You are on page 1of 13

Automatica 50 (2014) 452–464

Contents lists available at ScienceDirect

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

Consensus of high-order multi-agent systems with large input and


communication delays✩
Bin Zhou a,1 , Zongli Lin b
a
Center for Control Theory and Guidance Technology, Harbin Institute of Technology, P.O. Box 416, Harbin, 150001, China
b
Charles L. Brown Department of Electrical and Computer Engineering, University of Virginia, P.O. Box 400743-4743, Charlottesville, VA 22904-4743, USA

article info abstract


Article history: We study in this paper the consensus problem for multi-agent systems with agents characterized by high-
Received 13 September 2012 order linear systems with time delays in both the communication network and inputs. Provided that the
Received in revised form open-loop dynamics of the agents is not exponentially unstable, but may be polynomially unstable, and
10 September 2013
the communication topology contains a directed spanning tree, a truncated predictor feedback approach
Accepted 12 November 2013
Available online 17 December 2013
is established to solve the consensus problem. It is shown that, if the delays are constant and exactly
known, the consensus problems can be solved by both full state feedback and observer based output
Keywords:
feedback protocols for arbitrarily large yet bounded delays. If it is further assumed that the open-loop
Consensus dynamics of the agents only contains zero eigenvalues, the delays are allowed to be time-varying and
High-order multi-agent systems unknown. Numerical examples are worked out to illustrate the effectiveness of the proposed approaches.
Time delays © 2013 Elsevier Ltd. All rights reserved.
Truncated predictor feedback

1. Introduction challenging tasks that cannot be accomplished if they operate


uncooperatively. As a result, consensus via cooperative control
Consensus refers to a group of agents under appropriate of a group of agents has received considerable attention from
distributed control policies reaching an agreement on certain various scientific communities in the past several decades (see
quantities of interest (Ishii, Tempo, & Bai, 2012; Lv, Chen, & di Duan, Wang, Chen, & Huang, 2008, Hong, Chen, & Bushnell, 2006,
Bernardo, 2010; Lv, Chen, & Yu, 2011; Ren & Cao, 2011). Consensus Ishii & Tempo, 2010, Li & Jiang, 2009, Li & Xie, 2011, Lv et al., 2011,
is a fundamental problem in cooperative control of multi-agent Qin, Gao, & Zheng, 2011, Ren & Cao, 2011, Su et al., 2009, Su, Wang,
systems and is closely related with other high-level problems such & Chen, 2010, You & Xie, 2011b and the references therein).
as flocking (Su, Wang, & Chen, 2009; Yu, Chen, & Cao, 2010a) While the existing literature deals with consensus problem
and formation control (Fax & Murray, 2004), which find many for multi-agent systems captured by low order dynamics such
applications in engineering such as sensor networks, spacecraft as single integrator, double integrator, and oscillators, consensus
formation flying and cooperative surveillance (Chen & Duan, 2008; for multi-agent systems with high-order dynamics has received
Duan, Chen, & Huang, 2007; Ren, Beard, & Atkins, 2007). Indeed, little attention, except for multi-agent systems whose agents are
a group of autonomous agents connected by a communication or modeled by integrator chains of length greater than two (see,
sensing network can coordinate with each other to perform some for example, Ren, Moore, & Chen, 2006). Recently, the consensus
problem for multi-agent systems with agents modeled by high-
order linear dynamic systems is considered in Seo, Shim, and
✩ This work was supported in part by the National Natural Science Foundation Back (2009) by dynamic output feedback. The consensus problems
of China under grant numbers 61104124, 61273028, 61273105 and 61322305, for both continuous-time and discrete-time linear multi-agent
by the Fundamental Research Funds for the Central Universities under Grants
systems with directed communication topologies are addressed
HIT.BRETIII.201210 and HIT.BRETIV.201305, by the Foundation for Innovative
Research Group of the National Natural Science Foundation of China under Grant in Li, Liu, Lin, and Ren (2011), where distributed reduced-order
61021002, and by the National Science Foundation of the United States under grant observer-based consensus protocols are developed. For single
number CMMI-1129752. The material in this paper was partially presented at the input high-order nonlinear multi-agent systems with unknown
2013 American Control Conference (ACC2013), June 17–19, 2013, Washington, DC, dynamics, an adaptive cooperative tracking control scheme is
USA. This paper was recommended for publication in revised form by Associate
Editor Antonis Papachristodoulou under the direction of Editor Frank Allgöwer.
considered in Zhang and Lewis (2012). Synchronization of identical
E-mail addresses: binzhoulee@163.com, binzhou@hit.edu.cn (B. Zhou), general linear systems on a digraph containing a spanning tree is
zl5y@virginia.edu (Z. Lin). studied in Zhang, Lewis, and Das (2011) by establishing both state
1 Tel.: +86 451 87556211; fax: +86 45186418034. feedback and observer based output feedback protocols. For more
0005-1098/$ – see front matter © 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.automatica.2013.12.006
B. Zhou, Z. Lin / Automatica 50 (2014) 452–464 453

related work on consensus of high-order multi-agent systems, see We emphasize that, for observer based output feedback
Li et al. (2011), Ren et al. (2006), Seo et al. (2009), Su and Huang consensus, as the protocol is distributed, the observer error
(2012b), You and Xie (2011a) and the references cited there. dynamics and the dynamics of the agents are coupled. As a result,
Delay effect on the convergence of consensus protocols is an an intricate Lyapunov analysis of the stability of the closed-loop
important issue to be considered. One source of time delay in multi-agent system has to be performed. We also emphasize that,
multi-agent systems is the communication from one agent to while most of the literature on consensus of multi-agent systems
another, which is named as communication delay. Another source with time delays deals with the analysis of robustness with respect
of time delay is related with the processing and connecting time to delays, that is, to estimate the bounds on the time delays under
for the packets arriving at each agent, which is called input which a pre-designed consensus protocol continues to achieve
delay (Tian & Liu, 2009). Consensus of multi-agent systems with consensus in the presence of time delays that are generally of
communication and/or input delays has been extensively studied small size (see, for example, Lin et al., 2007, Munz et al., 2010,
in the literature (see Lin, Jia, Du, & Yuan, 2007, Olfati-Saber & 2011a, Munz, Papachristodoulou, & Allgower, 2012, Olfati-Saber &
Murray, 2004, Tian & Liu, 2009 and the references therein). In Murray, 2004, Qin et al., 2011, Rudy & Olgac, 2011 and Yu et al.,
most of these studies, the time delays are assumed to be unknown 2010b), in the present paper, we take time delays into account in
and the main purpose is to find the upper bounds on the time our design of consensus protocol and we allow the delays to be
delays such that the consensus can still be achieved in the presence arbitrarily large. Particularly, we adjust a design parameter in the
of time delay. For some simple agent dynamics, for example, protocols in accordance with the size of the delays.
those of single integrators or single oscillators, establishing the The remaining part of this paper is organized as follows. The
necessary and sufficient conditions on the maximal allowable time problem formulation is given in Section 2. The state feedback con-
delay is possible under a prescribed protocol by analyzing the sensus and observer based output feedback consensus problems
roots of certain characteristic equations (see, for example, Lin are respectively solved in Sections 3 and 4. Numerical examples
et al., 2007, Olfati-Saber & Murray, 2004, Rudy & Olgac, 2011 are provided in Section 5 to validate the effectiveness of the pro-
and Yu, Chen, & Cao, 2010b, for details). By using the generalized posed protocols. Finally, Section 6 concludes the paper.
Nyquist criterion, some set-valued conditions are established Notation: The notation used in this paper is fairly standard. For a
in Munz, Papachristodoulou, and Allgower (2010) to guarantee complex matrix A with appropriate dimensions, AT , AH , λ(A), ∥A∥
the consensus of general single-input–single-output linear multi- are respectively its transpose, conjugated transpose, eigenvalue
agent systems with delays and explicit conditions for determining set, and norm. For a positive scalar τ , let Cn,τ = C ([−τ , 0] , Rn )
the convergence rate of single-integrator multi-agent systems are denote the Banach space of continuous vector functions mapping
also proposed there. In Munz, Papachristodoulou, and Allgower the interval [−τ , 0] into Rn with the topology of uniform conver-
(2011a), robust static output-feedback controllers are designed to gence, and let xt ∈ Cn,τ denote the restriction of x(t ) to the inter-
achieve consensus in networks of heterogeneous agents modeled val [t − τ , t] translated to [−τ , 0], that is, xt (θ) = x (t + θ ) , θ ∈
as nonlinear systems of relative degree two in the presence of [−τ , 0]. For two matrices A and B, we use A ⊗ B to denote their
heterogeneous communication delays. Very recently, the high- Kronecker product. For two integers p and q with p ≤ q, the sym-
order consensus problem for heterogeneous multi-agent systems bol I [p, q] refers to the set {p, p + 1, . . . , q}. For a complex num-
with unknown communication delays is studied in Tian and Zhang ber s, we use Re {s} and |s| to denote respectively its real part and
(2012) and a necessary and sufficient condition is given for the model. Finally, throughout this paper, if not specified, for a series
existence of a high-order consensus solution to heterogeneous of vectors xi , i ∈ I [1, N] with appropriate dimensions, we denote
multi-agent systems.
T
x = xT1 xT2 xTN

··· .
This paper studies the consensus problem for multi-agent sys-
tems with agents characterized by high-order linear systems with
time delays. The time delays can be in both the communication 2. Problem formulation and preliminaries
network and the inputs of the agents. Since the delays are allowed
to be arbitrarily large, under the assumption that the communica- We consider a continuous-time high-order multi-agent system
tion topology among the agents contains a directed spanning tree, described by
we first employ the well-known predictor feedback (which is also
ẋi (t ) = Axi (t ) + Bui (t − τcon ) ,

known as the finite spectrum assignment and model reduction ap- (1)
yi (t ) = Cxi (t ), i ∈ I [1, N] ,
proach, Manitius & Olbrot, 1979, Zhou, Lin, & Duan, 2012) to design
both distributed state feedback and observer based output feed- where xi ∈ Rn , ui ∈ Rm and yi ∈ Rp are the state, the control
back protocols such that the delays are completely compensated. and the output of Agent i, respectively, N ≥ 1 is a given integer
However, as the predictor feedback protocols require the exact denoting the number of agents, τcon ≥ 0 is the input delay, and
information of the network and the relative input signals among (A, B, C ) is a given matrix triple. Let the communication topology
the agents, they may suffer some implementation problems. To among these agents be characterized by a weighted directed graph
overcome this problem, under the additional condition that the G (N ,E , A
 ), where N is the node set, E is the edge set, and
open-loop dynamics of the agents is at most polynomially unstable, A = αij ∈ RN ×N is the weighted adjacency matrix. Denote the
we show that the consensus problems can also be solved by trun-  
corresponding Laplacian by L = lij ∈ RN ×N .
cated predictor feedback protocols which need neither the exact
In the full information case, we assume that Agent i collects the
information of the network nor the relative input signals among
delayed state information of its neighboring agents by the rule
the agents. The allowable delays under the truncated predictor 
zi (t ) = αij xi (t − τcom ) − xj (t − τcom )
 
feedback based protocols can also be arbitrarily large, yet bounded.
We also show that, if the open-loop dynamics of the agents only j∈Ni
contains zero eigenvalues, the time delays are allowed to be time-
N
varying and unknown, namely, for an arbitrarily given positive 
= lij xj (t − τcom ) , i ∈ I [1, N] , (2)
number τ ∗ , the established explicit distributed protocols (depen-
j =1
dent on τ ∗ ) achieve consensus for arbitrary time delays bounded
by τ ∗ . Numerical examples are worked out to illustrate the effec- where τcom ≥ 0 represents the communication delay and Ni = {j :
tiveness of the proposed protocols. αij ̸= 0}. In the partial information case, we assume that Agent i
454 B. Zhou, Z. Lin / Automatica 50 (2014) 452–464

collects the delayed output information of its neighboring agents Notice that multi-agent systems with agents characterized by a
by the rule chain of integrators or a single oscillator (see, for example, Lin et al.,
2007 and Yu et al., 2010b) satisfy Assumption 1 automatically.
N
 However, Assumption 1 allows A to have nonzero eigenvalues with
wi ( t ) = lij yj (t − τcom ) , i ∈ I [1, N] . (3) algebraic and geometry multiplicities larger than 1, and is thus
j =1 weaker than the existing assumptions, for example, Assumption
1 in Su and Huang (2012a). We notice that such an assumption is
For future use, we assume that the initial conditions of the multi-
also made in Wang, Saberi, Stoorvogel, Grip, and Yang (2013).
agent system (1) are xi (θ) = (xi )0 ∈ Cn,τ and ui (θ ) = (ui )0 ∈
Our second assumption is concerned with the communication
Cm,τ , θ ∈ [−τ , 0], where τ = τcon + τcom . In this paper, we are
topology among the agents.
interested in the design of the state feedback protocol ui (t ) =
ui (zi ) , i ∈ I [1, N], and observer based output feedback protocol
Assumption 2. The communication topology G (N , E , A) con-
ui (t ) = ui (wi , ωi ) , ω̇i = h (ωi , wi ) , i ∈ I [1, N], that achieve
tains a directed spanning tree.
consensus of the multi-agent system (1). To this end, we first
introduce the concept of consensus. This assumption is necessary for guaranteeing a solution to the
consensus problem even in the absence of delay (see, for example,
Definition 1. The high-order Li et al., 2010 and Seo et al., 2009). Under Assumption 2, there
 linear multi-agent system (1) achie-
ves consensus if limt →∞ xi (t ) − xj (t ) = 0, ∀i, j ∈ I [1, N]. exists a nonsingular matrix T ∈ CN ×N , whose first column is 1N ,

[1, 1, . . . , 1]T ∈ RN , such that (Ren & Cao, 2011)
0
 
Remark 1. By definition, if the high-order multi-agent system (1)
achieves consensus, then there exists a signal s(t ) ∈ Rn , which may  λ2 δ2   
be unbounded, such that limt →∞ ∥xi (t ) − s(t )∥ = 0, ∀i ∈ I [1, N]. −1
 .. .. , 0
 0
T LT =  . . 0 DL
The signal s(t ) is referred to as the reference trajectory.
 
 λ N −1 δ N −1 

The problems we are to solve can then be formally stated as


λN
follows. , JL , (5)

where λi are such that Re {λi } > 0, i ∈ I [2, N], and δi , i ∈


Problem 1 (State Feedback Consensus). Design a state feedback
I [2, N − 1], equals either 1 or 0. For future use, we define δN = 0.
protocol ui (t ) = ui (zi (t )) = Fzi (t ), where zi (t ) is defined in (2)
At the end of this section, we introduce the following lemma
and F ∈ Rm×n is a constant matrix, such that the high-order linear which plays the central role in the stability analysis in the sequel.
multi-agent system (1) with time delays in both the inputs and the The detailed proof is provided in Appendix A.1 for clarity.
communication network achieves consensus.
Lemma 1. Assume that (A, B, C ) satisfies Assumption 1 and λ ∈ C
Problem 2 (Observer Based Output Feedback Consensus). Design a is a given scalar such that Re {λ} > 0. Let F = −µBT P (γ ) where
finite dimensional stable observer based output feedback protocol, P (γ ) = P > 0 is the unique positive definite solution to the following
expressed in the form of parametric algebraic Riccati equation (ARE)

ω̇i (t ) = AF ωi (t ) + BF wi (t ), ωi (0) ∈ Rn ,

(4) AT P + PA − PBBT P = −γ P . (6)
ui (t ) = CF ωi (t ) + DF wi (t ), i ∈ I [1, N] ,
Let H be such that A + HC is Hurwitz, (θ , ϑ) be a pair of constant real
where AF ∈ Rn×n is Hurwitz, such that the linear continuous-time numbers and µ∗ = Re1{λ} .
high-order multi-agent system (1) with time delays in both the
(1) If τ is constant yet can be arbitrarily large, then for any µ ≥ µ∗ ,
inputs and the communication network achieves consensus and
there exists a scalar γ ∗ = γ ∗ (µ, τ , |λ| , H ) > 0 such that the
limt →∞ ∥ωi (t )∥ = 0, ∀i ∈ I [1, N].
linear time delay system
We will impose some assumptions on the linear continuous- ϕ̇(t ) = Aϕ(t ) + λBF ϕ (t − τ ) + θ λBF e (t − τ ) ,

time high-order multi-agent system (1). First, to ensure that the (7)
ė(t ) = (A + HC ) e(t ) − ϑλBF (ϕ (t − τ ) − e (t − τ ))
consensus value reached by the agents will not tend to infin-
ity exponentially, namely, the reference trajectory s(t ) defined in is asymptotically stable for all γ ∈ (0, γ ∗ ], where F = F eAτ .
Remark 1 is not exponentially diverging, the matrix A should not (2) If, in addition, all the eigenvalues of A are zero and the delay τ
contain eigenvalues in the open right-half plane (Li, Duan, Chen, is possibly unknown, time varying yet bounded uniformly by τ∗ ,
& Huang, 2010; Seo et al., 2009). Hence we should assume that then for any µ ≥ µ∗ , there exists a scalar γ ∗ = γ ∗ (µ, τ∗ ,
all the eigenvalues of A are located on the closed left-half plane. |λ| , H ) > 0 such that the linear time-delay system
In this case, we can perform a state transformation such that the
ϕ̇(t ) = Aϕ(t ) + λBF ϕ (t − τ ) + θ λBFe (t − τ ) ,

asymptotically stable modes and the modes on the imaginary axis (8)
ė(t ) = (A + HC ) e(t ) − ϑλBF (ϕ (t − τ ) − e (t − τ ))
are de-coupled, and, consequently, only the consensus problem for
multi-agent systems associated with these modes on the imagi- is asymptotically stable for all γ ∈ (0, γ ∗ ].
nary axis is required to be solved (see Remark 6 in Su and Huang
(2012a) for details). Hence, without loss of generality, we will im- 3. Consensus by the state feedback protocol
pose the following assumption on the multi-agent system (1).
In this section we present solutions to Problem 1 by develop-
Assumption 1. The matrix pair (A, B) is controllable, the matrix ing TPF based state feedback protocols. Both delay-dependent and
pair (A, C ) is observable, and all the eigenvalues of A are on the delay-independent protocols will be established. To deduce our
imaginary axis. main results, we first introduce the following lemma.
B. Zhou, Z. Lin / Automatica 50 (2014) 452–464 455

Lemma 2. The multi-agent system in (1) achieves consensus by the Remark 3. We would like to point out that the predictor feedback
following state feedback protocol based protocol (9) (as well as the observer based protocol given
  in Lemma 3 later) does not require the matrix A to satisfy
N  0

 Assumption 1, namely, A can be any square matrix. Moreover, if
ui ( t ) = F e zi (t ) + lij e −As
Buj (t + s) ds , (9) L is exactly known and the relative input ui − uj , i, j ∈ I [1, N]
j =1 −τ
is accessible (we would like to point out that the accessibility of
where F ∈ Rm×n is such that A + λi BF , i ∈ I [2, N], are all Hurwitz. the relative input is indeed assumed in some existing literature, for
example, You & Xie, 2011a), this protocol is implementable, though
Proof. We consider an artificial protocol as computationally expensive.

ui (t ) = Fzi (t + τ ) , i ∈ I [1, N] . (10) We next develop a truncated version of the predictor feedback
based protocol in (9) under Assumption 1. Under this assumption,
Then the closed-loop dynamics of the multi-agent system (1) and there exists a parameterized feedback gain F = F (γ ) : R+ →
the protocol (10) can be represented as, for all t ≥ τ , Rm×n , which is of order 1 with respect to γ , namely (Zhou, Lin, &
ẋ(t ) = (T ⊗ In ) (IN ⊗ A) T −1 ⊗ In x(t )
  Duan, 2011),

+ (T ⊗ In ) (JL ⊗ BF ) T −1 ⊗ In x(t ), 1
 
(11) lim ∥F (γ )∥ = 0, lim ∥F (γ )∥ < ∞. (18)
γ →0+ γ →0+ γ
where we have used (5). The above equation can be rewritten as
Then the predictor feedback based protocol ui (t ), i ∈ I [1, N], itself
χ̇(t ) = (IN ⊗ A) χ (t ) + (JL ⊗ BF ) χ (t ), (12) is ‘‘of order 1’’ with respect to γ , namely,
in which 1
lim ∥ui (t )∥ = 0, lim ∥ui (t )∥ < ∞,
T T γ →0+ γ →0+ γ
χ , χ1T χ2T χN , T − 1 ⊗ In x .
   
··· (13)
∀t ≥ 0, i ∈ I [1, N] . (19)
Since the first column of T is 1N , if limt →∞ ∥χi (t )∥ = 0, ∀i ∈
I [2, N], then it follows from x(t ) = (T ⊗ In ) χ (t ) that xi (t ) →
0
As a result, by virtue of (18) and denoting νi (t ) = F
N
i=1 lij −τ
χ1 (t ) , s(t ), ∀i ∈ I [1, N], as t approaches infinity and the e−As Buj (t + s), we have
consensus is achieved. On the other hand, in view of (12) and (5),
the dynamics of χi (t ), i ∈ I [2, N], obey the following equations 1 1
N
  0
∥νi (t )∥ ≤ lim ∥F (γ )∥
 −As 
lim lij e B
χ̇i (t ) = Aχi (t ) + λi BF χi (t ) + δi BF χi+1 (t ), (14) γ →0+ γ 2
γ →0+ γ j =1 −τ

where i ∈ I [2, N] and χN +1 (t ) ≡ 0. Hence the problem of consen- 1 


uj (t + s) ds

× lim
sus is solved by the protocol (10) as A + λi BF , i ∈ I [2, N], are all γ →0+ γ
Hurwitz. < ∞, (20)
We next show that the artificial protocol (10) is equivalent to N  0 −As
(9). In view of (2), we can write namely, the second term F i=1 lij −τ e Buj (t + s) in (9) is at
  least ‘‘of order 2’’ with respect to γ . This indicates that, no mat-
N 0
ter how large the value of τ is, the distributed term F
N
i=1 lij

ui (t ) = F lij xj (t + τcon ) , i ∈ I [1, N] . (15) −τ
j =1
e−As Buj (t + s) in (9) is dominated by the term F eAτ zi (t ) in (9) and
thus might be safely neglected in ui (t ) when γ is sufficiently small
On the other hand, by using the system equation in (1), we can pre- (Zhou et al., 2012). As a result, the predictor feedback based proto-
dict xj (t + τcon ) from xj (t − τcom ) as col in (9) can be truncated as

0
ui (t ) = F eAτ zi (t ), i ∈ I [1, N] , (21)

xj (t + τcon ) = eAτ xj (t − τcom ) + e−As Buj (t + s) ds, (16)
−τ which we refer to as the TPF based protocol.
by which and (2) the artificial protocol (15) can be further ex-
pressed as (9). The proof is completed.  Remark 4. Compared with the predictor feedback protocol in
(9), the TPF based protocol ui (t ) in (21) possesses the following
Remark 2. Under Assumption 2, it is always possible to find a advantages:
matrix F such that A + λi BF , i ∈ I [2, N], are all Hurwitz. In fact, F
(1) The TPF based protocol ui (t ) in (21) is independent of the
can be chosen as any optimal control type feedback gains, namely,
signals uj , j ∈ I [1, N] , j ̸= i, which is not the case for the
F = −µBT P, where µ is chosen such that
predictor feedback based protocol in (9).
(2) The TPF based protocol ui (t ) in (21) is static, while the predictor
 
1
µ ≥ max , µ∗ (L), (17) feedback based protocol ui (t ) in (9) is dynamic as it involves
i∈I[2,N] Re {λi }
the history information of uj (t ), j ∈ I [1, N] in the interval
in which Re {λi } > 0, i ∈ I [2, N], are guaranteed by Assumption 2, [t − τ , t].
and P solves the ARE AT P + PA − PBBT P = −Q , where Q > 0 is (3) The TPF based protocol ui (t ) in (21) does not require the exact
given. This can be verified by (A + λi BF )H P + P (A + λi BF ) ≤ −Q . information of L, while the predictor feedback based protocol
ui (t ) in (9) does.
From the proof of Lemma 2 we can see that the protocol in (9) is
obtained by predicting the future states from the past ones. Hence The following theorem shows that the TPF based protocol ui (t )
it will be referred to as predictor feedback based protocol. in (21) indeed solves Problem 1.
456 B. Zhou, Z. Lin / Automatica 50 (2014) 452–464

Theorem 1. Let Assumptions 1 and 2 be satisfied and P (γ ) be the Remark 6. In the TPF based protocols proposed in Theorems 1
unique positive definite solution to the parametric ARE (6). Then, and 2, the graph G (N , E , A) is not necessarily known exactly. In
for any bounded delays τcon and τcom that are arbitrarily large and fact, only the information of the bound on the eigenvalues of L is
exactly known, and for any µ ≥ µ∗ (L), there exists a number γ ∗ = required. For example, if it is known that the ith eigenvalue of L lies
γ ∗ (µ, τ , {|λi |}Ni=2 ) > 0 such that Problem 1 is solved by the TPF based in a compact bounded set Ωi , i ∈ I [2, N], then the scalar µ∗ (L) in
protocol ui (t ) in (21) where Theorems 1 and 2 can be replaced by

F = −µBT P (γ ) , ∀µ ∈ [µ∗ (L), ∞), ∀γ ∈ (0, γ ∗ ].


 
(22) 1
µ (Ω ) =

sup . (28)
λi ∈Ωi , i∈I[2,N ] Re {λi }
Proof. Similarly to the proof of Lemma 2, with the TPF based
protocol in (21), we can show that the consensus problem is solved
if the following series of time-delay systems are asymptotically Remark 7. In the existing literature (see, for example, Lin et al.,
stable 2007, Olfati-Saber & Murray, 2004, Qin et al., 2011, Tian & Liu, 2009
and Yu et al., 2010b), the problem of finding the maximal allowable
χ̇i (t ) = Aχi (t ) + λi BF eAτ χi (t − τ ) delay with a prescribed protocol has been investigated. In this
+ δi BF eAτ χi+1 (t − τ ) , i ∈ I [2, N] , (23) literature, even for multi-agent systems with agents characterized
by a single or double integrator, the allowable delay cannot be
where χN +1 (t ) ≡ 0. In addition, all the states xi (t ), i ∈ I [1, N], arbitrarily large (see, for example, Olfati-Saber & Murray, 2004 and
converge to χ1 (t ), which satisfies χ̇1 (t ) = Aχ1 (t ). Clearly, stability Yu et al., 2010b). Here we have shown in Theorems 1 and 2 that
of (23) is equivalent to the stability of arbitrarily large bounded delays are allowable as long as the gain
in the protocol is adjusted low enough.
χ̇i (t ) = Aχi (t ) + λi BF eAτ χi (t − τ ) , i ∈ I [2, N] . (24)
However, the stability of the series of time-delay systems in (24) Remark 8. It is shown in Munz, Papachristodoulou, and Allgower
follows from Item 1 of Lemma 1 with (θ , ϑ) = (0, 0).  (2011b) that consensus in single-integrator multi-agent systems
can be guaranteed for arbitrary large delays. The same system and
Remark 5. It follows from the proof of Lemma 1 that the norm a similar result are also reported in Liu, Lu, and Chen (2010), where
of the gain F = −µBT P (γ ) decreases as the delay gets larger some mild conditions are established to guarantee the consensus.
and γ ∗ (µ, τ , {λi }Ni=2 ) decreases. Hence the norm of the gain F eAτ Extension of the results in Munz et al. (2011b) to a more general
in the protocol (21) decreases as the delay gets higher if A is case that the agents are characterized by non-identical single-
Lyapunov stable (namely,  all the eigenvalues on the imaginary input–single-output systems having only unstable poles at zero
axis are simple) since eAτ  = 1. If A contains at least one pair of is given in Munz et al. (2012). Theorem 2 can be considered as
imaginary eigenvalues whose algebraic multiplicity is greater than a generalization of these results since the systems considered in
1, numerical simulation also shows that the norm of the gain F eAτ Theorem 2 include the single-integrator system and/or single-
in the protocol (21) decreases as the delay gets larger. Hence, a input–single-output systems as very special cases. On the other
large delay will not lead to large control effort. On the other hand, hand, under the same assumptions made in this paper, delay-
as the term eAτ is very close to the identity matrix for very small dependent protocols are established in Wang et al. (2013) for
delays, the gain F eAτ is very close to F , which is frequently used in the consensus of high-order multi-agent systems with constant
the literature. In this case, our method is comparable to the existing communication delays. However, in that paper the delays are not
approaches. allowed to be arbitrarily large in the case that A has nonzero
In Theorem 1, the delays τcon and τcom should be known exactly, eigenvalues and only constant delays are allowed in the case that
which may not be the case all the time. In the following we present A has only zero eigenvalues.
an alternative delay-independent protocol which does not require
the exact information of τcon and τcom . Remark 9. As pointed out in Munz et al. (2011b), the convergence
rate tends to decrease as the delays increase, in other words,
Theorem 2. Let Assumptions 1 and 2 be satisfied and, in addition, all consensus is reached slower for larger delays. This is also the case
the eigenvalues of A be zero. Let P (γ ) be the unique positive definite for the other theorems in this paper and should be considered as
solution to the parametric ARE (6). Assume that the delays τcon and the inherent trade-off between the size of the allowable delay and
τcom are time-varying, unknown and arbitrarily large, yet bounded, the convergence rate of the consensus.
namely, there exists a scalar τ ∗ < ∞ such that
4. Consensus by an observer based output feedback protocol
sup {|τcon (t )|} + sup {|τcom (t )|} ≤ τ ∗ . (25)
t ∈R t ∈R
In this section, we give solutions to Problem 2 by proposing two
Then, for any µ ≥ µ∗ (L), there exists a number γ ∗ = γ ∗ (µ, τ ∗ , kinds of observer based output feedback protocols with the help of
{|λi |}Ni=2 ) such that Problem 1 is solved by the idea of TPF. To this end, we first introduce a lemma which is a
ui (t ) = Fzi (t ), i ∈ I [1, N] , (26) generalization of Lemma 2 to the observer based output feedback
case.
where
Lemma 3. The multi-agent system in (1) achieves consensus by the
F = −µBT P (γ ) , ∀µ ∈ [µ∗ (L), ∞), ∀γ ∈ (0, γ ∗ ]. (27) following observer based output feedback protocol
N

Proof. The proof is similar to that of Theorem 1 except for the 
use of Item 2 of Lemma 1 with (θ , ϑ) = (0, 0) and τ = τ (t ) , ω̇ ( ) ( )ω ( ) lij uj (t − τ ) − H wi (t ),



 i t = A + HC i t + B
τcon (t ) + τcom (t − τcon (t )). The details are omitted for brevity. 

j =1
N  0
(29)

 
A couple of remarks regarding Theorems 1 and 2 are given in 
 u ( t ) = F e ω (t ) + F e−As Blij uj (t + s) ds,
 i i
order.

j =1 −τ
B. Zhou, Z. Lin / Automatica 50 (2014) 452–464 457

where i ∈ I [1, N] , F ∈ Rm×n is such that A + λi BF , i ∈ I [2, N], are I [1, N], which together with limt →∞ ∥ei (t )∥ = 0, i ∈ I [1, N],
all Hurwitz and H ∈ Rn×p is such that A + HC is Hurwitz. In addition, indicate that limt →∞ ∥ωi (t )∥ = limt →∞ ∥ri (t ) − ei (t )∥ = 0, i ∈
the observer states ωi (t ) satisfy limt →∞ ∥ωi (t )∥ = 0, i ∈ I [1, N]. I [1, N].
The proof is finished by noting that the artificial protocol ui (t )
Proof. Consider the following new state variable defined in (38) is equivalent to the protocol ui (t ) defined in (29)
N with the help of Eq. (31). 

ri (t ) = lij xj (t − τcom ) , ∀t ≥ τ , i ∈ I [1, N] , (30)
Remark 10. The introduction of the new state variable ri (t ) de-
j =1
fined in (30) is to put the term B i=1 lij uj (t − τ ) into the dynamic
N
on which the dynamics of the multi-agent system (1) can be equation (31) so that an observer in the form of (29) can be con-
rewritten as, for t ≥ τ , structed to obtain the clean error system (37). On the other hand,
N
we can see that the protocol ui (t ) defined in (29) is also based on
the predictor feedback, which is similar to the state feedback case.

ṙi (t ) = Ari (t ) + B lij uj (t − τ ) , i ∈ I [1, N] . (31)
For this reason, we call (29) an observer based output predictor
j =1
feedback protocol.
In addition, the information for the feedback given in (3) reduces
We next establish a truncated version of the observer based
to
output predictor feedback protocol in (29). Similarly to the state
wi (t ) = Cri (t ), i ∈ I [1, N] , ∀t ≥ τ . (32) feedback case, if F (γ ) is such that (18) is satisfied, then the term

The relation in (30) can be written in compact form N 


 0
F e−As Blij uj (t + s) ds, (42)
r (t ) = (L ⊗ In ) x (t − τcom ) j =1 −τ

= (T ⊗ In ) (JL ⊗ In ) T −1 ⊗ In x (t − τcom ) ,
 
(33) in (29) is at least ‘‘of order 2’’ with respect to γ and can thus be
neglected provided γ is sufficiently small. On the other hand, as
which, by denoting
the term B i=1 lij uj (t − τ ) is at least ‘‘of order 1’’ with respect
N
ρ = T − 1 ⊗ In r , χ = T −1 ⊗ In x, to γ and the other three terms in the ωi dynamics in (29) are
   
(34)
independent of γ , it can also be neglected if γ is small enough.
can be further rewritten as Therefore, the observer based predictor feedback protocol in (29)
ρ(t ) = JL χ (t − τcom ) , ∀t ≥ τ . (35) can be truncated as
ω̇i (t ) = (A + HC ) ωi (t ) − H wi (t ),

It follows from (35) and (5) that ρ1 (t ) ≡ 0, ∀t ≥ τ , and (43)
ui (t ) = F eAτ ωi (t ), i ∈ I [1, N] ,
ρi (t ) = λi χi (t − τcom ) + δi χi+1 (t − τcom ) , (36)
which will be referred to as the observer based TPF protocol.
where t ≥ τ , i ∈ I [2, N], and χN +1 (t ) ≡ 0. Now, by denoting Compared with the observer based predictor feedback protocol
ei (t ) = ri (t ) − ωi (t ), i ∈ I [1, N], we get from (29) and (31) that (29), the observer based TPF protocol (43) also possesses the three
advantages listed in Remark 4.
ėi (t ) = (A + HC ) ei (t ), i ∈ I [1, N] , ∀t ≥ τ , (37)
Similarly to the state feedback case, the following theorem
which implies that limt →∞ ∥ei (t )∥ = 0, i ∈ I [1, N]. shows that the observer based TPF protocol (43) indeed solves
Now consider the following artificial protocol Problem 2 for arbitrarily large yet bounded delay τ .

ui (t ) = Fri (t + τ ) − F eAτ ei (t ), ∀t ≥ 0, (38) Theorem 3. Let Assumptions 1 and 2 be satisfied, H be such that A +
HC is Hurwitz, and P (γ ) > 0 be the unique positive definite solution
by which the equation in (31) can be written as to the parametric ARE (6). Then for any delays τcon and τcom that are
N exactly known and can be arbitrarily large yet bounded and for any
lij F rj (t ) − eAτ ej (t − τ ) , µ ≥ µ∗ (L), there exists a number γ ∗ = γ ∗ (µ, τ , {|λi |}Ni=2 , H ) such

ṙi (t ) = Ari (t ) + B
 
(39)
j =1
that Problem 2 is solved by (43), where ωi (0) ∈ Rn and

where i ∈ I [2, N] , ∀t ≥ τ . The above equation can be expressed F = −µBT P (γ ) , ∀µ ∈ [µ∗ (L), ∞), ∀γ ∈ (0, γ ∗ ]. (44)
in compact form
T
Proof. By denoting ~i (t ) = xTi (t ), ωiT (t ) , we can write the

ṙ (t ) = (IN ⊗ A) r (t ) + (L ⊗ BF ) r (t )
closed-loop multi-agent system consisting of (1) and (43) as
− L ⊗ BF eAτ e (t − τ ) ,
 
(40)
~(
˙ t ) = (IN ⊗ A1 ) ~(t ) + (IN ⊗ A2 ) ~ (t − τcon )
which, via the transformation in (34), is equivalent to − (L ⊗ A3 ) ~ (t − τcom ) , (45)
ρ̇1 (t ) = Aρ1 (t ),

where
ρ̇i (t ) = (A + λi BF ) ρi (t ) + δi BF ρi+1 (t ) (41)
BF eAτ
   
A 0
 −λi BF eAτ εi (t − τ ) − δi BF eAτ εi+1 (t − τ ) , A1 , , A2 ,
0
,
0 A + HC 0 0
(46)
where i ∈ I [1, N] , ε(t ) = T −1 ⊗ In e(t ), ρN +1 (t ) ≡ 0 and
   
0 0
εN +1 (t ) ≡ 0. As limt →∞ ∥e(t )∥ = 0, ρ1 (t ) ≡ 0, and A + λi BF , i ∈ A3 , .
HC 0
I [2, N], are all Hurwitz, it follows from (41) that limt →∞ ∥ρi (t )∥ =
0, i ∈ I [1, N]. Thus, we have from (36) that limt →∞ ∥χi (t )∥ = By denoting χ (t ) = T −1 ⊗ I2n ~(t ), we can express the dynamics
 
0, i ∈ I [2, N]. Consequently, by the property of T , we get from in (45) as
(34) that xi (t ) → χ1 (t ), ∀i ∈ I [1, N] as t approaches infinity,
namely, the consensus is achieved. Moreover, it follows from χ̇ (t ) = (IN ⊗ A1 ) χ (t ) + (IN ⊗ A2 ) χ (t − τcon )
limt →∞ ∥ρi (t )∥ = 0, i ∈ I [1, N], that limt →∞ ∥ri (t )∥ = 0, i ∈ − (JL ⊗ A3 ) χ (t − τcom ) , (47)
458 B. Zhou, Z. Lin / Automatica 50 (2014) 452–464

which, in view of the structures of A1 , A2 and A3 , is equivalent to based output feedback protocol

χ̇1 (t ) = A1 χ1 (t ) + A2 χ1 (t − τcon ) , ω̇i (t ) = (A + HC ) ωi (t ) − H wi (t ),


 
(58)
χ̇i (t ) = A1 χi (t ) + A2 χi (t − τcon ) − λi A3 χi (t − τcom ) (48) ui (t ) = F ωi (t ), i ∈ I [1, N] ,
−δi A3 χi+1 (t − τcom ) , i ∈ I [2, N] ,
where ωi (0) ∈ Rn and
where χN +1 (t ) ≡ 0. If limt →∞ ∥χi (t )∥ → 0, ∀i ∈ I [2, N], then, by
the property of T , we see that F = −µBT P (γ ) , ∀µ ∈ [µ∗ (L), ∞), ∀γ ∈ (0, γ ∗ ]. (59)

lim ∥~i (t ) − χ1 (t )∥ = 0, ∀i ∈ I [1, N] . (49) Proof. The proof is quite similar to the proof of Theorem 3 and
t →∞
is omitted for brevity. The only difference is that τcom should be
Notice that the first equation in (48) can be written as
assumed to be constant so that the transformation (54) (or the
transformation (30)) is valid. 
χ̇11 (t ) = Aχ11 (t ) + BF eAτ χ12 (t − τcon ) ,

(50)
χ̇12 (t ) = (A + HC ) χ12 (t ), Remark 11. Differently from the observer based output predictor
T feedback protocol (29), which makes the error dynamics and the
where χ1 (t ) = χ11 (t ), χ12 (t ) . Since the second subsystem in
 T T
dynamics of the original systems decoupled, as indicated by (37)
(50) is exponentially stable, it follows from (49) that, for all i ∈
and (39), the error dynamics and the dynamics of the original
I [1, N], systems under the observer based TPF protocol (43) are coupled,
lim ∥xi (t ) − χ11 (t )∥ = 0, lim ∥ωi (t )∥ = 0, (51) as indicated by (56). Hence an intricate Lyapunov analysis of the
t →∞ t →∞ stability of the closed-loop multi-agent system has to be developed
which implies that the consensus is achieved and the observer is in the proof of Lemma 1.
asymptotically stable. Therefore, in the remaining of the proof, we
need only to prove the stability of the second systems in (48), or Remark 12. We point out that we can obtain an explicit formula-
equivalently, the stability of tion of the functions s(t ) associated with Theorems 1–4. We how-
ever have not presented it here for the sake of brevity. In addition,
χ̇i (t ) = A1 χi (t ) + A2 χi (t − τcon ) − λi A3 χi (t − τcom ) , (52) as all the eigenvalues of A are on the imaginary axis, we can show
that there exists a number k > 0 such that
where i ∈ I [2, N].
T   x (θ ) 
  
By denoting χi (t ) = ϕi1 (t ), φi1T (t ) , ∀i ∈ I [2, N], we can
T
  ∗
∥s(t )∥ ≤ k 1 + t N −1 sup u (θ)
  + ∥ω(0)∥ (60)
write the systems in (52) as θ∈[−τ ,0]

ϕ̇i1 (t ) = Aϕi1 (t ) + BF eAτ φi1 (t − τcon ) , where N ∗ is the maximal algebraic multiplicity of the eigenvalues

(53)
φ̇i1 (t ) = (A + HC ) φi1 (t ) − λi HC ϕi1 (t − τcom ) , of A.
In the case that τ = τcon = τcom = 0, namely, the dynamics of
i φi1 (t ) and
where i ∈ I [2, N]. Let φi (t ) = λ− 1
the multi-agent system (1) becomes
ϕi (t ) = ϕi1 (t − τcom ) , i ∈ I [2, N] . (54) 
ẋi (t ) = Axi (t ) + Bui (t ),
(61)
Then system (53) simplifies to, for all i ∈ I [2, N], yi (t ) = Cxi (t ), i ∈ I [1, N] ,

and the partial information for feedback given in (3) becomes


ϕ̇i (t ) = Aϕi (t ) + λi BF eAτ φi (t − τ ) ,

(55)
φ̇i (t ) = (A + HC ) φi (t ) − HC ϕi (t ), N

wi ( t ) = lij yj (t ), i ∈ I [1, N] , (62)
which, with ei (t ) = ϕi (t ) − φi (t ), can be further rewritten as, for j =1
any i ∈ I [2, N],
we obtain the following corollary of Theorem 3.
ϕ̇i (t ) = Aϕi (t ) + λi BF eAτ ϕi (t − τ ) − λi BF eAτ ei (t − τ ) ,

ė (t ) = (A + HC ) ei (t ) + λi µBF eAτ (56) Corollary 1. Let Assumptions 1 and 2 be satisfied, H ∈ Rn×p be such
 i that A + HC is Hurwitz, µ be any positive number such that (17) is
× (ϕi (t − τ ) − ei (t − τ )) .
satisfied, and P (γ ) > 0 be the unique positive definite solution to the
The stability of (56) then follows from Item 1 of Lemma 1 where parametric ARE (6). Then for any µ ≥ µ∗ (L), there exists a number
(θ, ϑ) = (−1, −1). The proof is finished.  γ ∗ = γ ∗ (µ, {|λi |}Ni=2 , H ) such that the consensus of the multi-agent
system in (61)–(62) is achieved by the following observer based
If the delays in the inputs and communication network are output feedback protocol
unknown and/or time-varying, we can obtain the following
ω̇i (t ) = (A + HC ) ωi (t ) − H wi (t ),

result, which can be regarded as the output feedback version of
(63)
Theorem 2. ui (t ) = F ωi (t ), i ∈ I [1, N] ,

Theorem 4. Let Assumptions 1 and 2 be satisfied, H be such that where ωi (0) ∈ Rn and
A + HC is Hurwitz, and, in addition, all the eigenvalues of A be zero. Let
P (γ ) > 0 be the unique positive definite solution to the parametric F = −µBT P (γ ) , ∀µ ∈ [µ∗ (L), ∞), ∀γ ∈ (0, γ ∗ ]. (64)
ARE (6). Assume that τcon and τcom are unknown and arbitrarily large,
yet bounded, τcon is possibly time-varying and τcom is constant. Let Remark 13. Consensus of the multi-agent system in (61)–(62) by
dynamic output feedback was solved in Seo et al. (2009) where a
sup {|τcon (t )|} + τcom ≤ τ < ∞. ∗
(57)
t ∈R dynamic output feedback protocol in the form of

Then for any µ ≥ µ∗ (L), there exists a number γ ∗ = γ ∗ (µ, τ ∗ , ω̇i (t ) = (A + HC + BF ) ωi (t ) − H wi (t ),



{|λi |}Ni=2 , H ) such that Problem 2 is solved by the following observer (65)
ui (t ) = F ωi (t ), i ∈ I [1, N] ,
B. Zhou, Z. Lin / Automatica 50 (2014) 452–464 459

is proposed, where F = −BT P with P (ε) > 0 being the solution to

AT P + PA − τ0 PBBT P + ε In = 0, ε ∈ (0, 1], (66)

with τ0 = mini∈I[2,N] {Re {λi }}. Our dynamic protocol (63) in Corol-
lary 1 possesses a simpler structure than the dynamic protocol (65)
as we do not require that the term corresponds to the term BF ωi (t )
in (65). More importantly, as seen in the Appendix and Example 1,
the analytical expression of the gains in (63) can be obtained by
solving a linear Lyapunov equation, whereas the ARE in (66) has to
be solved numerically for each given value of ε .

The following result is a corollary of Lemma 3.

Corollary 2. Let Assumption 2 be satisfied, (A, B) be controllable and


(A, C ) be observable. Then the consensus of the multi-agent system
in (61)–(62) is achieved by the following observer based output
feedback protocol
 Fig. 1. Weighted communication network topology.
 N
ω̇i (t ) = (A + HC ) ωi (t ) + lij BF ωj (t ) − H wi (t ),


(67) The network in Fig. 1 clearly contains a directed spanning tree. In
j =1
ui (t ) = F ωi (t ), fact, the eigenvalue set of L can be computed asλ(L) ={0, 3.5698,


3.2151 ± 1.3071i}. We choose µ = maxi∈I[2,4] 1
= 0.3310.
where F ∈ Rm×n is such that A + λi BF , i ∈ I [2, N], are all Hurwitz Re {λi }

and H ∈ Rn×p is such that A + HC is Hurwitz. In addition, the observer


states ωi (t ) satisfy limt →∞ ∥ωi (t )∥ = 0, i ∈ I [1, N]. We construct the observer based output TPF protocol to solve
the consensus problem. By solving the parametric ARE (6), the
Remark 14. In Li et al. (2010) the following observer based output feedback gain defined in (44) can be computed as F eAτ = −µ [f1i ] ,
feedback protocol i ∈ I [1, 4], where

2γ 3 cos (τ ω0 ) γ4

N   
2γ 2 − sin (τ ω0 ) ,

ω̇i (t ) = (A + BF ) ωi (t ) + lij HC ωj (t ) − H wi (t ),

 f11 = −
ω0 2ω02

(68)


j =1


ui (t ) = F ωi (t ), 2γ 3 sin (τ ω0 ) γ4

 
  
+ 2γ 2 − cos (τ ω0 ) ,

f12 =


ω0 2ω02


is proposed to solve the consensus problem for the multi-agent


2γ 3 τ cos (τ ω0 ) γ4
  
system (61)–(62), where F ∈ Rm×n is such that A + BF is Hurwitz

− 2γ 2 − τ sin (τ ω0 )

f13 =


and H ∈ Rn×p is such that A +λi HC , i ∈ I [2, N], are all Hurwitz. It is
 4 ω0 2ω02
(72)
very interesting to notice that our observer based output feedback γ γ2

cos (τ ω0 ) − 4γ sin (τ ω0 ) ,

protocol (67) in Corollary 2 can be regarded as a dual form of (68).
 + +4
2ω0 ω0


 3


2γ 3 τ sin (τ ω0 ) γ4

  
γ 2
τ cos (τ ω0 )

5. Two numerical examples f = + 2 −

14

ω0 2ω02



γ γ

4 2

  
sin (τ ω0 ) + 4γ cos (τ ω0 ) .

Example 1. We consider four identical agents whose dynamics is
 + +4
ω03
ω0
given by
Let the observer gain H be chosen as
ω0
 
0 1 0
−ω0 0 0 1
ẋi (t ) =  xi ( t )
T
37.2391 −0.6667 29.5892 ,

0 0 0 ω0  H = 11 (73)
0 0−ω0 0
  which is such that λ (A + HC ) = {−4, −3, −2 ± i}. The observer
0
0 based output TPF protocol can then be constructed according
+   ui (t − τcon ) , (69) to (43).
0 √
1 For the simulation purpose, we choose ω0 = 2/ 3, γ = 0.1,
and the initial conditions for the four agents as x1 (θ ) = [3, −4,
yi (t ) = −1 0 xi (t ), i ∈ I [1, 4] ,
 
0 0 (70) 2, 4]T , x2 (θ ) = [−2, 3, −3, 2]T , x3 (θ ) = [4, −2, 2, 3]T , x4 (θ ) =
where ω0 > 0 is a constant. It follows that Assumption 1 is fulfilled [2, 3, −3, 2]T , and ui (θ ) = 0, i ∈ I [1, 4] , ∀θ ∈ [−τ , 0]. The
since all the eigenvalues of A are at ±ω0 i whose algebraic multi- initial conditions for the observer are respectively assigned as
plicity is 2, (A, B) is controllable and (A, C ) is observable. Assume ω1 (0) = [6, −3, 4, 1]T , ω2 (0) = [4, −3, 5, 1]T , ω3 (0) = [−4, 4,
that τcon = 32 s and τcom = 12 s, and consequently, τ = 2s. Let the 7, 2]T and ω4 (0) = [5, 2, −4, −2]T .
communication network be given in Fig. 1, which is characterized The differences between the states of Agent 1 and those of the
by the Laplacian other agents are recorded in Fig. 2. The observer states are also
  plotted in Fig. 3. In these figures, xij , i, j ∈ I [1, 4], denotes the jth
3 0 −2 −1
elements of xi (the state of Agent i), and ωij denote the jth elements
−2 2 0 0 
L= . (71) of ωi . From these figures we clearly see that the consensus is
−1 −1 2 0 
0 0 −3 3 achieved by this observer based output TPF protocol.
460 B. Zhou, Z. Lin / Automatica 50 (2014) 452–464

Fig. 2. Example 1: Differences between the states of Agent 1 and those of the other agents.

Fig. 3. Example 1: States of the observer.

Example 2. We consider four agents characterized by the topology shown in Fig. 1. We assume that the delays in the
    communication network and inputs are respectively characterized
0 1 0 0 0 0 by
0 0 1 0 1 0 √ 
ẋi (t ) =  x (t ) +  u (t − τcon (t )) , (74)
0 0 0 0 i 0 1 i 1 t 1
0 0 0 0 1 1 τcon (t ) = + cos2
, τcom = , (76)
2 2 2
 
−1 0 0 1 which is such that |τcon (t )|+τcom ≤ τ ∗ = 2, ∀t ∈ R. We construct
yi (t ) = x (t ), i ∈ I [1, 4] . (75)
0 −1 0 1 i the observer based output TPF protocol to solve the consensus
problem.
Notice that the associated system matrix A is in a Jordan canonical
form. It follows that A has only zero eigenvalues whose geometry By solving the parametric ARE (6), the feedback gain F defined
multiplicity is 2 and maximal algebraic multiplicity is 3. Moreover, in (58) can be computed as in Box I.
we have that (A, B) is controllable and (A, C ) is observable. Let Here we also choose µ = 0.3310 as in Example 1 and γ =
the network between these four agents be again characterized by 0.075. By applying the function place in Matlab, we get the
B. Zhou, Z. Lin / Automatica 50 (2014) 452–464 461

Fig. 4. Example 2: Differences between the states of Agent 1 and those of the other agents.

γ4 2γ 3 γ (γ − 1) γ
 
γ2 + 1
 γ +2 1 2 γ2 + 1 γ 2 + 1  
2
F = −µ   (77)
 γ3 3+γ γ γ 2γ2 + 3 + 2γ + γ3 γ2 γ2 + 2 
 

γ2 + 1 γ2 + 1 γ2 + 1 γ2 + 1
Box I.

observer gain H such that λ (A + HC ) = {−4, −3, −2 ± i} as Appendix


T A.1. Proof of Lemma 1
18.0416 12.7979 5.0134 11.1312

H = . (78)
−5.0836 −0.5969 3.2795 −4.6864 Proof of Item 1. By the variation of constants formula, we obtain
from the first equation in (7) that, for all t ≥ 2τ ,
For the simulation purpose, we choose the initial conditions for  t
these four agents as the same as in Example 1. The differences Aτ
ϕ(t ) = e ϕ (t − τ ) + eA(t −s) (−λµBBT PeAτ ϕ (s − τ )
between the states of Agent 1 and those of the other agents and t −τ
the observer states are respectively recorded in Figs. 4 and 5.
From these figures we clearly see that the consensus of these four + θ λµBBT PeAτ e (s − τ ))ds, (79)
agents is also achieved by the proposed observer based output TPF substitution of which into the first system in (56) gives, for all
protocol. t ≥ 2τ ,
ϕ̇(t ) = A − λµBBT P ϕ(t ) − θ λµBBT PeAτ e (t − τ )
 
6. Conclusions
− (λµ)2 BBT P (π1 (t ) − π2 (t )) , (80)
This paper studies the consensus problem for multi-agent
where π1 (t ) and π2 (t ) are defined as
systems with agents characterized by high-order linear systems
with time delays existing in both the communication network   t
and inputs of the agents. Under the condition that the open- π (t ) = eA(t −s) BBT PeAτ ϕ (s − τ ) ds,

 1

t −τ
loop dynamics of the agents is at most polynomially unstable and  t (81)
the communication topology for the agents contains a directed A(t −s) Aτ
π2 (t ) = θ BB Pe e (s − τ ) ds.
T

 e
spanning tree, the truncated predictor feedback approach is t −τ
established to solve the consensus problems. It was proven that
The time-derivative of the Lyapunov function V1 (ϕ(t )) = ϕ H (t )
if the delays are exactly known, the consensus problems can be
P ϕ(t ) along the trajectories of (80) satisfies
solved by both full state feedback and observer based output
feedback protocol for arbitrarily large yet bounded delays. On the V̇1 (ϕ(t )) ≤ −γ ϕ H (t )P ϕ(t ) + (1 − 2µRe{λ}
other hand, the communication delays and input delays can be + 3κ |λ|4 µ4 )ϕ H (t )PBBT P ϕ(t )
allowed to be time-varying, arbitrarily large yet bounded, and nγ
even unknown, if the open-loop dynamics of the agents only + (π H (t )P π1 (t ) + π2H (t )P π2 (t )
contain zero eigenvalues. Numerical examples were worked out κ 1
+ θ 2 eH (t − τ ) eA τ PeAτ e (t − τ )),
T
to illustrate the effectiveness of the proposed approaches. (82)
462 B. Zhou, Z. Lin / Automatica 50 (2014) 452–464

Fig. 5. Example 2: States of the observer.

where we have used Lemma 4 in Appendix A.2. On the other hand, nγ θ 2 (n−1)γ τ H
by the Jensen inequality (Gu, 2000) and Lemma 4 in Appendix A.2, V̇3 (et ) = e e (t )Pe(t )
κ
we can compute nγ θ 2 (n−1)γ τ H
− e e (t − τ ) Pe (t − τ ) . (89)
 t
κ
π1H P π1 ≤ (nγ )2 τ e2(n−1)γ τ ϕ H (s)P ϕ(s)ds. (83) On the other hand, if we choose κ = 1
and µ ≥ 1
= µ∗ ,
3|λ|4 µ4 Re{λ}
t −2 τ
then
 t
2(n−1)γ τ
π H
2 P π2 ≤ θ (nγ ) τ e
2 2
e (s)Pe(s)ds.
H
(84) 1 − 2µRe {λ} + 3κ |λ|4 µ4 ≤ 0. (90)
t −2τ
Then we get from (85) and (88)–(90) that
Hence the inequality in (82) simplifies to
V̇1 (ϕ(t )) + V̇2 (ϕt , et ) + V̇3 (et )
V̇1 (ϕ(t )) ≤ −γ ϕ (t )P ϕ(t ) + 1 − 2µRe{λ} + 3κ |λ| µ
H 4 4 nγ
   
≤ − γ − 2τ (nγ )2 τ e2(n−1)γ τ ϕ H (t )P ϕ(t )
nγ κ
× ϕ H (t )PBBT P ϕ(t ) + (nγ )2 τ e2(n−1)γ τ
κ nγ θ 2
2τ (nγ )2 τ e2(n−1)γ τ + e(n−1)γ τ eH Pe.

t t + (91)
κ
  
× ϕ H (s)P ϕ(s)ds + θ 2 eH (s)Pe(s)ds
t −2τ t −2τ Now choose another Lyapunov function V4 (e(t )) = eH (t )Qe(t )
nγ θ 2 (n−1)γ τ H where Q > 0 satisfies
+ e e (t − τ ) Pe (t − τ ) . (85)
κ (A + HC )T Q + Q (A + HC ) = −In . (92)
Let V2 (ϕt , et ) and V3 (et ) be respectively defined as Then by using the second equation in (7) and Lemma 4 in
Appendix A.2, we have
 2τ  t
nγ 1
V2 (ϕt , et ) = (nγ )2 τ e2(n−1)γ τ ϕ (l)P ϕ(l)
 H
κ V̇4 (e(t )) ≤ − ∥e(t )∥2 + 4c (ϕ H (t − τ ) P ϕ (t − τ )
0 t −s 2
+ θ 2 eH (l)Pe(l) dlds,

(86)
+ eH (t − τ ) Pe (t − τ )), (93)
nγ θ 2 (n−1)γ τ t H

τ
where c = c (γ ) = |λ| µ B Q B nγ e( )γ ϑ . Choose finally

2 2 T 2 
 n − 1 2
V3 (et ) = e e (s)Pe(s)ds, (87)
κ t −τ the Lyapunov functional
t t
  
from which we can compute
V5 (ϕt , et ) = 4c ϕ H (s)P ϕ(s)ds + eH (s)Pe(s)ds , (94)
t −τ t −τ

V̇2 (ϕt , et ) = 2τ (nγ )2 τ e2(n−1)γ τ (ϕ H (t )P ϕ(t ) and it follows from (93) that
κ

+ θ 2 eH (t )Pe(t )) − (nγ )2 τ e2(n−1)γ τ 1
V̇4 (e(t )) + V̇5 (ϕt , et ) ≤ − ∥e(t )∥2 + 4c ϕ H (t )P ϕ(t )

κ
 t 2
ϕ (s)P ϕ(s) + θ 2 eH (s)Pe(s) ds,
 H 
× (88) + eH (t )Pe(t ) .

(95)
t −2τ
B. Zhou, Z. Lin / Automatica 50 (2014) 452–464 463

Now choose the total Lyapunov function as where we have again used Lemma 4 in Appendix A.2. Let
V (ϕt , et ) = V1 (ϕ(t )) + V2 (ϕt , et ) + V3 (et ) 
V (ϕ(t ), e(t )) = Vϕ (ϕ(t )) + ∥P ∥Ve (e(t )) . (107)
+ γ (V4 (e(t )) + V5 (ϕt , et )) , (96)
Then under the condition that
whose time-derivative, in view of (91) and (95), satisfies
V (ϕ (t + s) , e (t + s)) < ηV (ϕ(t ), e(t )) ,
1
V̇ (ϕt , et ) ≤ −γ f (γ ) ϕ (t )P ϕ(t ) −
H
γ g (γ ) ∥e(t )∥ ,
2
∀s ∈ −2τ ∗ , 0 ,
 
(97) (108)
2
where f (γ ) and g (γ ) are respectively related with where η > 1 is any prescribed number, we have
 
n ∥P ∥Ve (e (t + s)) , Vϕ (ϕ (t + s))
f (γ ) = 1 − 2τ (nγ )2 τ e2(n−1)γ τ max
κ 
− 4µ2 |λ|2 BT Q 2 B nγ e(n−1)γ τ ϑ 2 , ≤ V (ϕ (t + s) , e (t + s))

(98)
< ηV (ϕ(t ), e(t )) , ∀s ∈ −2τ ∗ , 0 .
 
n 2(n−1)γ τ
(109)
g (γ ) = 1 − 4θ τ 2
(nγ ) τ e
2
∥P ∥
κ
It follows from P ≤ ∥P ∥ In ≤ λ (Q ) Q and (109) that V̇ϕ (ϕ(t )) in
P ∥ ∥
n
− 2θ 2 e(n−1)γ τ ∥P ∥ min
κ  (103) and V̇e (e(t )) in (106) can be respectively continued as
− 8µ2 |λ|2 BT Q 2 B nγ e(n−1)γ τ ϑ 2 .

(99)
V̇ϕ (ϕ(t )) ≤ −γ ϕ H (t )P ϕ(t ) + γ h1 (γ ) ηV (ϕ(t ), e (t )) , (110)
As limγ →0+ P (γ ) = 0, there clearly exists a positive number
γ ∗ = γ ∗ (µ, τ , {|λ|}Ni=2 , H ) such that f (γ ) ≥ 12 , g (γ ) ≥ 12 , ∀γ ∈ 1
V̇e (e(t )) ≤ − ∥e(t )∥2 + γ h2 (γ ) ηV (ϕ(t ), e(t )) , (111)
(0, γ ∗ ]. Therefore it follows from (97) that 2
1 1 where h1 (γ ) and h2 (γ ) are respectively defined as
V̇ (ϕt , et ) ≤ − γ ϕ H (t )P ϕ(t ) − γ ∥e(t )∥2 ,
2 4 2n  2 2 2
∀µ ∈ [µ∗ , ∞), ∀γ ∈ (0, γ ∗ ], ∀t ≥ 2τ , (100) h1 (γ ) = 9τ∗ n γ + 3τ∗2 |λ|2 µ2 n2 γ 2
κ  √ √
θ 2 ∥P ∥ θ 2 ∥P ∥
 
and the asymptotic stability then follows from the Lyapunov × 1+ + , (112)
stability theorem. The proof is completed. λmin (Q ) λmin (Q )

Proof of Item 2. For any t ≥ 2τ∗ , integrating both sides of (8) from
 
∥P ∥
h2 (γ ) = 4ϑ n |λ| µ B Q B .
2

2 2 T 2 

t − τ to t gives ϕ (t − τ ) = ϕ(t ) − δ(t ), where 1+ (113)
λmin (Q )
 t
δ(t ) = (Aϕ(s) − λBF ϕ (s − τ ) + θ λBFe (s − τ )) ds. (101) It follows from (110)–(111) that
t −τ
V̇ (ϕ(t ), e(t )) ≤ −γ (1 − h (γ ) η) V (ϕ(t ), e(t ))
Then we can rewritten the first equation in system (8) as  
1
ϕ̇(t ) = (A − λBF ) ϕ(t ) + λBF (δ(t ) + θ e (t − τ )) , (102) − −γ ∥P ∥Ve (e(t )) , (114)
2
along whose trajectories the time-derivative of Vϕ (ϕ(t )) = ϕ H (t ) √
where h (γ ) = h1 (γ ) + ∥P ∥h2 (γ ). Hence, it follows from
P ϕ(t ) can be evaluated as
limγ →0 h (γ ) = 0 and (114) that there exists a number γ ∗ =
2nγ  γ ∗ (µ, τ∗ , {|λi |}Ni=2 ) ∈ 0, 12 such that
 
V̇ϕ (ϕ(t )) ≤ −γ ϕ H (t )P ϕ(t ) + δ H (t )P δ(t )
κ
1
+ θ 2 eH (t − τ ) Pe (t − τ ) , V̇ (ϕ(t ), e(t )) ≤ − γ V (ϕ(t ), e(t )) ,

(103)
2
where we have used Lemma 4 in Appendix A.2 and noticed that, if ∀µ ∈ [µ∗ , ∞), ∀γ ∈ (0, γ ∗ ], ∀t ≥ 2τ∗ . (115)
κ = 21 2 , then
|λ| µ
The stability of (8) then follows from the Razumikhin stability
1 − 2µRe{λ} + κµ |λ| ≤ 0, 2 2
∀µ ∈ [µ , ∞). ∗
(104) theorem (Hale, 1977). The proof is completed.

By using Lemma 4 in Appendix A.2, we can compute


A.2. Properties of solutions to a parametric ARE
 t
δ (t )P δ(t ) ≤ 3τ∗ |λ| µ n γ
H 2 2 2 2
(ϕ (s − τ ) P ϕ (s − τ )
H

t −τ
In this subsection we recall the following results from Zhou,
Lin, and Duan (2010), Zhou et al. (2012) regarding properties of
+ θ e (s − τ ) Pe (s − τ ))ds
2 H

 t solutions to the parametric ARE (6).

+ 9τ∗ n2 γ 2 ϕ H (s)P ϕ(s)ds. (105)


Lemma 4. Assume that the matrix pair (A, B) ∈ Rn×n , Rn×m is
 
t −τ
controllable and all the poles of A are on the imaginary axis. Then the
Now compute the time-derivative of Ve (e(t )) = eH (t )Qe(t ), where parametric ARE (6) has a unique positive definite solution P (γ ) =
Q solves (92), along the trajectories of the second equation in W −1 (γ ), where W (γ ) is the unique positive definite solution to
system (8) as γ
+ A + γ2 In W
 T 
the following Lyapunov equation W A + I
2 n
1 = BBT . Moreover, limγ →0+ P (γ ) = 0, ddγ P (γ ) > 0, ∀γ > 0,
V̇e (e(t )) ≤ − ∥e(t )∥2 + 4ϑ 2 nγ |λ|2 µ2 BT Q 2 B
 
2 T
tr BT P (γ ) B = nγ , P (γ ) BBT P (γ ) ≤ nγ P (γ ) , eA t P (γ ) eAt ≤
 
× ϕ H (t − τ ) P ϕ (t − τ )

e(n−1)γ t P (γ ). In particular, if all the eigenvalues of A are zero, then
+ eH (t − τ ) Pe (t − τ ) , AT PA ≤ 3 (nγ )2 P.

(106)
464 B. Zhou, Z. Lin / Automatica 50 (2014) 452–464

References Tian, Y., & Liu, C. (2009). Robust consensus of multi-agent systems with diverse
input delays and asymmetric interconnection perturbations. Automatica, 45,
Chen, G. R., & Duan, Z. S. (2008). Network synchronizability analysis: a graph- 1347–1353.
Tian, Y., & Zhang, Y. (2012). High-order consensus of heterogeneous multi-
theoretic approach. Chaos: An Interdisciplinary Journal of Nonlinear Science,
agent systems with unknown communication delays. Automatica, 48(6),
18(3), 037102.
1205–1212.
Duan, Z. S., Chen, G. R., & Huang, L. (2007). Complex network synchronizability: Wang, X., Saberi, A., Stoorvogel, A. A., Grip, H. F., & Yang, T. (2013). Consensus in
analysis and control. Physical Review E, 76(5), 056103. the network with uniform constant communication delay. Automatica, 49(8),
Duan, Z., Wang, J., Chen, G., & Huang, L. (2008). Stability analysis and decentralized 2461–2467.
control of a class of complex dynamical networks. Automatica, 44(4), You, K., & Xie, L. (2011a). Coordination of discrete-time multi-agent systems via
1028–1035. relative output feedback. International Journal of Robust and Nonlinear Control,
Fax, J., & Murray, R. (2004). Information flow and cooperative control of vehicle 21(13), 1587–1605.
formations. IEEE Transactions on Automatic Control, 49(9), 1465–1476. You, K., & Xie, L. (2011b). Network topology and communication data rate for
Gu, K. (2000). An integral inequality in the stability problem of time-delay systems. consensusability of discrete-time multi-agent systems. IEEE Transactions on
In The 39th IEEE conference on decision and control (pp. 2805–2810). Sydney, Automatic Control, 56(10), 2262–2275.
Australia. Yu, W., Chen, G., & Cao, M. (2010a). Distributed leader–follower flocking control for
Hale, J. (1977). Theory of functional differential equations. Springer-Verlag. multi-agent dynamical systems with time-varying velocities. Systems & Control
Hong, Y., Chen, G., & Bushnell, L. (2006). Distributed observers design for leader- Letters, 59(9), 543–552.
following control of multi-agent networks. Automatica, 44(3), 846–850. Yu, W., Chen, G., & Cao, M. (2010b). Some necessary and sufficient conditions
Ishii, H., & Tempo, R. (2010). Distributed randomized algorithms for the PageRank for second-order consensus in multi-agent dynamical systems. Automatica, 46,
computation. IEEE Transactions on Automatic Control, 55(9), 1987–2002. 1089–1095.
Ishii, H., Tempo, R., & Bai, E.-W. (2012). A web aggregation approach for distributed Zhang, H., & Lewis, F. L. (2012). Adaptive cooperative tracking control of higher-
randomized PageRank algorithms. IEEE Transactions on Automatic Control, order nonlinear systems with unknown dynamics. Automatica, 48, 1432–1439.
http://dx.doi.org/10.1109/TAC.2012.2190161. Zhang, H., Lewis, F. L., & Das, A. (2011). Optimal design for synchronization
Li, Z., Duan, Z., Chen, G., & Huang, L. (2010). Consensus of multiagent systems and of cooperative systems: state feedback, observer and output feedback. IEEE
synchronization of complex networks: a unified viewpoint. IEEE Transactions Transactions on Automatic Control, 56(8), 1948–1952.
on Circuits and Systems I: Regular Papers, 57(1), 213–224. Zhou, B., Lin, Z., & Duan, G. (2010). Global and semi-global stabilization of linear
Li, Q., & Jiang, Z. P. (2009). Global analysis of multi-agent systems based on Vicsek’s systems with multiple delays and saturations in the input. SIAM Journal on
model. IEEE Transactions on Automatic Control, 54(12), 2876–2881. Control and Optimization, 48(8), 5294–5332.
Li, Z., Liu, X., Lin, P., & Ren, W. (2011). Consensus of linear multi-agent systems Zhou, B., Lin, Z., & Duan, G. (2011). L∞ and L2 low gain feedback: their properties,
with reduced-order observer-based protocols. Systems & Control Letters, 60(7), characterizations and applications in constrained control. IEEE Transactions on
510–516. Automatic Control, 56(5), 1030–1045.
Li, T., & Xie, L. (2011). Distributed consensus over digital networks with limited Zhou, B., Lin, Z., & Duan, G. (2012). Truncated predictor feedback for linear systems
bandwidth and time-varying topologies. Automatica, 47(9), 2006–2015. with long time-varying input delays. Automatica, 48(10), 2387–2399.
Lin, P., Jia, Y., Du, J., & Yuan, S. (2007). Distributed consensus control for second-order
agents with fixed topology and time delay. In 2007 Chinese control conference
(pp. 577–581). Hunan, China.
Liu, X., Lu, W., & Chen, T. (2010). Consensus of multi-agent systems with unbounded Bin Zhou is a professor of the Department of Control
time-varying delays. IEEE Transactions on Automatic Control, 55(10), 2396–2401. Science and Engineering at the Harbin Institute of
Lv, J., Chen, G., & di Bernardo, M. (2010). On some recent advances in Technology. He was born in Luotian County, Huanggang,
synchronization and control of complex networks. In Proceedings of 2010 IEEE Hubei Province, PR China on July 28, 1981. He received
international symposium on circuits and systems (pp. 3773–3776). the Bachelor’s degree, the Master’s Degree and the Ph.D.
Lv, J., Chen, G., & Yu, X. (2011). Modelling, analysis and control of multi-agent degree from the Department of Control Science and
systems: a brief overview. In Proceedings of 2011 IEEE international symposium Engineering at Harbin Institute of Technology, Harbin,
on circuits and systems (pp. 2103–2106). China in 2004, 2006 and 2010, respectively. He was
Manitius, A., & Olbrot, A. (1979). Finite spectrum assignment problem for systems a Research Associate at the Department of Mechanical
with delays. IEEE Transactions on Automatic Control, 24(4), 541–552. Engineering, University of Hong Kong from December
Munz, U., Papachristodoulou, A., & Allgower, F. (2010). Delay robustness in 2007 to March 2008, a Visiting Fellow at the School of
consensus problems. Automatica, 46(8), 1252–1265. Computing and Mathematics, University of Western Sydney from May 2009 to
Munz, U., Papachristodoulou, A., & Allgower, F. (2011a). Robust consensus August 2009, and a Visiting Research Scientist at the Department of Electrical and
controller design for nonlinear relative degree two multi-agent systems with Computer Engineering, University of Virginia from July 2012 to August 2013. He
communication constraints. IEEE Transactions on Automatic Control, 56(1), joined the School of Astronautics, Harbin Institute of Technology in February 2009.
145–151. His current research interests include constrained control, time-delay systems,
Munz, U., Papachristodoulou, A., & Allgower, F. (2011b). Consensus in multi-agent
nonlinear control, and control applications in astronautic engineering. In these
systems with coupling delays and switching topology. IEEE Transactions on
areas, he has published about 100 papers, over 70 of which are in archival journals.
Automatic Control, 56(12), 2976–2982.
He is a reviewer for American Mathematical Review and is an active reviewer for a
Munz, U., Papachristodoulou, A., & Allgower, F. (2012). Delay robustness in non-
number of journals and conferences. He was selected as the ‘‘New Century Excellent
identical multi-agent systems. IEEE Transactions on Automatic Control, 57(6),
Talents in University’’, the Ministry of Education of China in 2011. He received
1597–1603.
Olfati-Saber, R., & Murray, R. (2004). Consensus problems in networks of agents the ‘‘National Excellent Doctoral Dissertation Award’’ in 2012 from the Academic
with switching topology and time-delays. IEEE Transactions on Automatic Degrees Committee of the State Council and the Ministry of Education of PR China.
Control, 49(9), 1520–1533. He is currently an associate editor on the Conference Editorial Board of the IEEE
Qin, J., Gao, H., & Zheng, W. X. (2011). Second-order consensus for multi-agent Control Systems Society.
systems with switching topology and communication delay. Systems & Control
Letters, 60(6), 390–397.
Ren, W., Beard, R., & Atkins, E. (2007). Information consensus in multi-vehicle
cooperative control. IEEE Control Systems Magazine, 27(2), 71–82. Zongli Lin is a professor of Electrical and Computer
Ren, W., & Cao, Y. (2011). Communications and control engineering series, Distributed Engineering at University of Virginia. He received his
coordination of multi-agent networks. London: Springer-Verlag. B.S. degree in mathematics and computer science from
Ren, W., Moore, K., & Chen, Y. (2006). High-order consensus algorithms in Xiamen University, Xiamen, China, in 1983, his Master
cooperative vehicle systems. In Proceedings of the 2006 IEEE international of Engineering degree in automatic control from Chinese
conference on networking, sensing and control (pp. 457–462). Academy of Space Technology, Beijing, China, in 1989, and
Rudy, C., & Olgac, N. (2011). An exact method for the stability analysis of linear his Ph.D. degree in electrical and computer engineering
consensus protocols with time delay. IEEE Transactions on Automatic Control, from Washington State University, Pullman, Washington,
56(7), 1734–1740. in 1994. His current research interests include nonlinear
Seo, J. H., Shim, H., & Back, J. (2009). Consensus of high-order linear systems using control, robust control, and control applications. He was
dynamic output feedback compensator: low gain approach. Automatica, 45, an Associate Editor of the IEEE Transactions on Automatic
2659–2664. Control (2001–2003), IEEE/ASME Transactions on Mechatronics (2006–2009) and
Su, Y., & Huang, J. (2012a). Stability of a class of linear switching systems with IEEE Control Systems Magazine (2005–2012). He was an elected member of the
applications to two consensus problems. IEEE Transactions on Automatic Control, Board of Governors of the IEEE Control Systems Society (2008–2010) and has served
57(6), 1420–1430. on the operating committees and program committees of several conferences.
Su, Y., & Huang, J. (2012b). Two consensus problems for discrete-time multi-agent He currently chairs the IEEE Control Systems Society Technical Committee on
systems with switching network topology. Automatica, 48(9), 1988–1997. Nonlinear Systems and Control and serves on the editorial boards of several journals
Su, H., Wang, X., & Chen, G. (2009). A connectivity-preserving flocking algorithm and book series, including Automatica, Systems & Control Letters, Science China
for multi-agent systems based only on position measurements. International Information Sciences, and Springer/Birkhauser book series Control Engineering.
Journal of Control, 83(7), 1334–1343. He is a Fellow of the Institute of Electrical and Electronics Engineers (IEEE), the
Su, H., Wang, X., & Chen, G. (2010). Rendezvous of multiple mobile agents with International Federation of Automatic Control (IFAC) and the American Association
preserved network connectivity. Systems & Control Letters, 59(5), 313–322. for the Advancement of Science (AAAS).

You might also like