AsG is in gen- equality in (9), i.e., to construct any closed-loop system
eral larger than AG and, therefore, controller design based (cf. Corollary 5), and thus, it enhances feasibility of (13).
on AsG is generally more conservative than a design based In particular, if the data are persistently exciting and there
on AG . Nevertheless, AsG admits a simpler parametrization exists a controller which stabilizes all matrices in AsG with a
and can be translated directly into a standard robust control common Lyapunov function, then (12) and (13) are feasible.
format. Further, as we will see in Section IV, considering AsG Hence, Corollary 6 contains two main sources of conser-
instead of AG leads to meaningful robust controllers also for vatism: a) the difference between AsG and ΣK X,U and b)
practical examples. Providing an exact quantification of the the fact that a common Lyapunov function is employed for
conservatism induced by this replacement will be the subject stabilization, similar to simple model-based robust controller
of future research. design methods. Nevertheless, Corollary 6 provides compu-
In the following, we exploit that the parametrization AsG tationally tractable conditions, based directly on open-loop
is equivalent to a particular lower linear fractional trans- data, to design controllers with stability guarantees.
formation (LFT) (compare [14, Chapter 10]). To be more
Remark 7. Although (13) is not an LMI, it is routine
precise, the matrices in AsG can be described as a lower
to transform it into one following the same steps as in
LFT of a nominal closed-loop system depending on G with
model-based robust state-feedback design (compare [15],
the disturbance W , i.e.,
     [16]). To be more precise, after performing a congruence
xk+1
=
X+ G Bw xk
, transformation on (13) with diag(X −1 , I) and applying the
z̃k −G 0 w̃k (11) Schur complement twice, the nonlinear matrix inequality (13)
w̃k = W z̃k , leads to the LMI
−M > Sw>
M > X+ >
M>
 
where W ∈ W. It follows from Theorem 4 that, if G satisfies −Y
−Sw M >
XG = I, the above LFT contains all potential closed-loop Qw Bw 0  ≺0
 (14)
systems under control with state-feedback K = U G. The  X+ M Bw −Y 0 
−1
following result exploits this fact by using robust control M 0 0 −Rw
methods to design a stabilizing controller parameter G for in the variables Y = X −1 , M = GX −1 . Further, multi-
the LFT (11), which hence stabilizes all elements of ΣK
X,U . plying (12) by Y from the right yields the linear equality
Corollary 6. If there exist X  0, G ∈ RN ×n such that constraint XM = Y. Together with the LMI (14), this leads
to a semidefinite program whose feasibility can be checked
XG = I (12) using standard solvers. The stabilizing state-feedback gain
as well as can then be recovered as K = U M Y −1 .
 >    Corollary 6 suggests a valuable alternative to sequential
∗ ∗ −X 0 0 0 I 0
 ∗ ∗   0 X 0 0  X+ G Bw  system identification and stabilizing robust control. In par-
     ≺ 0, ticular, in the presence of deterministic noise, identification-
 ∗ ∗   0 0 Qw Sw  0 I 
∗ ∗ 0 0 >
Sw Rw −G 0 based methods are usually either computationally intractable,
(13) overly conservative, or they admit no guarantees from finite
data. Essentially, Corollary 6 is a computationally tractable
then A+BK with K = U G is stable for all (A, B) ∈ ΣX,U . alternative to robust controller design based on set member-
Proof. This follows from an application of known robust ship estimation, which relies on an explicit construction of
control methods to the system (11) (cf. [15], [16]). the set ΣX,U [7]. Further alternatives include unfalsification-
based control, which typically requires infinitely long data
Corollary 6 applies known robust control methods to and a prescribed controller structure for closed-loop guar-
design state-feedback controllers which robustly stabilize all antees [8], [17], or a stochastic setting, where recent work
elements of AsG . If K is designed according to Corollary 6, has addressed finite-time guarantees on system identification
then (4) holds and hence, Theorem 4 leads to ΣK X,U = AG with sequential robust control [3], [4], [5], [6]. The latter
which thus implies ΣK X,U = A G ⊆ A s
G . This guarantees results are based on sophisticated statistical analysis and
stability of all closed-loop matrices ΣK X,U that are consistent many of them rely on restrictive assumptions, such as the
with the measured data. Similar to Theorem 4, Corollary 6 availability of multiple independent data trajectories, each of
does not require persistently exciting data explicitly. Thus, which only supplies one data tuple to the estimator. On the
it may be possible to find a controller K which stabilizes contrary, our approach relies on simple matrix manipulations
all elements of ΣK X,U , even if persistence of excitation does combined with existing robust control methods and requires
not hold, i.e., if the data is not sufficiently rich for system only a single data trajectory of finite length. Despite these
identification. Similar phenomena were analyzed for system advantages, the presented approach requires state measure-
analysis and control from noise-free data in [13], where ments which may be restrictive in practice. Extending the
also full row rank of X was sufficient to design stabilizing results in this paper to input-output data is an important
controllers from data. aspect for future research.

1535
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:16:15 UTC from IEEE Xplore. Restrictions apply.
 Remark 8. For the state-feedback stabilization problem hold, then, for any (A, B) ∈ ΣX,U ,
under additive state measurement noise, [12] provides suffi- i) A + BK with K = U G is stable,
cient conditions for closed-loop stability. However, this result ii) (1) with uk = Kxk satisfies quadratic performance with
relies on assumptions that cannot be verified from measured index P .
data. Moreover, in contrast to the approach of [12], an
extension of Corollary 6 to more general (robust) control Proof. The result follows from known robust control meth-
objectives is straightforward. ods (cf. [15], [16]).

C. Robust state-feedback for performance Corollary 10 applies robust control methods to design
controllers which guarantee robust closed-loop performance
Next, we consider the system (1) including the perfor- for all matrices in AsG and hence, according to Theorem 4,
mance channel w 7→ z. The goal is to use data {xk , uk }N for all closed-loop matrices ΣK X,U consistent with the data
k=0
of (1), affected by noise satisfying Assumption 3, in order (compare the discussion after Corollary 6). This implies that
to design K such that the closed-loop matrix AK is stable the closed-loop channel w 7→ z satisfies quadratic perfor-
and the following quadratic performance specification on (1) mance over an infinite time-horizon for arbitrary disturbance
is guaranteed for all AK ∈ ΣK X,U . inputs which are not required to satisfy a bound of the form
Definition 9. We say that the closed-loop system (1) with W ∈ W, compare Definition 9. In order to achieve this
state feedback uk = Kxk satisfies quadratic performance goal, a data trajectory of finite length and the (finite-horizon)
Q S assumption Ŵ ∈ W on the disturbance generating the data
with index P = , where R  0, if there exists an
S> R are sufficient. It is straightforward to extend Corollary 10
ε > 0 such that to design controllers with performance guarantees for the
X∞ 
wk
>   
Q S wk
∞
X channel wp 7→ z, where wp is an exogenous input different
≤ −ε wk> wk (15) from the noise perturbing the initial data trajectory, i.e., wp 6=
zk S > R zk
k=0 k=0 w. Further, following the same steps as in Remark 7, the
for all w ∈ `2 . nonlinear matrix inequality (17) can be transformed into (19),
which is an LMI in the variables Y = X −1 , M = GX −1 for
Important special cases of Definition 9 are Q =
a fixed multiplier λ. Thus, the proposed feasibility problem
−γ 2 I, S = 0, R = I for the H∞ -control problem and
can be solved via a line-search over λ.
Q = 0, S = −I, R = 0 for closed-loop strict passivity. Note
that the disturbance w enters the present problem setting D. Systems with partial model knowledge
in two different ways. First, it perturbs the measured input- We conclude the section by presenting an extension of the
state trajectory during the initial data generation for which proposed framework to systems with mixed data-driven and
w is bounded as W ∈ W. Second, it enters the control model-based components. To this end, we consider systems
objective of achieving quadratic performance of the channel of the form
w 7→ z. For instance, a desired H∞ -performance of this  
    xk
channel corresponds to a robustness objective for the closed xk+1 A1 A2 Bw1 B1 
 x̃k+1  =  A3 A4 Bw2 B2  x̃k , (20)

loop with respect to noise.
 wk 
Similar to Section III-B, the uncertain closed loop of (1), zk C1 C2 D w D
including the performance channel w 7→ z, can be written uk
as a lower LFT. To be more precise, for a state-feedback where the matrices A1 and B1 are unknown, but all other
gain K = U G, where G satisfies I = XG, a superset of the matrices occurring in (20) are known. Further, a single open-
−1
uncertain closed loop from w to z can be parametrized as loop data trajectory {xk , x̃k , uk }N
k=0 , which is perturbed by

xk+1
 
X+ G B w Bw

xk
 some unknown disturbance realization Ŵ ∈ W, is available.
 zk  =  C + DU G Dw 0  wk , In the following, we consider the closed loop of (20) under
z̃k −G 0 0 w̃k control with state-feedback uk = K1 xk + K2 x̃k . Suppose
there exist matrices G1 ∈ RN ×n , G2 ∈ RN ×ñ , where n and
w̃k = W z̃k , (16) ñ are the dimensions of xk and x̃k , respectively, such that
for W ∈ W. The above system contains two disturbance 
I 0
  
X  
inputs: w to model the performance channel w 7→ z, = G1 G2 . (21)
K1 K2 U
representing the control objective of closed-loop quadratic  
performance, and w̃ to model the uncertainty originating Multiplying (21) from the left by A1 B1 , we obtain
from the noisy data, similar to the LFT (11). The following
A1 + B1 K1 = (X+ − A2 X̃ − Bw1 Ŵ )G1 ,
result derives state-feedback controllers with robust perfor-
mance for (16). B1 K2 = (X+ − A2 X̃ − Bw1 Ŵ )G2 .

Corollary 10. If there exist X  0, G ∈ RN ×n , λ > 0, such These relations allow us to replace all occurrences of the
that (17) and unknown matrices A1 and B1 in the closed-loop dynamics.
Thus, following the same steps as in the previous sections,
XG = I (18) we obtain the LFT (22) with W ∈ W, which parametrizes

1536
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:16:15 UTC from IEEE Xplore. Restrictions apply.
 a superset of the uncertain closed loop dynamics of (20) where it is assumed that Atr and Btr are not available. We
under the above state-feedback. Note that this LFT depends generate data {xk , uk }N
k=0 of length N = 20 by sampling
only on known matrices and the open-loop data trajectory the input uk uniformly from [−1, 1]2 and the disturbance
−1
{xk , x̃k , uk }N
k=0 . The structure of (22) resembles that of ŵ uniformly from the ball kŵk2 ≤ w̄, where w̄ = 0.02.
the LFT (16) and therefore, robust controllers for the mixed This implies the disturbance bound Ŵ ∈ W for Qw =
system (20) can be derived by proceeding as in Section III-C. −I, Sw = 0, Rw = w̄2 I. In the following, we compute a
In contrast to the previous sections, the condition (21) state-feedback gain via Corollary 10 to achieve
 robust closed-
−γ 2 I 0

requires not only that X  has full row rank but also that loop quadratic performance with index P = for
>
N ≥ n+ñ. Moreover, if X > U > has full row rank, i.e., 0 I
the data-driven component of (20) is persistently exciting, a possibly small γ > 0, i.e., a small H∞ -norm of w 7→ z.
and N ≥ n + ñ, then, for any matrices K1 and K2 , there Following the procedure described in Remark 7, we verify
exist matrices G1 and G2 satisfying (21), i.e., any controller that (17) and (18) are feasible for γ = 2.4 and we obtain a
−2.45 −1.29 −2.4
can be constructed. corresponding controller as K = ,
−0.61 −0.03 −2.18
Remark 11. Our original motivation for considering the which leads to a closed-loop H∞ -norm of 2.3. In contrast,
above mixed data-driven and model-based configuration the minimal achievable H∞ -norm using a nominal (model-
comes from H∞ -loop-shaping: The H∞ -control problem is based) state-feedback is 2.2. Thus, the proposed approach
usually not solved for the performance channel w 7→ z yields a controller with guaranteed performance close to the
directly, but rather for the channel w 7→ z f , where z f is the ideal case with full model knowledge, despite noisy measure-
output of a filter with input z. In this scenario, the known ments. For larger noise levels w̄ ≥ 0.04, the design problem
components of (20) are mainly that of the filter, whereas the is infeasible since it addresses performance guarantees for
unknown matrices (A1 , B1 ) are equal to (Atr , Btr ) from (1). all matrices in the set ΣKA,B , which grows with w̄.
Notably, A2 = 0 holds in this case and hence, it can be
seen from (22) that measured data of the filter state x̃ is In the following, we analyze the influence of the data
not required. To conclude, by iteratively refining the filter length N on the feasibility of (17) and (18) for the above
dynamics and solving the robust performance design problem design problem. To keep the signal-to-noise ratio (approx-
for the LFT (22), we can systematically perform loop- imately) constant, we modify the bound w̄ of the noise
shaping for the system (1), without knowledge of (Atr , Btr ). generating the data linearly with N , i.e., w̄ = 0.0220 N . For
each data horizon 4 ≤ N ≤ 20, we perform 100 experiments
IV. E XAMPLE to generate data for the controller design, each with different
In this section, we apply the results of Section III to the (random) inputs u and disturbances ŵ as described above.
robust H∞ -control problem for an unstable example system. Figure 1 shows the number of successful designs depending
We consider System (1) with on N . It can be observed that the feasibility of (17) and (18)
    is enhanced if N increases, and N ≥ 15 suffices to success-
−0.5 1.4 0.4 0.1 −0.3
fully design a controller from 100 out of 100 experiments.
Atr = −0.9 0.3 −1.5 , Btr = −0.1 −0.7 ,
Intuitively, this can be explained by noting that, with an
1.1 1 −0.4 0.7 −1
increasing number of data points, the size of ΣX,U decreases,
Bw = I3 , C = I3 , D = 0, Dw = 0, and Corollary 10 provides robust performance guarantees

 >   
I 0 0 −X 0 0 0 0 0 I 0 0

 X+ G Bw Bw 


 0 X 0 0 0 0 
 X+ G Bw Bw 

 0 I 0   0 0 Q S 0 0  0 I 0 
≺0 (17)
S>
   
 C + DU G Dw 0   0 0 R 0 0  C + DU G Dw 0 
    
 0 0 I   0 0 0 0 λQw λSw  0 0 I 
>
−G 0 0 0 0 0 0 λSw λRw −G 0 0

(CY + DU M )> (RDw + S > ) −λM > Sw

>
M > X+>
(CY + DU M )> M>
 
−Y
 ∗ > > > >
 Q + SDw + Dw S + Dw RDw 0 Bw 0 0 

 ∗ >
∗ λQw Bw 0 0 
≺0

 ∗ (19)
 ∗ ∗ −Y 0 0 

 ∗ ∗ ∗ ∗ −R−1 0 
−1
∗ ∗ ∗ ∗ ∗ −(λRw )

1537
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:16:15 UTC from IEEE Xplore. Restrictions apply.
 for (a superset of) the uncertain closed loop matrices ΣK
X,U is extended to a setting with partial model knowledge, and
consistent with the data. Moreover, even for N as low as the design procedures are applied successfully to an unstable
4, in which case the data are not persistently exciting, the example system. The proposed approach leads to end-to-
design is successful in more than 50% of the scenarios. end guarantees for the closed loop, using a single noisy
open-loop data trajectory of finite length, and is thus a
promising alternative to sequential system identification and
100 robust control. Future research should extend the results of
this paper to robust data-driven output-feedback control.
90
R EFERENCES
[1] Z.-S. Hou and Z. Wang, “From model-based control to data-driven
80
control: Survey, classification and perspective,” Information Sciences,
vol. 235, pp. 3–35, 2013,
[2] B. Recht, “A tour of reinforcement learning: The view from contin-
70
uous control,” Annual Review of Control, Robotics, and Autonomous
Systems, 2018.
60 [3] N. Matni and S. Tu, “A tutorial on concentration bounds for system
identification,” arXiv preprint arXiv:1906.11395, 2019.
[4] N. Matni, A. Proutiere, A. Rantzer, and S. Tu, “From self-tuning
50 regulators to reinforcement learning and back again,” arXiv preprint
4 6 8 10 12 14 16 18 20 arXiv:1906.11392, 2019.
[5] R. Boczar, N. Matni, and B. Recht, “Finite-data performance guaran-
tees for the output-feedback control of an unknown system,” in Proc.
57th IEEE Conf. on Decision and Control, 2018, pp. 2994–2999.
Fig. 1. Number of successful designs for which (17) and (18) are [6] S. Dean, H. Mania, N. Matni, B. Recht, and S. Tu, “On the sample
feasible for the present example, depending on the data length N . For each complexity of the linear quadratic regulator,” Foundations of Computa-
horizon N , 100 experiments are carried out with varying random inputs and tional Mathematics, 2019, https://doi.org/10.1007/s10208-019-09426-
disturbances to generate data for controller design according to Corollary 10. y.
[7] M. Milanese and A. Vicino, “Optimal estimation theory for dynamic
Finally, we comment on the computational complexity systems with set membership uncertainty: an overview,” Automatica,
vol. 27, no. 6, pp. 997–1009, 1991.
of the feasibility problem stated in Corollary 10. After [8] R. L. Kosut, “Uncertainty model unfalsification for robust adaptive
its reformulation (cf. Remark 7), the problem contains an control,” Annual Reviews in Control, vol. 25, pp. 65–76, 2001.
LMI with 2(n + mw ) + pz + N rows, i.e., it is of size [9] J. C. Willems, P. Rapisarda, I. Markovsky, and B. De Moor, “A note
on persistency of excitation,” Systems & Control Letters, vol. 54, pp.
35 × 35 for the above example with N = 20, as well as 325–329, 2005.
an equality constraint of size n × n = 9. Moreover, the [10] A. Romer, J. Berberich, J. Köhler, and F. Allgöwer, “One-shot verifica-
matrix variables1 Y and M are of size n × n = 3 × 3 and tion of dissipativity properties from input-output data,” IEEE Control
Systems Letters, vol. 3, no. 3, pp. 709–714, 2019.
N × n = 20 × 3, respectively. The complexity of standard [11] J. Berberich and F. Allgöwer, “A trajectory-based framework for
LMI solvers scales cubically with the number of decision data-driven system analysis and control,” in Proc. European Control
variables. Thus, the proposed controller design method scales Conference, 2020, to appear, preprint online: arXiv:1903.10723.
[12] C. De Persis and P. Tesi, “Formulas for data-driven control: Stabiliza-
cubically with the data length N and proportionally to n6 if tion, optimality and robustness,” arXiv:1903.06842, 2019.
n is the system dimension, similar as in model-based robust [13] H. J. van Waarde, J. Eising, H. L. Trentelman, and M. K. Camlibel,
controller design. “Data informativity: a new perspective on data-driven system analysis
and control,” arXiv:1908.00468, 2019.
V. C ONCLUSION [14] K. Zhou, J. C. Doyle, and K. Glover, Robust and optimal control.
Prentice-Hall, Inc., Englewood Cliffs, N.J., 1996.
The present paper provides direct, data-driven design pro- [15] C. Scherer and S. Weiland, Linear Matrix Inequalities in Control,
cedures for state-feedback gains, which achieve guaranteed 3rd ed. New York: Springer-Verlag, 2000.
[16] C. Scherer, “Robust mixed control and linear parameter-varying con-
closed-loop stability and performance, using noisy input- trol with full-block scalings,” in Advances in Linear Matrix Inequality
state data. Based on a data-driven parametrization of the Methods in Control. SIAM: Philadelphia, 2000, pp. 187–207.
closed-loop matrices that are consistent with the data, known [17] M. G. Safonov and T.-C. Taso, “The unfalsified control concept and
learning,” IEEE Transactions on Automatic Control, vol. 42, no. 6, pp.
robust control methods can be applied. The parametrization 843–847, 1997.

1 Note that Y is symmetric and has therefore only n(n+1)

2
free decision
variables.

    
xk+1 (X+ − A2 X̃)G1 A2 + (X+ − A2 X̃)G2 Bw1 Bw1 xk
 x̃k+1   A3 + B2 U G1 A 4 + B 2 U G2 Bw2 0   x̃k 
 zk  =  C1 + DU G1
    
C2 + DU G2 Dw 0  wk  (22)
z̃k −G1 −G2 0 0 w̃k
w̃k = W z̃k

1538
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on December 18,2023 at 13:16:15 UTC from IEEE Xplore. Restrictions apply.