Received August 24, 2020, accepted September 11, 2020, date of publication September 22, 2020, date of current

version October 2, 2020.

Digital Object Identifier 10.1109/ACCESS.2020.3025966

Neural Network-Based Distributed Finite-Time

Tracking Control of Uncertain Multi-Agent
Systems With Full State Constraints
QIUTONG JI1 , GANG CHEN 1,2 , (Member, IEEE), AND QIURUI HE1
1 Collegeof Automation, Chongqing University, Chongqing 400044, China
2 Key Laboratory of Dependable Service Computing in Cyber Physical Society, Ministry of Education, Chongqing University, Chongqing 400044, China

Corresponding author: Gang Chen ([email protected])

This work was supported in part by the Natural Science Foundation of Chongqing under Project cstc2020jcyj-msxmX0264, in part by the
Fundamental Research Funds for the Central Universities under Project 2020CDJ-LHZZ-001, and in part by the National Natural Science
Foundation of China under Grant 61673077.

ABSTRACT This paper addresses the distributed tracking control problem of pure-feedback multi-agent
systems with full state constraints under a directed graph in finite time. By introducing the nonlinear
mapping technique, the system with full state constraints is converted into the form without state constraints.
Furthermore, by combining fractional dynamic surface and radial basis function neural networks, a novel
finite-time adaptive tracking controller is conducted recursively. In light of Lyapunov stability theory, it is
proven that all signals of multi-agent systems are semi-globally uniformly ultimately bounded in finite
time and the full states satisfy the constraints. Lastly, numerical simulations are supplied to demonstrate
the effectiveness of the proposed control strategy.

INDEX TERMS Multi-agent systems, finite-time tracking control, full state constraints, pure-feedback form.

I. INTRODUCTION The output-constrained leaderless and leader-follower con-

In recent years, the multi-agent system (MAS) has attracted sensus protocols for second-order nonlinear MASs were pro-
considerable attention since it is widely used in lots of posed in [13], [14] by applying BLFs. The work in [15]
applications such as power grids [1], transportation net- focused on the distributed control of strict-feedback nonlinear
works [2], and wireless sensor networks [3]. In practi- systems with state constraints, in which the unknown nonlin-
cal applications, the constraints often appear owing to the ear dynamics existing in the systems were approximated by
physical limitations or the safe operation of systems [4]. fuzzy logic systems (FLSs). In [16], the BLF was adopted
For example, velocity and position constraints were consid- to deal with consensus problems for heterogenous high-order
ered simultaneously for multi-agent systems in [5]. To deal nonlinear MAS with output constraints.
with constraints, Barrier Lyapunov Functions (BLFs) had For the BLF-based control method, it is inconvenient to
been extensively applied [6]–[11]. In [9], the asymmetri- make new designs of controllers to adapt to changes of Lya-
cal BLFs for strict-feedback nonlinear systems with output punov functions. As a result, a novel nonlinear mapping (NM)
constraints were investigated. By introducing an auxil- methodology was proposed. By introducing the NM method,
iary signal to the asymmetric time-varying BLF, an adap- the output-constrained system in strict-feedback form was
tive tracking controller for time-varying output-constrained converted into a new one without constraints [17]. The design
nonlinear systems with input delay was studied in [10]. process employed traditional Lyapunov functions and thus
A novel universal asymmetric BLF was proposed for the proofs of results were more concise and understandable.
multiple-input multiple-output (MIMO) nonlinear systems However, the uncertain nonlinear functions were assumed to
with output constraints or without constraints in [11]. More- satisfy the linear-in-the-parameters condition in [17]. To han-
over, the BLF was extended to multi-agent systems [12]–[16]. dle the unknown dynamics of systems, adaptive neural net-
works (NNs) or FLSs have been widely applied. In [18]
The associate editor coordinating the review of this manuscript and and [19], the adaptive NN control was developed for approx-
approving it for publication was Guangdeng Zong . imating unknown nonlinear functions and the asymmetric

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 8, 2020 174365
 Q. Ji et al.: NN-Based Distributed Finite-Time Tracking Control of Uncertain MASs With Full State Constraints

NM method was generalized to deal with tracking control systems in backstepping design and achieve the finite-time
problems of full state-constrained strict-feedback systems control, the fractional dynamic surface is introduced and
and pure-feedback systems, respectively. Although control employed. And the adaptive neural networks are used to deal
problems with input constraints, output constraints or full with unknown nonlinearities of the MAS.
state constraints for different categories of nonlinear systems The rest of this paper is organized as follows. Section II
have been extensively studied by employing BLF or NM highlights the main theoretical concepts including some pre-
methods, the tracking control issues of MAS with full state liminaries and problem formulation. The design of adap-
constraints have not been fully investigated. tive controller and the stability analysis are conducted in
It is also worth noting that the design of finite-time dis- section III. Section IV presents some simulation examples.
tributed control laws such that every state reaches consen- The last Section V concludes this paper.
sus in a fast way is very useful for the MAS. Compared
with the asymptotical convergence algorithm, the finite-time II. PRELIMINARIES AND PROBLEM DESCRIPTION
consensus control has significant advantages such as faster A. SOME CONCEPTS OF GRAPH THEORY
convergence rate, robustness against uncertainties as well Consider a MAS including M agents and the corresponding
as better disturbance rejection [20]–[22]. In [22], adaptive communication topology is expressed by G = (V, E), where
finite-time consensus control problems were addressed for V = {v1 , v2 , . . . , vM } is a nonempty set of nodes and E ⊆
the first-order MAS, in which the linearly parameterized V × V is the set of edges. The  set of neighbours of node
method was used to estimate the unknown nonlinear dynam- vj is denoted as Ni = {vj ∈ V (vj , vi ) ∈ E, j 6 = i }, where vi
ics. The works [23] and [24] discussed the finite-time con- can obtain information from node vj . The adjacency matrix
sensus control of second-order and high-order MASs with is denoted as A = [aij ] with aij > 0 when (vj , vi ) ∈ E or else
P the in-degree matrix D = diag{d1 , . . . , dM },
disturbances, respectively. However, the results presented in aij = 0. Define
those papers did not consider full state constraints. Based where di = j∈Ni aij refers to the in-degree of node vi . The
on BLFs and the finite-time stability theory, the finite-time Laplacian matrix can be described as L = D−A. The graph Ḡ
tracking controller for strict-feedback nonlinear systems with is applied to represent the topology among the M + 1 agents
full state constraints and dead-zone was investigated in [25]. with a leader node v0 . The Laplacian matrix L̄ with regard to
Compared with [25], the unknown nonlinear dynamics were Ḡ is expressed as
considered in [26], where an adaptive finite-time fuzzy con-  
L + B −b
troller was designed to solve the full state-constrained track- L̄ = ,
ing control problem of strict-feedback nonlinear systems by 0 0
introducing the tan-type BLF into the backstepping proce- where b = (b1 , . . . , bM )T and B = diag (b1 , . . . , bM ) with
dure. Nevertheless, it is noted that the finite-time tracking bi = 1 if the leader directly connects with the ith follower,
control of the uncertain multi-agent system with full state and bi = 0 otherwise.
constraints is rarely considered.
Inspired by the above articles, the distributed finite-time B. PROBLEM DESCRIPTION
tracking control of the pure-feedback MAS with full state Consider a MAS comprising the leader v0 and M followers
constraints under a directed graph is investigated in this paper. with full state constraints. The dynamics of M followers is
Based on the finite-time stability theory, it is proven that all modeled by
signals of multi-agent systems are semi-globally uniformly 
ultimately bounded in finite time and the full states satisfy ẋi,m = fi,m (x̄i,m , xi,m+1 ), m = 1, . . . , n − 1

the constraints. The main contributions are summarized as ẋi,n = fi,n (x̄i,n ) + gi,n (x̄i,n )ui (1)
follows: 

yi = xi,1 , i = 1, . . . , M ,
1) Compared with the works [22], [24]–[26], the pure-
feedback MAS with full state constraints is considered. Fur- where x̄i,m = [xi,1 , . . . , xi,m ]T ∈ Rm (m = 1, . . . , n). xi,m ∈
thermore, the NM method is extended to the more general R, ui ∈ R, yi ∈ R are system states, control input, and
multi-agent systems. It is challenging and difficult to analyze output of the ith follower, respectively. fi,m (.), m= 1, . . . , n
the stability of the uncertain multi-agent systems. By apply- and gi,n (x̄i,n ) are unknown smooth functions. All states xi,m
ing the NM technique, the full state-constrained MAS in are expected to always remain in an open predefined set
pure-feedback form is firstly converted into a novel one xi,m = {xi,m : −bm1 < xi,m < bm2 }, where bm1 and bm2
without constraints. Moreover, the proposed control design are known positive constants.
process is very simple because it can employ traditional An objective of this paper is to design a distributed adaptive
Lyapunov functions rather than redesigning the Lyapunov consensus protocol for MAS (1) such that the outputs of
functions. followers yi follow the specified desired trajectory yr in finite
2) By combining backstepping and fractional dynamic sur- time, and all states always remain in the constraints xi,m ∈
face control, a novel distributed finite-time tracking strategy xi,m , m = 1, . . . , n.
is proposed, which is different from the works [17]–[19]. Assumption 1: The directed communication topology Ḡ
To avoid the explosion of complexity for the pure-feedback has a directed spanning tree and the leader is the root node.

174366 VOLUME 8, 2020

Q. Ji et al.: NN-Based Distributed Finite-Time Tracking Control of Uncertain MASs With Full State Constraints

Assumption 2: The expected trajectory yr is continuous where

with |yr | < B1 < min {bm1 , bm2 }, where B1 is a known Fi,m (z̄i,m , zi,m+1 ) = qi,m (zi,m )fi,m (x̄i,m , xi,m+1 )−zi,m+1 (7)
positive constant. In addition, the derivative |ẏr | satisfies
|ẏr | ≤ r < ∞ with r being an unknown positive constant. Fi,n (z̄i,n ) = qi,n (zi,n )fi,n (x̄i,n ) (8)
Assumption 3: The function gi,n (x̄i,n ) is bounded, and Gi,n (z̄i,n ) = gi,n (x̄i,n ). (9)
there exists constant gi,n0 satisfying 0 < gi,n0 ≤ gi,n (x̄i,n ).
 
Remark 2: From (9) and Assumption 3, Gi,n (z̄i,n ) ≥
Without loss of generality, we assume that 0 < gi,n0 ≤
gi,n0 > 0 can be easily obtained. Therefore, the control gain
gi,n (x̄i,n ) in this paper.
sign is positive.
Remark 1: Assumption 1 is generally made in the case of
Remark 3: After the system transformation, the specific
leader-following tracking. Assumption 2 is the description of
object of this study is changed to solve distributed tracking
the expected trajectory, which is widely applied in the design
problems of the transformed MAS (6) so that the output yzi
of controllers. These assumptions are frequently used in the
of the followers will track the reference trajectory yd =
existing relevant literature.
[log(b11 + yr )/(b12 − yr )]. Besides, from (3) we know that
the system states xi,m , m = 1, . . . , n will keep in the bounded
C. PRELIMINARIES
 d̄ set i,m if the boundedness of zi,m is guaranteed.
Lemma 1 [27]: For ā, b̄ ∈ R, if c̄, d̄ > 0, then |ā|c̄ b̄ ≤

 c̄+d̄ In this paper, the radial basis function neural net-
c̄/ c̄ + d̄ |ā|c̄+d̄ + d̄/ c̄ + d̄ b̄ .


works (RBFNNs) are employed to estimate unknown contin-
Lemma 2 [27]: For ā, b̄ ∈ R, if 0 < h = h1 /h2 ≤ 1, uous functions ψi,m (m = 1, . . . , n) as follows
h
then āh − b̄h  ≤ 21−h ā − b̄ , where h1 , h2 are positive odd

∗T
ψi,m (ξi,m ) = Wi,m φi,m ξi,m + δi,m ξi,m ,



integers. (10)
LemmaPn 3 [28]:
h ForPnāi ∈ R,h i = 1, 2, . . . , n, 0 < h ≤ 1, where m = 1, . . . , n, φi,m (ξi,m ) denotes the basis function
then |āi | ≤ i=1 |āi | . ∗
i=1 vector; ξi,m denotes the input of neural network; Wi,m is the
Lemma 4 [29]: For ā, b̄ ∈ R, and 0
< h = h1 /h2 ≤
ideal unknown weight vector; δi,m (ξi,m ) is the approximation
1, then āh b̄ − ā ≤ 1+h 1
b̄1+h − ā1+h , where h1 , h2 are


error such that |δi,m (ξi,m )| ≤ ε̄i,m . Meanwhile, let χi,m =
positive odd integers. ∗ 2
Lemma 5 [30]: Consider the system ẋ = f (x, u). Sup- i,0 Wi,m , χ̃i,m = χi,m − χ̂i,m , where χ̂i,m is the estimation
g−1
pose that there exist continuous function V (x), scalars λ > 0, of χi,m and χ̃i,m denotes the estimation error with m =
0 < α < 1, 0 < η < ∞, 0 < θ < 1, such that 1, . . . , n, gi,0 = min{1, gi,n0 }.
V̇ ≤ −λV α (x) + η, then for the given initial state x (0),
the trajectories of the closed-loop system converge into a B. CONTROLLER DESIGN
bounded set  = {x |V α (x) ≤ η/ (1 − θ ) λ } in the finite Based on the transformed system (6), the fraction dynamic
settling time satisfying T ∗ ≤ V 1−α (x (0)) /λθ (1 − α) .


surface control and the backstepping design technique
are employed to design distributed finite-time tracking
III. CONTROLLER DESIGN AND STABILITY ANALYSIS controllers.
A. SYSTEM TRANSFORMATION Firstly, local error surface ei,1 , ei,m and boundary layer
To deal with full state constraints, the NM technique [17] is errors Si,m are respectively defined as
introduced. In this case, the considered dynamics (1) with
M
full state constraints can be converted into the pure-feedback X
ei,1 = aij (zi,1 − zj,1 ) + bi (zi,1 − yd ) (11)
nonlinear system without state constraints. Define
j=1
bm1 + xi,m
M bm1 , bm2 , xi,m = log , m = 1, . . . , n, (2) ei,m = zi,m − αi,mf (12)


bm2 − xi,m
Si,m = αi,mf − αi,m , m = 2, . . . , n, (13)
where bm1 and bm2
are known positive constants. Let zi,m =
M bm1 , bm2 , xi,m . From (2), it is easy to get that where αi,mf refers to the output of the fraction filters as
bm2 ezi,m− bm1 follows.
xi,m = z
. (3)
l
e +1
i,m
i,m α̇i,mf = αi,m − αi,mf , αi,mf (0) = αi,m (0), (14)
Therefore, we obtain
where 0 < i,m < ∞ is a constant and l = (2a − 1)/(2a + 1)
żi,m = qi,m (zi,m )ẋi,m (4) with a ∈ Z + /{1} denotes the fraction.
ezi,m + e−zi,m + 2 The detailed design processes are as follows.
qi,m (zi,m ) = . (5)
bm1 + bm2 Step 1: In what follows, the backstepping method is used
Then, thedynamics (1) can be rewritten as follows to derive the controller. First of all, the Lyapunov function
candidate is constructed as
żi,m = Fi,m (z̄i,m , zi,m+1 ) + zi,m+1

M
żi,n = Fi,n (z̄i,n ) + qi,n (zi,n )Gi,n (z̄i,n )ui (6) X 1 1
V1 = e2 . (15)
yzi = zi,1 , m = 1, . . . , n − 1, 2 bi + di i,1


i=1

VOLUME 8, 2020 174367

 Q. Ji et al.: NN-Based Distributed Finite-Time Tracking Control of Uncertain MASs With Full State Constraints

∗T
According to (10), let ψi,m (ξi,m ) = Wi,m φi,m ξi,m +


The time derivative of V1 can be obtained as below
δi,m ξi,m = Fi,m − α̇i,mf . Taking the time derivative of Vm
M


X 1
V̇1 = ei,1 ėi,1 by utilizing (6), (12), (13), we get
bi + di
i=1
M  M
ei,1
X X
= (bi + di )(Fi,1 + zi,2 ) V̇m = ei,m ėi,m
bi + di i=1
i=1
XM  M
X
ai,j (Fj,1 + zj,2 ) − bi ẏd . ei,m Fi,m + zi,m+1 − α̇i,mf


− (16) =
j=1 i=1

M
P  M 
Let ψi,1 (ξi,1 ) = Fi,1 − 1
X
bi +di ai,j (Fj,1 + zj,2 ) + bi ẏd . = ∗T
ei,m Wi,m φi,m (ξi,m ) + δi,m (ξi,m ) + ei, m+1
j=1
i=1
Then, substituting (12) and (13) into (16) and employing 
RBFNNs in the general form of (10), we have +Si,m+1 + αi,m+1 . (25)
M
X
∗T
φi,1 (ξi,1 )+δi,1 (ξi,1 )+ei,2 +αi,2 +Si,2 .


V̇1 = ei,1 Wi,1 ∗ 2

Let χi,m = g−1
i,0 Wi,m . Then, utilizing Young’s inequality
i=1
(17) once more, we get

where |δi,1 (ξi,1 )| ≤ ε̄i,1 . By employing Young’s inequality, gi,0 χi,m T ηi,m
2
∗T
ei,m Wi,m φi,m ≤ 2
φ φi,m e 2
+ (26)
we obtain 2ηi,m i,m i,m
2
gi,0 χi,1 T ηi,1
2
∗T
φi,1 ≤ φ φ 2 l
ei, 1 Wi,1 e + (18)

i,1 i,1 i,1 l
1+l ki,m1
2ηi,12 2 1 1+l
ei,m δi,m ≤ ε̄i,m
1+l
+ e l (27)
ki,m1 i,m
l
1+l

l
1+l ki,11 1 1+l
ei, 1 δi,1 ≤ ε̄i,1
1+l
e l
l
+ (19)

l
1+l ki,11 i,1 1+l ki,m2 1 1+l
l ei,m ei,m+1 ≤ e1+l + e l (28)
ki,m2 i,m
 i,m+1
h
ki,12
1+l
1+l 1 1+l
ei, 1 ei, 2 ≤ e1+l
i,2 + e l (20)  l
1+l ki,12 i,1 l
1+l ki,m3 1 1+l

l
l ei,m Si,m+1 ≤ 1+l
Si,m+1 + e l , (29)
1+l ki,13 1 1+l 1+l ki,m3 i,m
ei, 1 Si,2 ≤ 1+l
Si,2 + e l , (21)
1+l ki,13 i,1 where ηi,m , ki,m1 , ki,m2 , and ki,m3 are positive design con-
∗ 2
stants.
where χi,1 = g−1 i,0 W i,1 , ηi,1 and ki,11 , ki,12 , ki,13 are The virtual control law αi,m+1 is designed as
positive constants defined by the designer. Now, we construct
the virtual control αi,2 which is described as
1 χ̂i,m T
αi,m+1 = −ki,m0 eli,m − ki,m ei,m
l
− 2
φi,m φi,m ei, m , (30)
2ηi,m
χ̂i,1 T
1
αi,2 = −ki,10 eli,1 − ki,1 ei,1 − 2 φi,1
l
φi,1 ei, 1 , (22)
2ηi,1 where ki,m0 > 0, ki,m > 0 are design parameters with ki,m =
1 1 1
where ki,10 > 0, ki,1 > 0 are design parameters with ki,1 = ki,m1 + ki,m2 + ki,m3 .
1 1 1 Substituting (30) as well as Young’s inequalities (26) - (29)
ki,11 + ki,12 + ki,13 . into (25), it then follows that
Substituting (18) - (22) into (17), we obtain
M 
M 
gi,0 χ̃i,1 T
X gi,0 χ̃i,m T
V̇m ≤ − ki,m0 el+1
i,m + φi,m φi,m e2i,m + ci,m
X
V̇1 ≤ − ki,10 e1+l
i,1 + 2
φi,1 φi,1 e2i, + ci,1 2
2ηi,m
2η 1
i,1 i=1
i=1
h l
l l ( l ki,m2 ) 1+l ( l ki,m3 ) 1+l

( l ki,12 ) 1+l ( 1+l l
ki,13 ) 1+l

+ 1+l ei,2 + Si,2 (23) + 1+l ei,m+1 + 1+l Si,m+1 (31)
1+l 1+l 1+l 1+l
l
η2 h l η2 ( l
ki,m1 )
(
with ci,1 = 2i,1 + 1+l1+l ε̄i,1
ki,11 )
1+l
. with ci,m = i,m
2 +
1+l
1+l ε̄i,m
1+l
.
Step m(2 ≤ m ≤ n − 1): The Lyapunov function Vm for Step n: In this step, actual control laws ui will be designed.
m = 2, . . . , n − 1 is chosen as Considering a Lyapunov function as
M
X e2i,m M
e2i,n
.
X
Vm = (24) Vn = . (32)
2 2
i=1 i=1

174368 VOLUME 8, 2020

Q. Ji et al.: NN-Based Distributed Finite-Time Tracking Control of Uncertain MASs With Full State Constraints

From (12), we have ei,n = zi,n − αi,nf . According to (10), According to (30), it is obvious that α̇i,m+1 is the function of
∗T
let ψi,n (ξi,n ) = Wi,n φi,n ξi,n + δi,n ξi,n = Fi,n − α̇i,nf . variables ei,1 , . . . , ei,m , χ̂i,1 , . . . , χ̂i,m . Thus, from (40),




l
Si,m+1
Then, the derivative of V̇n is  
Ṡi,m+1 +  ≤ Ji,m+1 (ei,1 , . . . , ei,m , χ̂i,1 , . . . , χ̂i,m ),
 
M  i,m+1 
X
V̇n = ei,n ėi,n (41)
i=1
where Ji,m+1 (.) is a continuous function, m = 1, . . . , n − 1.
M
X Construct the overall Lyapunov function candidate as
ei,n Fi,n + qi,n Gi,n ui − α̇i,nf


=
i=1
n
X M X
X n
gi,0 χ̃i,m
2 M X
X n−1 2
Si,m+1
V = Vm + + . (42)
M 2γi,m 2
X m=1 i=1 m=1 i=1 m=1
∗T
φi,n (ξi,n ) + δi,n (ξi,n ) + qi,n Gi,n ui . (33)


= ei,n Wi,n
Applying (23), (31), (37), (38), and (41), the time derivative
i=1
of V can be obtained as
∗ 2

n M X n
Let χi,n = g−1
i,0 Wi,n . By Young’s inequality, we get
X X gi,0 χ̃i,m
V̇ = V̇m + (−χ̂˙ i,m )
γi,m
gi,0 χi,n T ηi,n
2 m=1 i=1 m=1
∗T
ei,n Wi,n φi,n ≤ 2
φ i,n φi,n e 2
i,n + (34) n−1
M X
2ηi,n 2 X
 l + (Si,m+1 Ṡi,m+1 )
ki,n1 l i=1 m=1
(1+l)gi,n0 1+l gi,n0 l
1+l
ei,n δi,n ≤ ε̄i,n + ei,n , (35)
 
M X
n−1 l l
1+l ki,n1 X ( 1+l ki,m2 )
≤ −(ki,m0 − )e1+l
i,m

where gi,n0 , ηi,n , and ki,n1 are positive constants. 1+l
i=1 m=1
Construct the actual controller ui as follows  
M l l
! ( 1+l ki,n2 )
χ̂i,n T
X
1 )el+1
1
ui = −ki,n0 eli,n − ki,n ei,n
l
− φi,n φi,n ei, n , (36) + −(ki,n0 gi,n0 −
i,n

qi,n 2 1+l
2ηi,n i=1

M X
n
where ki,n0 > 0, ki,n = ki,n1 1
and ηi,n > 0 are design X gi,0 σi,m l
+ ( χ̂ χ̃i,m + ci,m )
parameters. Substituting (34) - (36) into (33) yields γi,m i,m
i=1 m=1
M
!
l+1 gi,0 χ̃i,n T
n−1
M X l+1
X Si,m+1
φi,n φi,n ei,n + ci,n ,
2 X
V̇n ≤ −ki,n0 gi,n0 ei,n +
 
2 + (− + Si,m+1  Ji,m+1 )
i=1
2ηi,n i,m+1
i=1 m=1
(37) l
n−1  l
M X
( 1+l ki,m3 )
X 
1+l
.
 lki,n1
l
η2 (1+l)gi,n0
+ Si,m+1 (43)
where ci,n = 2i,n + ε̄i,n
1+l
. 1+l
1+l i=1 m=1
The updating laws of the unknown parameter χi,m for i = According to Lemmas 1, 2, and 4, it can be shown that
1, . . . , M are designed as 1  
χ̂i,m
l
χ̃i,m ≤ − 2l−1
− 2 (l−1)(l+1)
χ̃i,m
1+l
γi,m T (1 + l)2
χ̂˙ i,m = −σi,m χ̂i,m
l
+ 2 φi,m φi,m e2i, , m = 1, . . . , n (38)
2ηi,m m
2(l−1)
 
1 l
+ 1−2 l−1
+ + χi,m
1+l
(44)
with χ̂i,m (0) > 0, σi,m and γi,m are positive design 1+l 1+l 1+l
parameters. From (41), all the variables in the function Ji,m+1 (.) are
bounded, and thus we assume thatthere exists an upper bound
C. STABILITY ANALYSIS on Ji,m+1 (.) such that Ji,m+1 (·) ≤ Ni,m+1 . By Young’s
From the boundary layer errors Si,m (13) and the definition of inequality, we have that
fraction filter for αi,(m+1)f (14), we have 1+l
l
lNi,m+1
Si,m+1  Ji,m+1 ≤ λi S 1+l + .
 

l l (45)
αi,m+1 − αi,(m+1)f Si,m+1 i,m+1 1
α̇i,(m+1)f = =− . (39) (1 + l) [(1 + l) λi ] l
i,m+1 i,m+1 l l
( ki,m3 )
1
By choosing i,m+1 = λi + 1+l1+l +i,m+1 ∗ (m = 1, . . . , n−
Furthermore,
1) and substituting (44) and (45) into (43), it follows that
Ṡi,m+1 = α̇i,(m+1)f − α̇i,m+1 M Xn−1

l

l
l X ( 1+l ki,m2 ) 1+l
Si,m+1 V̇ ≤ −(ki,m0 − )ei,m 
=− − α̇i,m+1 . (40) 1+l
i,m+1 i=1 m=1

VOLUME 8, 2020 174369

 Q. Ji et al.: NN-Based Distributed Finite-Time Tracking Control of Uncertain MASs With Full State Constraints

 
M l l
X ( 1+l ki,n2 ) where
+ −(ki,n0 gi,n0 − )el+1
i,n
  
1+l 1 1 gi,0
i=1 kv = max , , . (53)
M X
n 
2(bi + di ) 2 2γi,m
gi,0 σi,m
X 
+ − (2l−1 − 2(l−1)(l+1) )χ̃i,m
1+l 1+l
γi,m (1 + l)2 Let c̃ = kd /kv 2
. From (50) and (52), it follows that
i=1 m=1
1+l
XM Xn−1   V̇ (t) ≤ −c̃V 2 (t) + C. (54)
∗ 1+l
+ −i,m+1 Si,m+1 + C, (46)
i=1 m=1 In the light of Lemma 5, T ∗ satisfies
1+l
where V (t0 )1− 2
1+l T∗ ≤ (55)
M X
X n M X
X n−1
lNi,m+1l (1 − η1 )c̃(1 − 1+l
2 )
C= ci,m +
(1 + l) [(1 + l) λi ] l
1
with 0 < η1 < 1. Then, we have
i=1 m=1 i=1 m=1
2(l−1)  2
M X n
gi,0 σi,m (1 − 2l−1 +
" l
# 
+ 1+l )
C 1+l
= ς.
X 1+l
+ χi,m
1+l
. (47) V (t) ≤ (56)
γi,m (1 + l) η1 c̃
i=1 m=1
T
Let E1 = e1,1 , . . . , eM ,1 . Then, steady-state errors of M

Let

l l l l
 agents are derived as follows.
 ( 1+l ki,m2 ) ( 1+l ki,n2 ) 
, ki,n0 gi,n0 −
q q
k0 = min ki,m0 − kE1 k ≤ 2 max{bi + di }V (t) ≤ 2 max{bi + di }ς (57)
 1+l 1+l  i i
T T
gi,0 σi,m (2l−1 − 2(l−1)(l+1) ) Let Em = e1,m , . . . , eM ,m , Sm = S1,m , . . . , SM ,m , and
 
k1 = T
γi,m (1 + l)2 χ̃m = χ̃1,m , . . . , χ̃M ,m . Then, we have


k2 = i,m+1 .
∗
(48) p
kEm k ≤ 2V (t) ≤ 2ς , m = 2, . . . , n
p
Then, (46) can be represented as p
kSm k ≤ 2V (t) ≤ 2ς , m = 2, . . . , n
p
M X
n M X
n γmax γmax
r r
kχ̃m k ≤ 2 ς , m = 1, . . . , n, (58)
X X
V̇ ≤ −k0 e1+l χ̃i,m
1+l V (t) ≤ 2
i,m − k1 gmin gmin
i=1 m=1 i=1 m=1
where γmax = max γi,m , gmin = min{gi,0 }. Under the pro-

XM Xn−1
1+l posed finite-time control schemes (22), (30), (36), and (38),
−k2 Si,m+1 + C. (49)
i=1 m=1 the convergent region 1 of tracking errors is described by
Let kd = min{k0 , k1 , k2 }, and further we get the following 1 = {kE1 k ≤ 2 max{bi + di }ς ,
q
inequality i
kEm k ≤ 2ς , kSm k ≤ 2ς , m = 2, . . . , n}.
p p
M X
X n M Xn−1 (59)
X
V̇ ≤ −kd e1+l
i,m + 1+l
Si,m+1 Simultaneously, parameter estimation errors converge to the
i=1 m=1 i=1 m=1 neighborhood 2 given by
M Xn 
γmax
X r
+ χ̃i,m
1+l
+ C. (50) 2 = {kχ̃m k ≤ 2 ς , m = 1, . . . , n} (60)
i=1 m=1 gmin
We further prove that there exist a finite time T ∗ > 0 and a in the finite time T ∗ .
bounded constant 0 < ς < ∞ such that V (t) < ς when It should be noted that there may be a singularity of α̇i,m+1
t ≥ T ∗ . Note that in (41) due to the existence of eli,m (0 < l < 1) in the control
XM X n M X
X n−1 XM X n
! schemes. To avoid the singularity problem, l is redefined as
V (t) ≤ kv 2
ei,m + 2
Si,m+1 + χ̃i,m .
2 follows:
(2a − 1)

i=1 m=1 i=1 m=1 i=1 m=1

 , ei,m ∈/ 0
(51) l ei,m = (2a + 1) (61)
1, otherwise

Using Lemma 3, we get that
where 0 = ei,m ∈ 1 ei,m  < ε, ε > 0 with ε being a
  
1+l  M n M X
n−1
1+l 2 XX X
V 2 (t) ≤ kv e1+l
i,m + 1+l
Si,m+1 small constant.
/ 0 , l ei,m = (2a − 1) (2a + 1) .


i=1 m=1 i=1 m=1 Remark 4: If ei,m ∈
M X
X n  Based on the fractional dynamic surface control, the proposed
+ χ̃i,m
1+l
, (52) finite-time tracking strategy works, and
ei,m converges to the
i=1 m=1 neighborhood 1 . Otherwise, l ei,m = 1. That is to say,

174370 VOLUME 8, 2020

Q. Ji et al.: NN-Based Distributed Finite-Time Tracking Control of Uncertain MASs With Full State Constraints

ei,m is
practically bounded in 1 . From (61), it is obvious

l ei,m is not always an integer. And thus, the control scheme
is essentially a fraction feedback control.
Based on the above analysis, from (56), (57), and (58), they
show that V (t), ei,m , χ̃m (m = 1, . . . , n), Si,m (m = 2, . . . , n)
are bounded. Furthermore, αi,mf , αi,m (m = 2, . . . , n) are
also uniformly ultimately bounded. From (11), we have
E1 = (L + B) (z1 − yd ). Then, it is obvious that kz1 k ≤ FIGURE 1. Communication topology between leader and followers.
kE1 k (λmin (L + B)) + kyd k is bounded. Based on (12)
and (13), zi,m (m = 2, . . . , n) are also bounded. From (3), all
system states are remained in the constrains xi,m ∈ xi,m , m = γi,1 = 2, γi,2 = 2, σi,1 = 0.5, σi,2 = 0.1, i,2 =
1, . . . , n. In addition, the error sets 1 and 2 can converge 0.001, b11 = 1.1, b12 = 1.8, b21 = b21 = 1.8. In addi-
into arbitrary small regions of the origin by selecting suitable tion, the Gaussian
 functions of the RBFNNs
 
are given by
 2 2
design parameters. −
Ps
d=1 zi,md −cki,md /2 vki,md
Overall, the following conclusion can be drawn from the 8i,mk (·) = e , where zi,md ∈
above analysis. Rs is the input vector of the RBFNNs, ci,md and vi,md denote
Theorem 1: Consider the multi-agent system (1) under the center and the width of the RBFNNs respectively. The
Assumptions 1-3, with virtual control laws (22), (30), actual node numbers for RBFNNs are selected as 10, the centers
control laws (36) and adaptive laws given by (38). If the space
√ evenly in the intervals [−1, 1], and the width is equal
initial conditions satisfy xi,m (0) ∈ xi,m = {xi,l : −bm1 < to 2.
xi,m < bm2 }, all system states xi,m ∈ xi,m , i.e., the full
state constraints are not violated. By selecting suitable design
parameters, the local tracking errors converge into arbitrary
small regions of the origin in the finite time T ∗ given in (55).
Remark 5: By introducing the asymmetric NM technique,
the problems of MASs with full state constraints are more
easily addressed as compared with the BLF-based control
method because the NM-based control transforms the con-
strained system into a new system without state constraints
and then the design process can employ traditional Lyapunov
functions. In comparison, it is needed to redesign the Lya-
punov function if the Barrier Lyapunov Functions are applied.
In the work [16], the communication topology is undi-
rected. However, the communication topology of this paper
is directed which is different from [16]. On the other hand, FIGURE 2. The tracking performance with full state constraints.
the fraction dynamic surface based on RBFNNs is employed
to design the finite-time tracking strategy of the uncertain
multi-agent system. Compared with the nonfinite-time algo-
rithm, the proposed control laws have many advantages
including higher accuracy and faster convergence rate.

IV. SIMULATION RESULTS

In this section, numerical simulations are conducted to show
the effectiveness of the proposed control.
Example 1: Consider a group of multi-agent systems (four
followers and one leader) with the following dynamics.


ẋi,1 = xi,2 + 0.05 sin xi,2
2
 
ẋi,2 = 0.9 + 0.05e−xi,1 ui + 0.1 sin xi,1


(62)

The communication topology is shown in Fig. 1, where the FIGURE 3. The trajectory of system states xi ,2 with constraints.

leader node is labeled by 0. The desired trajectory yr is given

as yr = sin(0.5t). The initial conditions of the followers are Fig. 2 shows the tracking performance. By applying the
set as xi,1 (0) = [0.5, 1, 0.6, 1.3]T , xi,2 (0) = [0, 0, 0, 0]T . proposed control method, it shows that the outputs of the
The design parameters are selected as s = 8, ki,10 = followers can track the desired trajectory and always remain
10, ki,1 = 5, ki,22 = 6, ki,2 = 6, ηi,1 = 2, ηi,2 = 2, in xi,1 = {xi,1 : −1.1 < xi,1 < 1.8}. Fig. 3 shows

VOLUME 8, 2020 174371

 Q. Ji et al.: NN-Based Distributed Finite-Time Tracking Control of Uncertain MASs With Full State Constraints

FIGURE 4. The local errors of followers.

FIGURE 5. The estimations of the adaptive parameters.

FIGURE 6. Comparisons of xi ,2 with/without constraints.

that the states xi,2 of the followers also remain in xi,2 = xi,2 (0) = [1, 0.5, 1.1, 0.3]T . The design parameters are
{xi,2 : −1.8 < xi,2 < 1.8}. The tracking errors ei,1 and selected as s = 8, ki,10 = 9, ki,1 = 1, ki,22 = 9, ki,2 = 1,
ei,2 are illustrated in Fig. 4. It is clear that the tracking errors ηi,1 = 2, ηi,2 = 2, γi,1 = 2, γi,2 = 2, σi,1 = 0.5, σi,2 = 0.1,
decline fast and remain in a small bound around zero. Fig. 5 is i,2 = 0.001, b11 = 4.8, b12 = 4.3, b21 = b21 = 5.1. The
the estimations of the unknown parameters χi,1 , χi,2 . Under other parameters are the same as Example 1.
the same communication topology and the same dynamics
of the multi-agent system with the same design parameters
mentioned above, the results of the proposed NM-based con-
trol schemes are compared with the conventional control
scheme without the NM method. As Fig. 6 shows, there is
a significant difference between system states xi,2 with or
without constraints. As we see, it is apparent that the proposed
control based on the NM method can guarantee that the
system states always stay in the constraints xi,2 = {xi,2 :
−1.8 < xi,2 < 1.8}.
Example 2: Another example is carried out to further
demonstrate the effectiveness of the proposed algorithm.
In this example, the communication topology is identical to
Example 1. The dynamic model is presented as

ẋi,1 = xi,2 FIGURE 7. The consensus tracking performance of Example 2.

ẋi,2 = − sin(xi,1 ) − 0.25xi,2 +1.5cos(2.5t) + ui (63)

Fig. 7 displays the outputs of all agents. It can be observed
The trajectory is given as yr = 0.5 sin(0.5t) + 0.5 sin(t). And that the desired tracking performance with state constraints
the initial conditions are set as xi,1 (0) = [0.5, 1, 0.6, 1.3]T , is obtained. Fig. 8 shows the constrained states xi,2 of the

174372 VOLUME 8, 2020

Q. Ji et al.: NN-Based Distributed Finite-Time Tracking Control of Uncertain MASs With Full State Constraints

finite-time control enjoys a stronger stability performance

than the nonfinite-time scheme. In addition, the effect of
interference is considered. Fig. 10 demonstrates the robust-
ness properties of the proposed algorithm under the interfer-
ence with the amplitude of 10 and the duration of 2 seconds
at 10 s. It shows that the proposed control schemes ensure the
good robustness of the system. Based on the above simula-
tion results, we get that the good performance of proposed
controllers is achieved and the local tracking errors converge
into the small regions of the origin in the finite time while full
state constraints are not violated.

V. CONCLUSION
FIGURE 8. System states xi ,2 of Example 2. In this paper, the distributed finite-time tracking problem is
discussed for a class of pure-feedback uncertain multi-agent
systems with full state constraints. By applying the NM
method, the full state-constrained MAS is converted into a
novel pure-feedback one without state constraints. Based on
fractional dynamic surface control as well as backstepping,
a distributed adaptive finite-time tracking control strategy is
proposed recursively and the proposed scheme is very simple.
It is proved that all the signals are semi-globally uniformly
ultimately bounded in finite time, and all the state constraints
are always guaranteed. In future studies, we will consider
extending the fractional dynamic surface control to a class
of nonlinear multi-agent systems with input constraints as
well as full state constraints. In addition, further research with
switching topologies will be undertaken.

FIGURE 9. Comparisons of e2,1 and e2,2 between finite-time and

nonfinite-time control.
REFERENCES
[1] M. M. Esfahani, A. Hariri, and O. A. Mohammed, ‘‘A multiagent-
based game-theoretic and optimization approach for market operation
of multimicrogrid systems,’’ IEEE Trans. Ind. Informat., vol. 15, no. 1,
pp. 280–292, Jan. 2019.
[2] R. Luo, T. J. J. van den Boom, and B. De Schutter, ‘‘Multi-agent dynamic
routing of a fleet of cybercars,’’ IEEE Trans. Intell. Transp. Syst., vol. 19,
no. 5, pp. 1340–1352, May 2018.
[3] X. Li, X. Hu, W. Li, and H. Hu, ‘‘A multi-agent reinforcement learning
routing protocol for underwater optical sensor networks,’’ in Proc. IEEE
Int. Conf. Commun. (ICC), Shanghai, China, May 2019, pp. 1–7.
[4] D.-J. Li, S.-M. Lu, Y.-J. Liu, and D.-P. Li, ‘‘Adaptive fuzzy tracking control
based barrier functions of uncertain nonlinear MIMO systems with full-
state constraints and applications to chemical process,’’ IEEE Trans. Fuzzy
Syst., vol. 26, no. 4, pp. 2145–2159, Aug. 2018.
[5] P. Lin, W. Ren, C. Yang, and W. Gui, ‘‘Distributed optimization with
nonconvex velocity constraints, nonuniform position constraints, and
nonuniform stepsizes,’’ IEEE Trans. Autom. Control, vol. 64, no. 6,
pp. 2575–2582, Jun. 2019.
[6] C. Xi and J. Dong, ‘‘Adaptive neural network-based control of uncertain
nonlinear systems with time-varying full-state constraints and input con-
FIGURE 10. The local errors e1,1 and e2,1 with the interference.
straint,’’ Neurocomputing, vol. 357, pp. 108–115, Sep. 2019.
[7] W. He, L. Kong, Y. Dong, Y. Yu, C. Yang, and C. Sun, ‘‘Fuzzy track-
ing control for a class of uncertain MIMO nonlinear systems with state
constraints,’’ IEEE Trans. Syst., Man, Cybern. Syst., vol. 49, no. 3,
followers. From Fig. 7 and Fig. 8, the system states xi,1 and pp. 543–554, Mar. 2019.
xi,2 are always remaining in the open sets xi,1 = {xi,1 : [8] W. He, Z. Yin, and C. Sun, ‘‘Adaptive neural network control of a marine
−4.8 < xi,1 < 4.3} and xi,2 = {xi,2 : −5.1 < xi,2 < 5.1}, vessel with constraints using the asymmetric barrier Lyapunov function,’’
IEEE Trans. Cybern., vol. 47, no. 7, pp. 1641–1651, Jul. 2017.
respectively. Fig. 9 shows the comparisons of the convergence [9] K. P. Tee, S. S. Ge, and E. H. Tay, ‘‘Barrier Lyapunov functions for the
results between the proposed finite-time control and the control of output-constrained nonlinear systems,’’ Automatica, vol. 45,
nonfinite-time control with general dynamic surface control, no. 4, pp. 918–927, Apr. 2009.
[10] W. Deng, J. Yao, and D. Ma, ‘‘Adaptive control of input delayed uncertain
i.e., l = 1 under the same initial conditions and the design nonlinear systems with time-varying output constraints,’’ IEEE Access,
parameters. From Fig. 9, it can be seen that the proposed vol. 5, pp. 15271–15282, 2017.

VOLUME 8, 2020 174373

 Q. Ji et al.: NN-Based Distributed Finite-Time Tracking Control of Uncertain MASs With Full State Constraints

[11] X. Jin, ‘‘Adaptive fixed-time control for MIMO nonlinear systems with [29] Y. Wang, Y. Song, M. Krstic, and C. Wen, ‘‘Fault-tolerant finite time
asymmetric output constraints using universal barrier functions,’’ IEEE consensus for multiple uncertain nonlinear mechanical systems under
Trans. Autom. Control, vol. 64, no. 7, pp. 3046–3053, Jul. 2019. single-way directed communication interactions and actuation failures,’’
[12] Y. Zhu and F. Zhu, ‘‘Barrier-function-based distributed adaptive control Automatica, vol. 63, pp. 374–383, Jan. 2016.
of nonlinear CAVs with parametric uncertainty and full-state constraint,’’ [30] Z. Zhu, Y. Xia, and M. Fu, ‘‘Attitude stabilization of rigid spacecraft with
Transp. Res. C, Emerg. Technol., vol. 104, pp. 249–264, Jul. 2019. finite-time convergence,’’ Int. J. Robust Nonlinear Control, vol. 21, no. 6,
[13] X. Cai, C. Wang, G. Wang, and D. Liang, ‘‘Distributed consensus control pp. 686–702, Apr. 2011.
for second-order nonlinear multi-agent systems with unknown control
directions and position constraints,’’ Neurocomputing, vol. 306, pp. 61–67,
Sep. 2018.
[14] Y. Yu, W. Wang, and K.-H. Jo, ‘‘Adaptive consensus control of output-
constrained second-order nonlinear systems via neurodynamic optimiza-
tion,’’ Neurocomputing, vol. 295, pp. 1–7, Jun. 2018. QIUTONG JI was born in Jiangsu, China, in 1996.
[15] Y. Zhang, H. Liang, H. Ma, Q. Zhou, and Z. Yu, ‘‘Distributed adaptive She is currently pursuing the M.S. degree with
consensus tracking control for nonlinear multi-agent systems with state the School of Automation, Chongqing University,
constraints,’’ Appl. Math. Comput., vol. 326, pp. 16–32, Jun. 2018. Chongqing, China.
[16] D. Shen and J.-X. Xu, ‘‘Distributed learning consensus for heterogenous
Her current research interests include dis-
high-order nonlinear multi-agent systems with output constraints,’’ Auto-
tributed control, cooperative adaptive control, and
matica, vol. 97, pp. 64–72, Nov. 2018.
[17] T. Guo and X. Wu, ‘‘Backstepping control for output-constrained nonlinear control of nonlinear systems.
systems based on nonlinear mapping,’’ Neural Comput. Appl., vol. 25,
nos. 7–8, pp. 1665–1674, Dec. 2014.
[18] T. Zhang, M. Xia, and Y. Yi, ‘‘Adaptive neural dynamic surface control of
strict-feedback nonlinear systems with full state constraints and unmodeled
dynamics,’’ Automatica, vol. 81, pp. 232–239, Jul. 2017.
[19] T. Zhang, M. Xia, Y. Yi, and Q. Shen, ‘‘Adaptive neural dynamic surface
control of pure-feedback nonlinear systems with full state constraints and
dynamic uncertainties,’’ IEEE Trans. Syst., Man, Cybern., Syst., vol. 47, GANG CHEN (Member, IEEE) received the Ph.D.
no. 8, pp. 2378–2387, Aug. 2017. degree in control engineering from Zhejiang Uni-
[20] B. Ning and Q.-L. Han, ‘‘Prescribed finite-time consensus tracking for versity, Hangzhou, China, in 2006. Since 2006,
multiagent systems with nonholonomic chained-form dynamics,’’ IEEE he has been with the College of Automation,
Trans. Autom. Control, vol. 64, no. 4, pp. 1686–1693, Apr. 2019. Chongqing University, Chongqing, China, where
[21] X. Lin and Y. Zheng, ‘‘Finite-time consensus of switched multia- he is currently a Professor. From 2009 to 2010,
gent systems,’’ IEEE Trans. Syst., Man, Cybern. Syst., vol. 47, no. 7, he was a Visiting Scholar with the Automation
pp. 1535–1545, Jul. 2017. and Robotics Research Institute, The University of
[22] Z. Tu, H. Yu, and X. Xia, ‘‘Decentralized finite-time adaptive consensus Texas at Arlington. His research interests include
of multiagent systems with fixed and switching network topologies,’’
distributed control, cooperative control, intelligent
Neurocomputing, vol. 219, pp. 59–67, Jan. 2017.
control, distributed optimization, cyberphysical systems, smart grids, non-
[23] X. Tian, H. Liu, and H. Liu, ‘‘Robust finite-time consensus control
for multi-agent systems with disturbances and unknown velocities,’’ ISA linear control, and control applications.
Trans., vol. 80, pp. 73–80, Sep. 2018.
[24] Y. Wang, Y. Song, M. Krstic, and C. Wen, ‘‘Adaptive finite time coordi-
nated consensus for high-order multi-agent systems: Adjustable fraction
power feedback approach,’’ Inf. Sci., vol. 372, pp. 392–406, Dec. 2016.
[25] H. Li, S. Zhao, W. He, and R. Lu, ‘‘Adaptive finite-time tracking control
QIURUI HE was born in Chongqing, China,
of full state constrained nonlinear systems with dead-zone,’’ Automatica,
vol. 100, pp. 99–107, Feb. 2019. in 1996. He is currently pursuing the M.S. degree
[26] J. Xia, J. Zhang, W. Sun, B. Zhang, and Z. Wang, ‘‘Finite-time adaptive with the School of Automation, Chongqing Uni-
fuzzy control for nonlinear systems with full state constraints,’’ IEEE versity, Chongqing.
Trans. Syst., Man, Cybern., Syst., vol. 49, no. 7, pp. 1541–1548, Jul. 2019. His current research interests include distributed
[27] C. Qian and W. Lin, ‘‘Non-lipschitz continuous stabilizers for nonlinear control, online convex optimization, microgrids,
systems with uncontrollable unstable linearization,’’ Syst. Control Lett., and storage systems.
vol. 42, no. 3, pp. 185–200, Mar. 2001.
[28] G. H. Hardy, J. E. Littlewood, and G. Pólya, Inequalities. Cambridge, U.K.:
Cambridge Univ. Press, 1952.

174374 VOLUME 8, 2020