You are on page 1of 8

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO.

11, NOVEMBER 2023 6951

Data-Driven Self-Triggered Control via Trajectory Prediction


Wenjie Liu , Jian Sun , Senior Member, IEEE, Gang Wang , Member, IEEE,
Francesco Bullo , Fellow, IEEE, and Jie Chen , Fellow, IEEE

Abstract—Self-triggered control, a well-documented technique proposed, including e.g., reinforcement learning-based control [3],
for reducing the communication overhead while ensuring desired model-free control [4], and extremum seeking control [5]. More results
system performance, is gaining increasing popularity. However,
a majority of existing self-triggered control methods require ex- on data-driven control can be found in [6]. Most recently, the result of
plicit system models. An end-to-end control paradigm known as fundamental lemma in [7] has attracted reviving interest. This lemma
data-driven control designs control laws directly from data and provides a nonparameteric representation of a linear time-invariant
offers a competing alternative to the routine system identification- system using a trajectory of the system. Inspired by this work, a number
then-control strategy. In this context, the present article puts forth
data-driven self-triggered control schemes for unknown linear sys-
of applications and generalizations have been made, including stabi-
tems using input–output data collected offline. Specifically, a data- lization and optimization [8], [9], [10], linear quadratic regulation [11],
driven model predictive control (MPC) scheme is proposed, which robust control [12], quantized control [13], model predictive control
computes a sequence of control inputs while generating a pre- (MPC) [14], [15], [16], [17], consensus control [18], [19], and control
dicted system trajectory. In addition, a data-driven self-triggering of complex networks [20].
mechanism is designed, which determines the next triggering time
using the solution of the data-driven MPC and the newly collected Yet, the aforementioned works employ periodic transmission proto-
measurements. Finally, both feasibility and stability are established cols, which may be resource inefficient for real-world systems in terms
for the proposed self-triggered controller, which are validated us- of processor usage, communication bandwidth, and energy. In cyber-
ing a numerical example. physical networked systems [21], for instance, whose communication
Index Terms—Data-driven control, data-driven model predictive network is shared by many devices, the communication bandwidth is al-
control (MPC), predicted control, self-triggered control. ways restricted for each device. To tackle this issue, a resource-efficient
scheduling approach for data transmissions, known as event-based
control, has been widely studied in the context of model-based control.
I. INTRODUCTION There are two effective event-based approaches, namely, event-
Thanks to recent advances in data acquisition and computing tech- triggered control and self-triggered control [22], [23]. In the former,
nologies, data-driven control has attracted considerable attention in the an event, e.g., a data transmission, is triggered only after some trig-
past years. Designing control laws directly from data, without resorting gering condition, in terms of the state/output measurements, is met.
to any system identification step, offers an appealing alternative to the This condition is usually continuously or periodically checked, and
traditional model-based control paradigm. This is because in real-world the sensor should be activated at these checking times. While a level
applications, it is always difficult or even impossible to acquire an of robustness against uncertainties and unmodeled dynamics can be
accurate system model [1], [2]. Several data-driven methods have been expected, having the sensor frequently operating results in waste of
resources. Related work can be found in [24]. Self-triggered control,
on the other hand, outputs the next sampling and transmission once
Manuscript received 16 October 2022; accepted 3 February 2023. a sampled measurement is received, eliminating the need for contin-
Date of publication 10 February 2023; date of current version 26 Oc- uously or periodically sampling, see e.g., [25], [26], [27], and [28].
tober 2023. The work was supported in part by the National Key R&D
Program of China under Grant 2021YFB1714800, in part by the National Therefore, the sensors in self-triggered control can be completely shut
Natural Science Foundation of China under Grant 61925303, Grant off between sampling times, which saves energy and prolongs service
62173034, and Grant 62088101, in part by the BIT Research and Inno- life of the sensor. This feature appears promising in the model-based
vation Promoting Project under Grant 2022YCXZ007, and in part by the case and motivates us to pursue the generalization of self-triggered from
Chongqing Natural Science Foundation under Grant 2021ZX4100027.
Recommended by Associate Editor I. Necoara. (Corresponding author:
model-based to data-driven settings.
Gang Wang.) Recently, Wang et al. [29] proposed a data-driven self-triggering
Wenjie Liu, Jian Sun, and Gang Wang are with the National Key Lab of mechanism for unknown state feedback systems. A data-based param-
Autonomous Intelligent Unmanned Systems and School of Automation, eterization of lifted system matrices is used to represent the sequence
Beijing Institute of Technology, Beijing 100081, China, and also with the of predicted states, which plays an important role in constructing
Beijing Institute of Technology Chongqing Innovation Center, Chongqing
401120, China (e-mail: liuwenjie@bit.edu.cn; sunjian@bit.edu.cn; gang- the self-triggering mechanism. However, when only input–output data
wang@bit.edu.cn). are available, neither the state feedback control law nor the lifted
Francesco Bullo is with the Mechanical Engineering Department matrix method can be applied directly. The goal of this article is to
and the Center of Control, Dynamical Systems and Computation, design self-triggered controllers for unknown linear systems using only
UC Santa Barbara, Santa Barbara, CA 93106 USA (e-mail: bullo@
engineering.ucsb.edu).
input–output data. An event of sampling and transmission is triggered
Jie Chen is with the National Key Lab of Autonomous Intelligent if a predicted output satisfies some triggering conditions. To predict
Unmanned Systems and School of Automation, Beijing Institute of Tech- future outputs while compensating for the network-induced noise in the
nology, Beijing 100081, China, and also with the Department of Control absence of a system model, a data-driven MPC scheme, combining that
Science and Engineering, Tongji University, Shanghai 201804, China of [30] and a special case of [14], is developed. Leveraging the predicted
(e-mail: chenjie@bit.edu.cn).
Color versions of one or more figures in this article are available at trajectory, a data-driven self-triggering mechanism, accounting for both
https://doi.org/10.1109/TAC.2023.3244116. the recursive feasibility of the data-driven MPC and the system stability,
Digital Object Identifier 10.1109/TAC.2023.3244116 is designed.

0018-9286 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: University of Chinese Academy of SciencesCAS. Downloaded on November 16,2023 at 15:26:50 UTC from IEEE Xplore. Restrictions apply.
6952 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 11, NOVEMBER 2023

In succinct form, the contributions of this work are summarized as


follows.
c1) A self-triggered MPC is proposed using noise-corrupted input–
output data, where the triggering mechanism relies on the optimal
solution of a data-driven MPC scheme.
c2) Both recursive feasibility of the data-driven MPC and the input-
to-state stability (ISS) of the self-triggered system are established.
Notations: Denote the set of real numbers (natural numbers) by R
(N), and define N0 := N ∪ {0}. Given α ∈ N, let Nα denote the set of
natural numbers greater than or equal to α. For a matrix M , if it has full
column rank (full row rank), its left pseudoinverse (right pseudoinverse)
is denoted by M † (M ‡ ). Given a vector x ∈ Rnx , x is its Euclidean
norm and x∞ is its infinity norm, and for a positive
√ definite matrix
P = P   0, define the weighnorm xP = x P x. Let M  be
the spectral norm of matrix M . Let λP (λP ) represent the minimum Fig. 1. Pictorial description of system (1) with data-driven self-
(maximum) eigenvalue of matrix P . Let [t1 , t2 ] denote the time interval triggered MPC.
from t1 to t2 with discrete times. A function α : [0, ∞) → [0, ∞) is said
to be of class K if it is continuous, strictly increasing, and α(0) = 0. A
function α : [0, ∞) → [0, ∞) is said to be of class K∞ if it is of class K
and also unbounded. A function β : [0, ∞) × [0, ∞) → [0, ∞) is said In networked control systems, the plant is connected with a sensor
to be of class KL if β(·, t) is of class K for each fixed t ≥ 0 and β(s, t) that takes the output measurement at sampling times. Sampled mea-
decreases to 0 as t → ∞ for each fixed s ≥ 0. surements are transmitted to a remote controller to construct control
N −1
The Hankel matrix associated with the sequence {xt }t=0 is denoted inputs. Frequent sampling and transmission inevitably results in waste
by of energy and network resources. To save the network bandwidth as well
⎡ ⎤ as prolong the life-cycle of the sensor, we incorporate a self-triggering
x0 x1 . . . xN −L
⎢ x1 module running at the controller side, to dictate when a new output
⎢ x2 . . . xN −L+1 ⎥ ⎥
HL (x) := ⎢ . .. .. ⎥ measurement should be sampled and transmitted.
⎣ .. .. ⎦
. . . Specifically, we consider that the output is transmitted over a network
xL−1 xL . . . xN −1 of limited bandwidth to the remote controller, while the controller-to-
plant channel is assumed perfect either over wired lines or networks
N −1
where x is used to represent the sequence {xt }t=0 for brevity. A with sufficient bandwidth. Furthermore, to model the network effect,
stacked window of a sequence, say {xt }tt=t
2
1
is given by x[t1 ,t2 ] = the output is corrupted by additive noise obeying nt  ≤ n̄ for some
[x
t1 · · · x  
t2 ] . known upper bound n̄ when transmitted over the network, i.e.,

II. PRELIMINARIES AND PROBLEM FORMULATION


ζt = yt + nt . (3)
A. Networked Control Systems
Consider the following discrete-time linear system: See Fig. 1 for a pictorial depiction of the system structure.
Let t ∈ N0 stand for the time at which the th event (i.e., sampling
xt+1 = Axt + But , t ∈ N0 (1a) and transmission) is triggered by the self-triggering module. Assume
yt = Cxt + Dut (1b) the sensor is equipped with a buffer of size η, which records the
most recent η historical measurements. Once an event is triggered,
nx nu ny
where xt ∈ R , ut ∈ R , and yt ∈ R are the state, control input, and the data packet consisting of the η buffered measurements is sent to
output, respectively. In this article, we make the following assumptions the controller. The self-triggering module computes the next triggering
on the system (1). time by evaluating the triggering function on a predicted trajectory
Assumption 1 (Controllability and observability): The pair (A, B) along with the received data packet and sends it back to the sensor.
is controllable, and (C, A) is observable. For example, at time t , the th transmission has just been made,
Assumption 2 (Unknown system model): The system matrices and our self-triggering module receives the data packet y[t −η,t −1] :=
(A, B, C, D) in (1) are unknown. Instead, some precollected input– {yt −η , yt −η+1 , · · ·, yt −1 }, and computes the next triggering time
output data, i.e., {upt , ytp }t=0
N −1
are available. t+1 , which is fed back to the sensor. During the interval [t−1 , t ),
Regarding the two assumptions, a remark comes ready. it is only required that the sensor samples output measurements for the
Remark 1: Assumption 1 is standard for stability analysis of linear subinterval [t − η, t ). Of course, if the intertriggering time t − t−1
systems. In the data-driven setting, it can be checked following [9, is smaller than η, i.e., t − t−1 ≤ η, one only needs to send the
Sec. V]. packet of the new t − t−1 measurements as the old η − (t − t−1 )
We further recall the definition of observability index [31]. ones have already been sent to the controller in the last packet.
Definition 1 (Observability index): The observability index of a lin- Furthermore,
ear system as in (1) is defined to be the smallest integer η ∈ {1, . . . , nx }
such that the observability matrix Θ has full rank nx , i.e.,
 B. Fundamental Lemma
rank(Θ) = rank [C  (CA) · · · (CAη−1 ) ] = nx . (2)
In this section, we briefly review the data-driven system represen-
Assumption 3 (Upper bound on the observability index): An integer tation based on the so-called fundamental lemma [7], which is key
η ∈ N such that η ≥ η is known. to derive our data-driven self-triggering predictive controllers. Before

Authorized licensed use limited to: University of Chinese Academy of SciencesCAS. Downloaded on November 16,2023 at 15:26:50 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 11, NOVEMBER 2023 6953

presenting the lemma, the definition of persistence of excitation is the output at the most recent triggered time. However, when the system
introduced first. matrices are unknown, there are two challenges: c1) How to design a
Definition 2 (Persistence of excitation): A sequence {ut ∈ self-triggered controller using only input–output data that are corrupted
N −1
Rnu }t=0 is persistently exciting of order L if rank(HL (u)) = by network-induced noise? c2) How to analyze the performance for
nu L. the data-driven self-triggered closed-loop system? We address c1) by
Based on Definition 2, it has been shown in [7] that any trajectory first putting forward a data-driven MPC scheme, which can compute
of system (1) can be expressed as a linear combination of precol- a window of control inputs and associated predicted outputs once a
lected input–output data {upt , ytp }t=0
N −1
, provided that the input sequence noise-corrupted output measurement is transmitted and received. Next,
p N −1
{ut }t=0 is persistently exciting. This result, formally summarized in based on the optimal solution of the MPC, a self-triggering mechanism
the following, is also known as the fundamental lemma [7]. is designed. Finally, recursive stability of the data-driven MPC along
Lemma 1 (Fundamental lemma): Consider a persistently exciting with the ISS of the self-triggered control system is established.
input sequence {upt }t=0
N −1
of order L + nx . A trajectory {ūt , ȳt }L−1
t=0 is
generated by system (1) if and only if there exists a vector g ∈ RN −L+1
A. Data-Driven MPC Scheme
such that the following holds:
Although Lemma 1 characterizes system trajectories in some sense,
HL (up ) ū
g = [0,L−1] . (4) the correspondence between the input–output trajectory and the coeffi-
HL (y p ) ȳ[0,L−1]
cient vector is not unique in general. In addition, the fundamental lemma
It can be concluded from Lemma 1 that to generate a predicted does not account for any noise, so it cannot be directly employed for
trajectory of length L, the precollected data should be persistently control and trajectory prediction. Building on existing works [14], [17],
exciting of at least order L. To validate this requirement, the next we propose a new data-driven MPC that, at each triggered time t , com-
assumption is imposed. putes a window of optimal control inputs and associated outputs, using
Assumption 4 (Precollected data): Let {upt }t=0 N −1
be any sequence the received noisy data ζ[t −η,t −1] . The control inputs are subsequently
persistently exciting of order L + nx + η. The output sequence employed one by one at times t ∈ [t , t+1 ). Let vectors ū(t ) :=
{ytp }i=0
N −1
is generated from system (1) offline with any initial condition [ū     
−η (t ) · · · ūL−1 (t )] and ȳ(t ) := [ȳ−η (t ) · · · ȳL−1 (t )] denote
x0 , and input sequence {upt }t=0
p N −1
. the predicted input and output from time t − η to t + L − 1,
For brevity, we use up and y p to represent the sequences {upt }t=0
N −1
and ξ̄(t ) = [ξ¯0 (t ) · · · ξ̄L

(t )] the corresponding extended state
and {ytp }t=0
N −1
in the following. Based on Assumption 3, we construct with
an extended state for t ≥ η as follows:
ū[i−η,i−1] (t )
ξ̄i (t ) =
u ȳ[i−η,i−1] (t )
ξt := [t−η,t−1] ∈ Rnξ (5)
y[t−η,t−1]
where i ∈ {0, 1, · · ·, L}. Mathematically, the following optimization
where nξ := (nu + ny )η. It follows iteratively from (1) that problem is solved at each self-triggered time t :
yt = CAη xt−η + CAη−1 But−η · · · + CBut−1 + Dut . JL∗ (ut , yt ) :=
Recalling from Definition 1, there exist some matrices Υ, Υ1 , and Υ2 L−1
λh
such that xt−η = Υ1 u[t−η,t−1] + Υ2 y[t−η,t−1] = Υξt . Hence, system min ūi (t ) − ue 2R + ȳi (t ) − y e 2Q + h(t )2
{g(t ),h(t )
i=−η

(1) can be transformed into ȳ(t ),ū(t )}

ξt+1 = Ãξt + B̃ut , t ∈ Nη (6a) + λg n̄g(t )2 + ξ¯L (t ) − ξ e 2P (8a)
p
yt = C̃ξt + D̃ut (6b) ū(t ) HL+η (u )
s.t. = g(t ) (8b)
ȳ(t ) + h(t ) HL+η (y p )
for suitable matrices (Ã, B̃, C̃, D̃) depending on (A, B, C, D). For
ū[−η,−1] (t ) u
subsequent analysis, let us denote the equilibrium point of the new = [t −η,t −1] (8c)
system (6) by ȳ[−η,−1] (t ) ζ[t −η,t −1]

u[0,η−1] ξ¯L (t ) ∈ Ξ (8d)


ξ e := (7)
y[0,η−1]
ūi ∈ U, i ∈ {1, 2, . . ., L − 1}. (8e)
in which ui = ue and yi = y e for all i = 0, 1, . . . , η − 1. As a matter of
fact, such an augmented system (state) has been commonly employed It is worth remarking that the data-driven robust MPC formulation
in studies of data-driven MPC to simplify the stability analysis; see (8) can be seen as a combination of [30] and a special case of [14].
e.g., [14]. However, the motivation, controller design, and associated stability
Remark 2: Since η ≥ η, it follows from the observability of (C, A) analysis in this work are different from those in [14] and [30]. Prese-
that (1) and (6) have equivalent input–output properties. This implies lected weighting matrices R  0 and Q  0 penalize the distances from
that the following proposed self-triggering mechanism is valid for all the predicted input and output to their equilibrium points. Constraint
η ≥ η. (8b) reflects the data-driven system representation in Lemma 1, where
g(t ) ∈ RN −L−η+1 stands for the vector g at time t . The vector
h(t ) = [h  
−η (t ) · · · hL−1 (t )] is added to compensate for the un-
III. DATA-DRIVEN SELF-TRIGGERING MPC
known network-induced noise nt in predicting future outputs, since
This section advocates a data-driven approach to designing a con- the left-hand side of constraint (8b) does not represent an input–output
troller and a self-triggering mechanism for unknown linear systems as trajectory of system (1). Penalties are also imposed on the use of g(t )
in (1). A model-based self-triggering mechanism determines the next and h(t ) in the objective function, to guarantee system stability and
transmission by comparing the predicted future output sequentially with improve robustness against noise, with weights λh > 0 and λg > 0

Authorized licensed use limited to: University of Chinese Academy of SciencesCAS. Downloaded on November 16,2023 at 15:26:50 UTC from IEEE Xplore. Restrictions apply.
6954 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 11, NOVEMBER 2023


balancing between minimizing different costs. Feasibility sets U and Ξ
√ 
τ +η−2
are prescribed and convex. Finally, (8d) is a terminal constraint, which ≤ ηn̄ + h∗[−η,−1] (t )  ρi (10b)
together with the last summand having P  0 in (8a) ensures that the i=τ −1
last predicted extended state ξ¯L (t ) (at the end of the window) stays in
Ξ (i.e., a small neighborhood of the desired equilibrium point). Overall, + h∗[τ −η,τ −1] (t ) (10c)
Problem (8) is convex and can be efficiently solved using off-the-shelf
solvers. where ρi := CAi+η Θ†  for i = τ − 1, τ , . . . , τ + η − 2.
The terminal constraint (8d) and the terminal cost ξ¯L (t ) − ξ e 2P The proof can be proceeded following the steps as in [14, Lemma
are also critical ingredients in the model-based MPC, see e.g., [32]. To 2], and is thus omitted here due to space limitations.
proceed, we make the following assumption on the terminal ingredients According to Lemma 2, an upper bound on the error between the
Ξ : {ξ ∈ Rnξ |ξ − ξ e P ≤ } with Ξ ⊂ Ξr and the corresponding predicted extended state and the actual extended state is characterized
matrix P . in terms of known or computable parameters h∗ (t ), n̄, and unknown
Assumption 5: There exist matrices P = P   0, K ∈ Rnu ×nξ , parameters ρi for i = τ , . . . , τ + η − 1. Before running the system
and a set Ξr : {ξ ∈ Rnξ |ξ − ξ e P ≤ r} ⊆ Uη × Rη such that for all online, upper bounds of parameters ρi can be obtained offline by solving
ξ ∈ Ξr , u = ue + K(ξ − ξ e ), and y = (C̃ + D̃K)ξ, the following the optimization problem proposed in [16, Sec. V.B] for i = 1,. . . , L−1
statements hold true:
Ji := max yi ∞ (11a)
1) u ∈ U, Ãξ + B̃u ∈ Ξr , and y,g
2) the following inequality holds
s.t. y[0,η−1] ∞ ≤ 1 (11b)

(Ã + B̃K)ξ2P ≤ ξ2P − Kξ2R − y2Q . Hi+1 (up )  0


(9) g = (11c)
Hi+1 (y p ) y[0,i]

The restrictiveness of this assumption as well as data-based method where Hi+1 (up ) denotes the first (i + 1)nu rows of matrix HL+η (up ),
for selecting the matrices P and K and the set Ξr has been discussed and likewise for HL+η (y p ), vector g ∈ RN −L−η+1 , and y[0,i] consists
in Proposition 10 and in the following [30, Assumption 5]. It will be of the i + 1 predicted outputs. It has been shown in [16] that Ji ≥
shown in the next section that these terminal ingredients have influence CAi+η Φ† , and hence, an upper bound for ρi is obtained, i.e., Ji ≥ ρi .
on the intertriggering time, and should be handled with care. Therefore, one gets an upper bound on ξe (t+1 ).
Remark 3 (Terminal ingredients): Generally, there are two ways for Remark 4 (Noisy offline data): Noise-free offline input–output data
guaranteeing the stability of an MPC scheme: i) providing bounds on are required for solving Problem (11). Indeed, this assumption can
the minimal required prediction horizon [33], [34] and ii) including be restrictive and may be violated in practice. Nevertheless, various
terminal ingredients, such as terminal cost functions or terminal region approaches have been proposed for denoising, such as
constraints. However, a larger prediction horizon requires more precol- i) filtering the collected input–output data with a high-pass or low-
lected data and increases the online computational burden. Hence, we pass filter, e.g., [35, Sec. 10.3.3],
consider in this article MPC with terminal ingredients. The results can ii) constructing Hankel matrices by averaging [11], or
be extended by adopting MPC schemes without terminal ingredients. iii) using the page matrix [36].
According to (8c), a lower bound on the prediction window as Designing self-triggering mechanisms from noisy input–output data
follows. is beyond the scope of this article and is left for future research.
Assumption 6 (Prediction horizon: The prediction horizon of the
data-driven MPC satisfies L ≥ η + 1. B. Self-Triggering Mechanism
Without loss of generality, we consider for simplicity the equi-
librium ξ e = [(ue ) (y e ) ] = 0 in the rest of this article. Let Leveraging the upper bound in Lemma 2, our self-triggering mech-
(ȳ ∗ (t ), ū∗ (t ), g ∗ (t ), h∗ (t )) denote any optimal solution of Problem anism is given in the following lemma.
(8) at t , thereby yielding the predicted extended state ξ̄ ∗ (t ). The Lemma 3 (Recursive feasibility): Let Assumptions 1–6 hold. As-
closed-loop control input ut is given by ut = ū∗t−t (t ) sequentially for sume further that (8) is feasible at t0 . For appropriate r > 0 and P  0
all t ∈ [t , t+1 − 1]. Let ut , yt , and ξt represent the actual input, satisfying Assumption 5, there exist constants λ̄g > λg > 0, λ̄h >
output, and extended state, respectively, at time t . The intertrigger- λh > 0, and n̄0 > 0 such that for all λg ≤ λg ≤ λ̄g , λh ≤ λh ≤ λ̄h ,
ing time between two consecutive self-triggered times is defined as and 0 ≤ n̄ < n̄0 , Problem (8) is feasible at all t ∈ N0 , if i)  in (8d)
τ := t+1 − t for  ∈ N0 , with t0 = 0. Observe that at most L − 1 obeys
future outputs can be obtained by solving (8) at time t . Therefore,  
λ 
the intertriggering time between any two consecutive self-triggered 1 − K RK r2 ≤ 2 ≤ r2 (12)
times obeys 1 ≤ τ ≤ L − 1 for all  ∈ N0 . With these definitions, the λ̄P
following lemma indicates that the error between the predicted extended and ii) the intertriggering time satisfies
state ξ̄ ∗ (t ) and the actual extended state ξt is bounded, and it can be
rigorously quantified using the optimizer of (8). τ = min {τ̂ , τ̌ , L − 1} (13)
Lemma 2 (Prediction error): Let Assumptions 1–6 hold. Consider
system (1) with the control input is given by ut = ū∗t−t (t ), t ∈ with integers τ̂ and τ̌ defined by
[t , t+1 − 1], where ū∗ (t ) is the optimal solution of Problem (8) at  
√
each self-triggered time t . For every  ∈ N0 , the error between ξt+1  τ +η−1
τ̂ := sup τ ∈ N( ηn̄ + h∗[−η,−1] (t )) ρi
and ξ̄τ∗ (t ) satisfies i=τ

r
+ h∗[τ −η,τ −1] (t ) ≤ − ξ¯τ∗ (t ) (14)
ξe (t+1 ) := ξt+1 − ξ¯τ∗ (t ) (10a) λ̄P

Authorized licensed use limited to: University of Chinese Academy of SciencesCAS. Downloaded on November 16,2023 at 15:26:50 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 11, NOVEMBER 2023 6955

and J¯L (ξt ). The optimizer of Problem (8) at t is denoted by


  τ −1   (ū∗ (t ), ȳ ∗ (t ), g ∗ (t ), h∗ (t )). The difference of the Lyapunov func-
 ∗ 2 h∗ (t )2
τ̌ := sup τ ∈ N2λ̄Q hi (t ) +λh ηn̄− tion between two consecutive self-triggered times satisfies
i=0

  V (ξt+1 ) − V (ξt )
 √  τ +η−2
+ λ̄P (r + ) ηn̄ + h∗[−η,−1] (t ) ρi = J¯L (ξt+1 ) + γW (ξt+1 ) − JL∗ (ξt ) − γW (ξt )
i=τ −1
  
‡ ≤ J¯L (ξt+1 ) − JL∗ (ξt ) + γ W (ξt+1 ) − W (ξt ) .
+h∗[τ −η,τ −1] (t ) + λg n̄ Huξ 2 (2 + λ̄P /λR )r2      
=T1 =T2
 τ −1
∗ 2 2
−g (t ) + λ̄Q 2ηn̄ +2h∗[−η,−1] (t )2 (ρi+η )2 Next, we derive upper bounds on these two terms. Part a: Upper-
i=0 bounding T1 . According to (8a), it can be obtained that
τ −1  L−1
≤σ ξ¯i∗ (t )2 
(15) T1 ≤ ūi (t+1 )2R + ȳi (t+1 )2Q
i=0 i=0
where matrix Huξ is defined by + λg n̄g(t+1 )2 + (λh /n̄)h(t+1 )2 + ξ¯L (t+1 )2P
HL+η (up )
Huξ = (16) L−1
 ∗
H1 {ξtp }t=η
N −L
− ūi (t )2R + ȳi∗ (t )2Q + λg n̄g ∗ (t )2
and constant σ balances between the system performance and the i=0

intertriggering time satisfying σ ∈ (0, σ̄) with σ̄ ∈ (0, 1).


Proof: First, for some  ∈ N0 , we assume that Problem (8) is feasi- + (λh /n̄)h∗ (t )2 + ξ¯L

(t )2P . (19)
ble at t and construct a candidate solution at t+1 . Then, leveraging a
carefully designed Lyapunov function, recursive feasibility of Problem In addition, since for all i = 0, . . . , L − 1, input is chosen such that
¯ +1 ), it can be deduced recursively from (9) that
ūi (t+1 ) = K ξ(t
(8) is proven by showing the decrease of the Lyapunov function.
Suppose that (8) is feasible at t for some  ∈ N0 . Denote a L−1

candidate solution of (8) at t+1 by ȳ(t+1 ), ū(t+1 ), g(t+1 ), ūi (t+1 )2R + ȳi (t+1 )2Q ≤ ξt+1 2P −ξ¯L (t )2P .
and h(t+1 ). It follows from (8c) that ūi (t + 1) = ū∗τ +i (t ) and i=0
ȳi (t+1 ) = ζt+1 +i for i = −η, . . . , −1. Using (10), one gets from (14) Therefore,
that ξt+1 P ≤ ξe (t+1 )P + ξ¯τ∗ (t )P ≤ r, namely ξt+1 ∈ Ξr .
L−1
Thus, it follows from Assumption 5 that if ūi (t+1 ) = Kξi+t+1 
ūi (t+1 )2R +ȳi (t+1 )2Q +ξ¯L (t+1 )2P − ξ¯L

(t )2P
for all i = 0, . . . , L − 1, its corresponding state ξt+1 +i ∈ Ξr for all
i=0
i = 0, . . . , L − 1. Let ȳi (t+1 ) = yi+t+1 for i = 0, · · ·, L − 1. Notic- 
ing from Lemma 1, the right-hand side of Constraint (8b) character- ≤ ξt+1 2P − ξ¯L

(t )2P ≤ λ̄P (r + )ξe (t+1 ). (20)
izes a real system input–output trajectory. Choose the slack variable
h(t+1 ) as h[−η,−1] (t+1 ) = n[t+1 −η,t+1 −1] and hi (t+1 ) = 0 for In addition, since ū[−η,−1] (t+1 ) = [I 0]ξt , it follows from (17) that:
i = 0, 1, . . . , L − 1, and set ‡
g(t+1 )2 ≤ Huξ 2 (ū[0,L−1] (t )2 + 2ξt+1 2 )
‡ u[−η,L−1] (t+1 ) ‡
g(t+1 ) = Huξ (17) ≤ Huξ 2 (2 + λ̄P /λR )ξt+1 2
ξt+1

with Huξ in (16). Constraint (8b) is satisfied. In addition, it follows ≤ Huξ 2 (2 + λ̄P /λR )r2 . (21)
recursively from (9) that: Note that h[−η,−1] (t+1 ) = n[t+1 −η,t+1 −1] and hi (t+1 ) = 0 for i =
ξ¯L (t+1 )2P ≤ ξ¯L−1 (t+1 )2P − ξ¯L−1 (t+1 )2K  RK 0, . . . , L − 1, one has that

≤ [1 − λK  RK /λ̄P ] ξ¯L−1 (t+1 )2P (λh /n̄)h(t+1 )2 ≤ λh ηn̄. (22)

≤ [1 − λK  RK /λ̄P ] ξ¯0 (t+1 )2P ≤ 2


L Substituting (20)–(22) into (19) yields
 
√  τ +η−2
where the last inequality is due to condition (12), and hence, Constraint T1 ≤ λ̄P (r + ) ηn̄ + h∗[−η,−1] (t ) ρi
(8d) is met. Therefore, if Problem (8) is feasible at t , it is feasible at i=τ −1

t+1 . 
Next, we construct a Lyapunov function to show the recursive + h∗[τ −η,τ −1] (t ) + λh (ηn̄ − h∗ (t )2 /n̄)
feasibility. Since (C, A) is observable, (C̃, Ã) is detectable. It follows 
from [37, Prop. 3.3 ] that, for a detectable system, there exists an ‡
+ λg n̄ Huξ 2 (2 + λ̄P /λR )r2 − g ∗ (t )2
input-output-to-state stability Lyapunov function W (ξ) = ξ  Pξ ξ such
that L−1
 ∗
2 2 2 − ūi (t )2R + ȳi∗ (t )2Q . (23)
W (Ãξ+ B̃u)−W (ξ) ≤ −(1/2)ξ +c1 u +c2 y (18) i=0

for suitable c1 , c2 > 0, Pξ  0, and all u ∈ Rnu , ξ ∈ Rnξ , and Part b: Upper-bounding T2 . It follows recursively from (18) that:
y = C̃ξ + D̃ξ. Consider the Lyapunov function V (ξt ) = JL∗ (ξt ) +
γW (ξt ) having γ = min{λQ , λR }/max{c1 , 2c2 }, where JL∗ (ξt ) W (ξt+1 ) − W (ξt )
is used to represent JL∗ (ut , yt ) for brevity. For the candidate 1
solution (ū(t+1 ), ȳ(t+1 ), g(t+1 ), h(t+1 )), denote its cost by ≤ − ξ[t ,t+1 −1] 2 + c1 ū∗[0,τ −1] (t )2 + c2 y[t ,t+1 ] 2
2
Authorized licensed use limited to: University of Chinese Academy of SciencesCAS. Downloaded on November 16,2023 at 15:26:50 UTC from IEEE Xplore. Restrictions apply.
6956 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 11, NOVEMBER 2023

1
≤ − ξ[t ,t+1 −1] 2 +c1 ū∗[0,τ −1] (t )2 +2c2 ȳ[0,τ

(t )2
 −1] 
Algorithm 1: Data-Driven Self-Triggered Predictive Control.
2

1: Input: Prediction window L ≥ η + 1; coefficients R  0,
+ 2c2 y[t ,t+1 ] − ȳ[0,τ (t )2 .
 −1] 
(24) Q  0, P  0, λg > 0, λh > 0, σ ∈ (0, 1),  > 0, and r > 0
Since γ = min{λQ , λR }/max{c1 , 2c2 }, we have that satisfying (12); noise bound n̄; data {up[0,N −1] , y[0,N
p
−1] } of
p p
(1) from initial condition x0 , with u[0,N −1] persistently
γ(c1 ū∗[0,τ −1] (t )2 +2c2 ȳ[0,τ

(t )2 )
 −1]  exciting of order L + η + 1.
τ −1 2: Construct Hankel matrices for the input, output, and the

≤ ū∗i (t )2R + ȳi∗ (t )2Q . (25) extended state trajectories, i.e., HL+η (up ), HL+η (y p ), and
 p
i=0 Huξ := [HL+η (up ) H1 (ξN 
−L−η+1 )] .
3: Compute ρi for i ∈ {1, 2, . . . , L − 1} from (11).
Substituting (25) into (24) and leveraging the result in Lemma 2, we
4: If t = t , do
arrive at
5: Use the past η measurements, i.e., u[t−η,t−1] and ζ[t−η,t−1] ,
τ −1
γ to solve Problem (8).
T2 ≤ − ξ[t ,t+1 −1] 2 + 2λ̄Q h∗i (t )2 Determine the next self-triggered time t+1 such that the
2 i=0
intertriggering time obeys (13)–(15).
τ −1
 ∗ Set ut = ū∗0 (t).
+ ūi (t )2R + ȳi∗ (t )2Q Set t = t+1
i=0
6: Else if t = t
 τ −1 7: Set ut = ū∗t−t (t ).
+ λ̄Q 2ηn̄2 + 2h∗[−η,−1] (t )2 (ρi+η )2 . (26) 8: End if
i=0
9: Set t = t + 1 and go back to 4.
Combining the results in Parts a and b, it holds that
τ −1
γ  ∈ N0 , the proposed scheme becomes trivial. In the literature, analysis
V (ξt+1 ) − V (ξt ) ≤ − ξ[t ,t+1 −1] 2 + 2λ̄Q h∗i (t )2
2 i=0 on the minimum intertriggering time is often ignored, e.g., [26] an
 
√  τ +η−2
[28]. Only a few works considered this problem by either posing the
+ λ̄P (r + ) ηn̄ + h∗[−η,−1] (t ) ρi assumption τ ≥ 2 [38], or proving τ ≥ 2 for all  ∈ N0 in [39].
i=τ −1
 In the present article, for small enough noise n̄, if parameters r,
P , and σ are carefully chosen, it can be deduced that τ ≥ 2 for
+ h∗[τ −η,τ −1] (t ) + λh (ηn̄ − h∗ (t )2 /n̄)
all  ∈ N0 . To see this, consider for simplicity there is no noise,
 i.e., n̄ = 0. If r and P are chosen such that ξt0 ∈ Ξr , then condi-

+ λg n̄ Huξ 2 (2 + λ̄P /λR )r2 − g ∗ (t )2 tion (14) holds for all τ ∈ {1, . . . , L − 1} with  ∈ N0 . This im-
 τ −1 plies that there exist n̄0 ≥ 0 and 1 < τ1 ≤ L − 1 such that if n̄ ∈
+ λ̄Q 2ηn̄2 + 2h∗[−η,−1] (t )2 (ρi+η )2 . (27) [0, n̄0 ), condition (14) holds for all τ̂ ∈ {1, . . . , τ1 }, and hence τ̂ =
i=0 τ1 ≥ 2. In addition, it can be extended from the proof of [40, Th. 2]
that even if the inputs in sequence ū∗[0,L−1] (t ) are applied sequen-
In addition, (13) indicates that τ ≤ τ̌ , and hence (27) obeys
tially for t ∈ {t , . . . , t + L − 1}, there exists γ3 ∈ (0, 1) such that
V (ξt+1 ) − V (ξt ) V (ξt +L−1 ) ≤ γ3 V (ξt ). This indicates that for small enough n̄, there
  τ −1  exist σ ∈ (0, 1) and 2 ≤ τ2 ≤ L − 1 such that condition (15) is satisfied
γ γ for all τ̌ ∈ {1, . . . , τ2 } and τ̌ = τ2 . Therefore, the intertriggering time
≤− −2σ ξt 2 + 2σηn̄2 + σξ¯i (t )2 − ξt +i 2
2 i=1
2 satisfies τ = min{τ1 , τ2 , L − 1} ≥ 2.
 τ −1 Merging the data-driven MPC and self-triggering schemes in
γ γ
≤− − 2σ ξt 2 + 2σηn̄2 + ξ¯e (t + i)2 Lemma 3, our data-driven self-triggered controller is presented in
2 2 i=1 Algorithm 1, with stability guarantees provided in the following.
Theorem 1 (ISS): Let Assumptions 1–6 hold. Consider system (1)
≤ −(γ/2 − 2σ)ξt 2 + α1 (n̄)
whose control input is generated by solving Problem (8) at each self-

where α1 (n̄) := n̄[JL∗ (t )/λh + L−1 η+i
)(η + JL∗ (t )/λh )] + triggered time t . If the sequence {t }∈N0 satisfies the conditions in
i=1 (ρ
2
2σηn̄ is a K∞ function, and σ̄ ≤ γ/4. Moreover, according to [14, Lemma 3, then system (1) achieves ISS.
Lemma 1], the Lyapunov function satisfies λPξ ξt 2 ≤ V (ξt ) ≤ Proof: Recall from Lemma 3 that V (ξt+1 ) ≤ γ2 V (ξt ) + α3 (n̄)
γ1 ξt 2 + α2 (n̄) for every  ∈ N0 and some constant γ1 > 0 with holds for any t . Recursively,  it can be deduced that
function α2 (·) ∈ K∞ . Hence, we have that V (ξt ) ≤ γ2 V (ξt0 ) + α3 (n̄) −1 i=0 2γ i
. In addition, note that
  λPξ ξt 2 ≤ V (ξt ) ≤ γ1 ξt 2 + α2 (n̄), we arrive at ξt 2 ≤
γ − 4σ γ − 4σ 
V (ξt+1 ) ≤ 1 − V (ξt ) + α2 (n̄) + α1 (n̄) (γ1 /λPξ )γ2 ξt0 2 + (α2 (n̄)γ2 + α3 (n̄) −1 i
i=0 γ2 ). Combining with
2γ1 2γ1
      the fact that xt−η = Υξt , the ISS is established [41, Def. 2.2].1 
=γ2 <1 =α3 (n̄)∈K∞ Remark 6 (Effect of noise on intertriggering time): It is true that
due to the network-induced noise nt , the sensor side receives ts+1 :=
where α3 (n̄) → 0 as n̄ → 0. This implies that there exists a value min{t ∈ N|t+1 + nt +1 } at t , which may different from the time
n̄0 > 0 such that for all n̄ ∈ [0, n̄0 ) the Lyapunov function decreases t+1 generated by Lemma 3 when n̄ ≥ 1. However, it can be observed
at every self-triggered time. Since Problem (8) is feasible at t0 , it is
feasible at t for all  ∈ N0 .  1 System (1) is ISS if there exist a KL-function β and a K -function α such

Remark 5 (Minimum intertriggering time): It is worth noting that that for each ξt0 and nt  ≤ n̄, the inequality xt  ≤ β(xt0 , t ) + α(n̄)
for discrete-time systems, if the intertriggering time τ = 1 for all holds true for all  ∈ N0 .

Authorized licensed use limited to: University of Chinese Academy of SciencesCAS. Downloaded on November 16,2023 at 15:26:50 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 11, NOVEMBER 2023 6957

decreases or the noise increases. Fig. 2(c) depicts the triggering time
instants with n̄ = 0.0015, N = 80, and σ = 0.99. Evidently, the sys-
tem converges to the setpoint with 37 measurement packets transmitted
to the controller. In addition, since η = 2, at most 2 × 37 = 74 outputs
are sent based on the packetized transmission protocol in Section II-A,
which is also much less than 200, and illustrates the effectiveness of the
self-triggering mechanism. Fig. 2(b) and (d) compares the cost function
JL∗ (t ) and the self-triggered times for different σ values.

V. CONCLUSION
This article puts forth a data-driven self-triggering control framework
for unknown linear systems through trajectory prediction. When only
input–output trajectories are available, a data-driven MPC scheme is
developed, which generates a sequence of control inputs once an event
(transmission) is triggered. The predicted output trajectory from the
MPC is further used to design a self-triggering law that is purely
data-based. Both feasibility as well as input-to-state stability were
established for the resulting data-driven self-triggered MPC. Finally,
numerical example is presented to validate the practical merits of the
Fig. 2. (a) Output trajectory, (b) intersampling time t+1 − t with σ =
proposed data-driven control methods.
0.88, (c) comparison of triggering times σ, and (d) comparison of cost
function J.
REFERENCES
[1] K. J. Åström and B. Wittenmark, “On self-tuning regulators,” Automatica,
vol. 9, no. 2, pp. 185–199, Mar. 1973.
from the proof of Lemma 3 that several scaling approximations are
[2] J. Chen, J. Sun, and G. Wang, “From unmanned systems to autonomous
made when deriving self-triggering mechanism, which may sacrifice intelligent systems,” Engineering, vol. 12, pp. 16–19, 2022.
the optimality on τ . In other words, the generated intertriggering [3] S. Adam, L. Busoniu, and R. Babuska, “Experience replay for real-time
time τ is smaller than or equal to the actual maximally tolerable reinforcement learning control,” IEEE Trans. Syst. Man Cybern. Syst.,
intertriggering time τ̄ . This gap can be deemed as a natural mitigation vol. 42, no. 2, pp. 201–212, Mar. 2012.
[4] F. Michel and J. Cédric, “Model-free control,” Int. J. Control, vol. 86,
strategy for the network-induced noise nt . Hence, there exists a constant no. 12, pp. 2228–2252, Jul. 2013.
n̄∗ > 1 such that for all n̄ ∈ (0, n̄∗ ), ISS still holds by transmitting the [5] J. I. Poveda and M. Krstić, “Nonsmooth extremum seeking control with
output at each ts . For n̄ ≥ n̄∗ , one can involve a slack on the left-hand user-prescribed fixed-time convergence,” IEEE Trans. Autom. Control,
side of both inequalities in (14) and (15), which balances between the vol. 66, no. 12, pp. 6156–6163, Dec. 2021.
[6] Z.-S. Hou and Z. Wang, “From model-based control to data-driven con-
system robustness against noise and the transmission frequency. Careful
trol: Survey, classification and perspective,” Inf. Sci., vol. 235, pp. 3–35,
investigation into this issue is left for future. Jun. 2013.
[7] J. C. Willems, I. Markovsky, P. Rapisarda, and B. L. M. De Moor, “A
IV. NUMERICAL EXAMPLES note on persistency of excitation,” Syst. Control Lett., vol. 56, no. 4,
pp. 325–329, May 2005.
Consider a linearized version of the four-tank system in, e.g., [14]. [8] C. De Persis and P. Tesi, “Formulas for data-driven control: Stabilization,
The system matrices are given by optimality, and robustness,” IEEE Trans. Autom. Control, vol. 65, no. 3,
pp. 909–924, Mar. 2020.
⎡ ⎤ ⎡ ⎤ [9] H. J. van Waarde, J. Eising, H. L. Trentelman, and M. K. Camlibel, “Data
0.927 0 0.041 0 0.017 0.001
⎢ 0 informativity: A new perspective on data-driven analysis and control,”
0.918 0 0.033⎥ ⎢
⎥ , B = ⎢0.001 0.023⎥

A=⎢ ⎣ 0
IEEE Trans. Autom. Control, vol. 65, no. 11, pp. 4753–4768, Nov. 2020.
0 0.924 0 ⎦ ⎣ 0 0.061⎦ [10] G. Bianchin, M. Vaquero, J. Cortés, and E. Dall’Anese, “Online stochas-
0 0 0 0.937 0.072 0 tic optimization for unknown linear systems: Data-driven synthesis and
controller analysis,” Aug. 2021, arXiv:2108.13040.
1 0 0 0 [11] C. De Persis and P. Tesi, “Low-complexity learning of linear quadratic reg-
C= , D = 0. ulators from noisy data,” Automatica, vol. 128, Jun. 2021, Art. no. 109548.
0 1 0 0
[12] H. J. van Waarde, M. K. Camlibel, and M. Mesbahi, “From noisy data
with the observability index η = 2. To this end, an input–output trajec- to feedback controllers: Nonconservative design via a matrix S-lemma,”
IEEE Trans. Autom. Control, vol. 67, no. 1, pp. 162–175, Jan. 2022.
tory of length N = 100 was obtained by means of simulating the open- [13] F. Zhao, X. Li, and K. You, “Data-driven control of unknown linear systems
loop system offline using a persistently exciting input sequence. Choose via quantized feedback,” in Proc. 4th Annu. Learn. Dyn. Control Conf.,
n̄ = 0.0015. Parameters of Algorithm 1 were set as follows: the predic- 2022, pp. 467–479.
tion horizon L = 11, cost matrices Q = I, R = 8 · 10−3 I, coefficients [14] J. Berberich, J. Köhler, M. A. Müller, and F. Allgöwer, “Data-driven model
λg n̄ = 10−6 and λh /n̄ = 500, the input constraint set U = [−2, 2]2 , predictive control with stability and robustness guarantees,” IEEE Trans.
Autom. Control, vol. 66, no. 4, pp. 1702–1717, Apr. 2021.
and the input–output equilibrium (ue , y e ) = ([1, 1] , [0.65, 0.77]) . [15] J. Coulson, J. Lygeros, and F. Dörfler, “Data-enabled predictive control: In
Over a simulation horizon of 200 time steps, Fig. 2(a) compares the the shallows of the DeePC,” in Proc. Eur. Control Conf., 2019, pp. 307–312.
output trajectory of the system using the periodic data-driven MPC [16] J. Berberich, J. Köhler, M. A. Müller, and F. Allgöwer, “Robust constraint
in [30], data-driven self-triggering MPC scheme with σ = 0.88 under satisfaction in data-driven MPC,” in Proc. IEEE Conf. Decis. Control,
2020, pp. 1260–1267.
different lengths of data, and noise levels. It can be observed that the [17] W. Liu, J. Sun, G. Wang, F. Bullo, and J. Chen, “Data-driven resilient
self-triggered system achieves a similar performance with the periodic predictive control under denial-of-service,” IEEE Trans. Autom. Control,
one. In addition, the system performance degrades as the datasize early access, Sep. 26, 2022, doi: 10.1109/TAC.2022.3209399.

Authorized licensed use limited to: University of Chinese Academy of SciencesCAS. Downloaded on November 16,2023 at 15:26:50 UTC from IEEE Xplore. Restrictions apply.
6958 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 68, NO. 11, NOVEMBER 2023

[18] X. Wang, J. Sun, F. Deng, G. Wang, and J. Chen, “Event-triggered [29] X. Wang, J. Berberich, J. Sun, G. Wang, F. Allgöwer, and J. Chen, “Model-
consensus control of heterogeneous multi-agent systems: Model- based and data-driven control of event- and self-triggered discrete-time LTI
and data-based approaches,” Sci. China Inf. Sci., pp. 1–13, 2022, systems,” IEEE Trans. Cybern., pp. 1–13, 2023, arXiv:2202.08019.
doi: 10.1007/s11432-022-3683-y. [30] J. Berberich, J. Köhler, M. A. Müller, and F. Allgöwer, “On the design of
[19] Y. Li, X. Wang, J. Sun, G. Wang, and J. Chen, “Data-driven consensus terminal ingredients for data-driven MPC,” IFAC-PapersOnLine, vol. 54,
control of fully distributed event-triggered multi-agent systems,” Sci. pp. 257–263, May 2021.
China Inf. Sci., vol. 66, no. 5, May 2023, Art. no. 152202. [31] J. O’Reilly, Mathematics in Science and Engineering, Observers for
[20] G. Baggio, D. S. Bassett, and F. Pasqualetti, “Data-driven control of Linear Systems. London, U.K.: Acad. Press, 1983.
complex networks,” Nature Commun., vol. 12, Mar. 2021, Art. no. 1429. [32] J. B. Rawlings, D. Q. Mayne, and M. M. Diehl, Model Predictive Con-
[21] Z. Pang, C. Bai, G. Liu, Q. Han, and X. Zhang, “A novel networked predic- trol: Theory and Design, 2nd ed. Santa Barbara, CA, USA: Nob Hill,
tive control method for systems with random communication constraints,” 2019.
J. Syst. Sci. Complex., vol. 34, no. 4, pp. 1364–1378, Aug. 2021. [33] L. Grüne, “NMPC without terminal constraints,” IFAC Proc. Volumes,
[22] M. Mazo and P. Tabuada, “On event-triggered and self-triggered control vol. 45, no. 17, pp. 1–13, 2012.
over sensor/actuator networks,” in Proc. IEEE Conf. Decis. Control, 2008, [34] G. Grimm, M. Messina, S. Tuna, and A. Teel, “Model predictive control:
pp. 435–440. For want of a local control Lyapunov function, All is not lost,” IEEE Trans.
[23] W. P. M. H. Heemels, K. H. Johansson, and P. Tabuada, “An introduction Autom. Control, vol. 50, no. 5, pp. 546–558, May 2005.
to event-triggered and self-triggered control,” in Proc. IEEE Conf. Decis. [35] M. Verhaegen and V. Verdult, Filtering and System Identification: A Least
Control, Maui, HI, USA, 2012, pp. 3270–3285. Squares Approach. Cambridge, U.K.: Cambridge Univ. Press, 2007.
[24] D. Antunes and W. P. M. H. Heemels, “Rollout event-triggered control: Be- [36] G. Wester and M. Lazar, “On the design of distributed DeePC with an
yond periodic control performance,” IEEE Trans. Autom. Control, vol. 59, application to frequency control in power systems,” M.S. thesis, Eindhoven
no. 12, pp. 3296–3311, Dec. 2014. Univ. Technology, Eindhoven, The Netherlands, Aug. 2021.
[25] Z. Sun, L. Dai, K. Liu, D. V. Dimarogonas, and Y. Xia, “Robust self- [37] C. Cai and A. R. Teel, “Input–output-to-state stability for discrete-time
triggered MPC with adaptive prediction horizon for perturbed nonlinear systems,” Automatica, vol. 44, no. 2, pp. 326–336, Sep. 2008.
systems,” IEEE Trans. Autom. Control, vol. 64, no. 11, pp. 4780–4787, [38] A. Eqtami, D. V. Dimarogonas, and K. J. Kyriakopoulos, “Event-triggered
Nov. 2019. control for discrete-time systems,” in Proc. Amer. Control Conf., 2010,
[26] J. Almeida, C. Silvestre, and A. M. Pascoal, “Self-triggered output pp. 4719–4724.
feedback control of linear plants in the presence of unknown distur- [39] K. Zhang, B. Zhou, and G. Duan, “Event-triggered and self-triggered
bances,” IEEE Trans. Autom. Control, vol. 59, no. 11, pp. 3040–3045, control of discrete-time systems with input constraints,” IEEE Trans. Syst.
Nov. 2014. Man Cybern., Syst., vol. 52, no. 3, pp. 1948–1957, Mar. 2022.
[27] G. Delimpaltadakis and M. Mazo, “Region-based self-triggered control [40] D. E. Quevedo, E. I. Silva, and G. C. Goodwin, “Packetized predic-
for perturbed and uncertain nonlinear systems,” IEEE Trans. Control Netw. tive control over erasure channels,” in Proc. Amer. Control Conf., 2007,
Syst., vol. 8, no. 2, pp. 757–768, Jun. 2021. pp. 1003–1008.
[28] M. Wakaiki, “Self-triggered stabilization of discrete-time linear systems [41] Z.-P. Jiang, A. R. Teel, and L. Praly, “Small-gain theorem for ISS systems
with quantized state measurements,” IEEE Trans. Autom. Control, early and applications,” Math. Control Signals Syst., vol. 7, pp. 95–120, 1994.
access, Mar. 15, 2022, doi: 10.1109/TAC.2022.3159262.

Authorized licensed use limited to: University of Chinese Academy of SciencesCAS. Downloaded on November 16,2023 at 15:26:50 UTC from IEEE Xplore. Restrictions apply.

You might also like