Automatica 42 (2006) 1177 – 1182

www.elsevier.com/locate/automatica

Brief paper
Tracking control for multi-agent consensus with an active leader
and variable topology夡
Yiguang Hong a,∗ , Jiangping Hu a , Linxin Gao b
a Key Laboratory of Systems and Control, Institute of Systems Science, Chinese Academy of Sciences, Beijing 100080, China
b Institute of Systems Science, Wenzhou University, Zhejiang, China

Received 16 March 2005; received in revised form 4 November 2005; accepted 16 February 2006
Available online 17 April 2006

Abstract
In this paper, we consider a multi-agent consensus problem with an active leader and variable interconnection topology. The state of the
considered leader not only keeps changing but also may not be measured. To track such a leader, a neighbor-based local controller together
with a neighbor-based state-estimation rule is given for each autonomous agent. Then we prove that, with the proposed control scheme, each
agent can follow the leader if the (acceleration) input of the active leader is known, and the tracking error is estimated if the input of the leader
is unknown.
䉷 2006 Elsevier Ltd. All rights reserved.

Keywords: Multi-agent systems; Consensus; State estimation; Active leader

1. Introduction formations was given with a Nyquist-type criterion in (Fax

& Murray, 2004). Moreover, by a Lyapunov-based approach,
In recent years, there has been an increasing research interest Olfati-Saber and Murray (2004) solved the average-consensus
in the control design of multi-agent systems. Many results have problem with directed interconnection graphs or time-delays.
been obtained with local rules applied to each agent in a consid- In reality, some variables of the agents and/or the leader in
ered multi-agent system. These neighbor rules for each agent a multi-agent system may not be able to be measured. Fax and
are based on the average of its own information and that of its Murray (2004) raised this important issue regarding observer
neighbors or its leader (Fax & Murray, 2004; Jadbabaie, Lin, design for multi-agent systems, and first tackled this problem.
& Morse, 2003; Lin, Broucke, & Francis, 2004; Olfati-Saber However, many works remain to be done for the distributed
& Murray, 2004; Savkin, 2004). For example, Jadbabaie et observer design of networks of multiple agents.
al. (2003) demonstrated that a simple neighbor rule makes all With this background, we consider a consensus problem
agents eventually move in the same direction despite the ab- with an active leader with an underlying dynamics. Here, some
sence of centralized coordination and each agent’s set of neigh- variables (that is, the velocity and maybe the acceleration) of
bors changing with time as the system evolves under a joint an active leader cannot be measured, and each agent only gets
connection condition. Also, with a similar technique, Lin et the measured information (that is, the position) of the leader
al. (2004) studied three formation strategies for groups of mo- once there is a connection between them. In this paper, we
bile autonomous agents. The stability analysis of multi-vehicle propose an “observer” by inserting an integrator into the loop
for each agent to estimate the leader’s velocity. To analyze the
夡 This paper was not presented at any IFAC meeting. This paper was
problem, a Lyapunov-based approach is developed. With the
recommended for publication in revised form by Associate Editor Marco proposed estimation rule and a selected Lyapunov function, the
Campi under the direction of Editor Ian Petersen.
∗ Corresponding author. Tel.: +86 10 6265 1449; fax: +86 10 6258 7343. leader-following problem can be solved if the leader’s input
E-mail addresses: [email protected] (Y. Hong), [email protected] is known, while the tracking error can also be analyzed if the
(J. Hu), [email protected] (L. Gao). input is unknown.
0005-1098/$ - see front matter 䉷 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.automatica.2006.02.013
1178 Y. Hong et al. / Automatica 42 (2006) 1177 – 1182

2. Problem formulation Meanwhile, the connection weight between agent i and the
leader, denoted by bi , is time-varying, too. We assume that there
To solve coordination problems, graph theory is helpful. An are fixed positive constants i (i = 1, . . . , n) such that
undirected graph G on vertex set V = {1, 2, . . . , n} contains 
V and a set of unordered pairs E = {(i, j ) : i, j ∈ V}, which i if agent i is connected to the leader at t,
bi (t) = (3)
are called G’s edges. If there is an edge between two vertices, 0 otherwise.
the two vertices are called adjacent. A graph is simple if it
has no self-loops or repeated edges. If there is a path between The next lemma was given to check the positive definiteness of
any two vertices of a graph G, then G is connected, otherwise a matrix (Horn & Johnson, 1985).
disconnected. A subgraph X of G is an induced subgraph if
two vertices of V(X) are adjacent in X if and only if they are Lemma 1. Suppose that a symmetric matrix is partitioned as
adjacent in G. An induced subgraph X of G that is maximal,  
E 1 E2
subject to being connected, is called a component of G. E= ,
Here, we consider a system consisting of n agents and a E2T E3
leader. In the sequel, the state of agent i is denoted by xi for
where E1 and E3 are square. E is positive definite if and only
i=1, . . . , n. With regarding the n agents as the vertices in V, the
if both E1 and E3 − E2T E1−1 E2 are positive definite.
relationships between n agents can be conveniently described
by a simple and undirected graph G, which is defined so that
The following result is well-known in algebraic graph theory
(i, j ) defines one of the graph’s edges in case agents i and j are
(Godsil & Royle, 2001) and establishes a direct relationship
neighbors. Ni (t) denotes the set of labels of those agents which
between the graph connectivity and its Laplacian.
are neighbors of agent i (i = 1, . . . , n) at time t. The weighted
adjacency matrix of G is denoted by A = [aij ] ∈ R n×n , where
Lemma 2. Let G be a graph on n vertices with Laplacian
aii =0 and aij =aj i
0 (aij > 0 if there is an edge between agent
L. Denote the eigenvalues of L by 1 (L), . . . , n (L) satisfying
i and agent j). Its degree matrix D = diag{d1 , . . . , dn }
∈ R n×n
1 (L)  · · · n (L). Then 1 (L) = 0 and 1 = [1, 1, . . . , 1]T ∈
is a diagonal matrix, where diagonal elements di = nj=1 aij
R n is its eigenvector. Moreover, if G is connected, 2 > 0.
for i = 1, . . . , n. Then the Laplacian of the weighted graph is
defined as
In this paper, all the considered agents move in a plane:
L = D − A, (1)
ẋi = ui ∈ R 2 , i = 1, . . . , n, (4)
which is symmetric. In what follows, we mainly concern a
graph Ḡ associated with the system consisting of n agents and where ui is the control input. The leader of this consid-
one leader. In fact, Ḡ contains n agents (related to graph G) and ered multi-agent system is active; that is, its state variables
the leader with directed edges from some agents to the leader. keep changing. Its underlying dynamics can be expressed as
By “the graph, Ḡ, of this system is connected”, we mean that follows:
at least one agent in each component of G is connected to the ⎧
leader. ⎨ ẋ0 = v0 ,
v̇ = a(t) = a0 (t) + (t), x0 , v0 ,  ∈ R 2 , (5)
For the multi-agent system under consideration, the rela- ⎩ 0
tionships between neighbors (and the interconnection topol- y = x0 ,
ogy) change over time. Suppose that there is an infinite se-
quence of bounded, non-overlapping, contiguous time-intervals where y(t) = x0 (t) is the measured output and a(t) is the (ac-
[ti , ti+1 ), i = 0, 1, . . . , starting at t0 = 0. celeration) input. Note that (5) is completely different from
Denote S = {Ḡ1 , Ḡ2 , . . . , ḠN } as a set of the graphs with the agent dynamics (4). In other words, the agents will track a
all possible topologies, which includes all possible intercon- leader with a different dynamics.
nection graphs (involving n agents and a leader), and denote In our problem formulation, the input a(t) may not be com-
P = {1, 2, . . . , N} as its index set. pletely known. We assume that a0 (t) is known and (t) is
To describe the variable interconnection topology, we de- unknown but bounded with a given upper bound ¯ (that is,
¯ The input a(t) is known if and only if ¯ = 0. On
(t)  ).
fine a switching signal
: [0, ∞) → P, which is piecewise-
constant. Therefore, Ni and the connection weight aij (i = the other hand, y = x0 is the only variable that can be obtained
1, . . . , n, j = 1, . . . , n) are time-varying, and moreover, Lapla- directly by the agents when they are connected to the leader.
cian Lp (p ∈ P) associated with the switching interconnec- Our aim here is to propose a decentralized control scheme for
tion graph is also time-varying (switched at ti , i = 0, 1, . . .), each agent to follow the leader (i.e., xi → x0 ).
though it is a time-invariant matrix in any interval [ti , ti+1 ). In Since v0 (t) cannot be measured even when the agents are
our problem, we assume that there are fixed positive constants connected to the leader, its value cannot be used in the control
ij (i = 1, . . . , n; j = 1, . . . , n) such that design. Instead, we have to estimate v0 during the evolution.
 Note that, each agent has to estimate v0 only by the informa-
 = j i if agents i and j are connected at t, tion obtained from its neighbors in a decentralized way. The
aij (t) = ij (2)
0 otherwise. estimate of v0 (t) by agent i is denoted by vi (t) (i = 1, . . . , n).
 Y. Hong et al. / Automatica 42 (2006) 1177 – 1182 1179

Therefore, for each agent, the local control scheme consists of follows:
two parts:  
0
˙ = F
 + g, g= (9)
−1 ⊗ 
• a neighbor-based feedback law:
⎡ ⎤ where
   
ui = − k ⎣ aij (t)(xi − xj ) + bi (t)(xi − x0 )⎦ x̄ −k(L
+ B
) In
= , F
= ⊗ I2 .
j ∈Ni (t) v̄ −k(L
+ B
) 0
+ vi , k > 0, i = 1, . . . , n, (6) 3. Main results
where Ni is the set consisting of agent i’s neighbor agents;
• a dynamic neighbor-based system to estimate v0 In this section, the convergence analysis of system (9) is
⎡ given for the consensus problem of multi-agent system (8). If
the information of the input a(t) can be used in local control
v̇i = a0 − k ⎣ aij (t)(xi − xj ) + bi (t) design, we can prove that all the agents can follow the leader,
j ∈Ni (t) though the leader keeps changing. If not, we can also get some
⎤ estimation of the tracking error. We first assume that the in-
terconnection graph Ḡ is always connected, though the inter-
× (xi − x0 )⎦ , i = 1, . . . , n, (7) connection topology keeps changing; and then we consider an
extended case.
for some positive constant  < 1. In fact, (7), can be viewed As mentioned above, Ḡ is connected if at least one agent
as an “observer” in some sense. in each of its component is connected with the leader. To be
specific, if there are m
1 components, then the Laplacian Lp
Note that ui in (6) is a local controller of agent i, which only (for any p ∈ P) of the graph associated with n agents have m
depends on the information from its neighbors, and, in fact, zero eigenvalues. For simplicity, we can rearrange the indices
when v0 = 0, a = 0, the proposed control law (6) is consis- of n agents such that Lp can be rewritten as a block diagonal
tent with the one given by Olfati-Saber and Murray (2004). In matrix:
⎛ 1 ⎞
addition, with the neighbor-based estimation rule in a form of Lp
observer (7) to estimate the leader’s velocity, each agent re- ⎜ L2p ⎟
⎜ ⎟
lies only on the locally available information at every moment. Lp = ⎜ . ⎟,
⎝ .. ⎠
In other words, each agent cannot “observe” or “estimate” the
m
Lp
leader directly based on the measured information of the leader
if it is not connected to the leader. In fact, it has to collect the
where each block matrix Lip is also a Laplacian of the corre-
information of the leader in a distributed way from its neighbor
sponding component. For convenience, denote Mp = Lp + Bp ,
agents.
where Lp is the weighted Laplacian and Bp (p ∈ P) is the di-
Take
agonal matrix as defined in Section 2. The next lemma is given
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
x1 v1 u1 for Mp .
. . .
x = ⎝ .. ⎠ , v = ⎝ .. ⎠ , u = ⎝ .. ⎠ .
Lemma 3. If graph Ḡp is connected, then the symmetric matrix
xn vn un
Mp associated with Ḡp is positive definite.
Regarding the switching interconnection graphs, the closed-
loop system can be expressed as Proof. We only need to prove the case when m = 1. Let
 1 , . . . , n be the eigenvalues of Laplacian Lp in the increas-
ẋ = u = −k(L
+ B
) ⊗ I2 x + kB
1 ⊗ x0 + v, ing order. From Lemma 2, 1 = 0 and i > 0, i
2. Denote n
(8)
v̇ = 1 ⊗ a0 − k(L
+ B
) ⊗ I2 x + k(B
1) ⊗ x0 , eigenvectors of Lp by i , i = 1, . . . , n, with 1 = 1, an eigen-
vector of Lp corresponding to 1 =
0. Then any nonzero vector
where Il ∈ R l×l (for any positive integer l) is the identity z ∈ R n can be expressed by z = ni=1 ci i for some constants
matrix and ⊗ denotes the Kronecker product,
: [0, ∞) → ci , i = 1, 2, . . . , n. Moreover, Bp  = 0 since there is at least
P={1, 2, . . . , N} is a piecewise-constant switching signal with one agent connected to the leader. Without loss of generality,
successive switching times, B
is an n × n diagonal matrix we assume bj > 0 for some j, and it is obvious T1 Bp 1
bj .
whose ith diagonal element is bi (t) at time t, L
is the Lapla- Therefore, in either the case when c2 = · · · = cn = 0 (so c1 = 0)
cian for the n agents. Note that, even in the case when the in- or the case when ci  = 0 for some i
2, we always have
terconnection graph is connected, bi (t) may be always 0 for
n
some i, and therefore, B
may not be of full rank.
Denote x̄ = x − 1 ⊗ x0 and v̄ = v − 1 ⊗ v0 . Because z T M p z = z T Lp z + z T Bp z
i ci2 Ti i + zT Bp z > 0
−k(L
+ B
) ⊗ I2 x + kB
1 ⊗ x0 = −k(L
+ B
) ⊗ I2 x̄ (in- i=2

voking Lemma 2), we can obtain an error dynamics of (8) as for z  = 0, which implies the conclusion. 
1180 Y. Hong et al. / Automatica 42 (2006) 1177 – 1182

Based on Lemma 3 and the fact that the set P is finite, in the following forms:

¯ = min{eigenvalues of Mp ∈ R n×n , ∀Ḡp is i,1 = (1 − 2 )ki (Mp ) + 


connected} > 0, (10)
+ [(1 − 2 )ki (Mp ) + ]2 − 4(1 − 2 )ki (Mp ) + 1,
is fixed and depends directly on the constants ij and i for i =
1, . . . , n, j =1, . . . , n given in (2) and (3). Its estimation is also i,2 = (1 − 2 )ki (Mp ) + 

related to the minimum nonzero eigenvalue of Laplacian Lp , − [(1 − 2 )ki (Mp ) + ]2 − 4(1 − 2 )ki (Mp ) + 1,
which has been widely studied in different situations (Merris,
1994). for i = 1, . . . , n. Clearly, the smallest eigenvalue of Qp will be
In some existing works (including Jadbabaie et al., 2003; found in the form of i,2 for some i.
Lin et al., 2004), the convergence analysis depends on theory Note that (11) implies ki (Mp ) > 1/4(1 − 2 ). In this case,
of nonnegative matrices or stochastic matrices. However, Fp i,2 increases as ki (Mp ) increases. Therefore, the minimum
of system (9) fails to be transformed easily to a matrix with eigenvalue of Qp will be no less than
some properties related to stochastic matrices, and therefore, 
the effective methods used in Jadbabaie et al. (2003) may not ¯ = (1 − 2 )k ¯ +  − [(1 − 2 )k ¯ − ]2 + 1 > 0, (17)
work. Here, we propose a Lyapunov-based approach to deal
with the problem. which is obtained by taking i (Mp )= ¯ with a given k satisfying
(11). In addition, since the eigenvalues of P are either min =
Theorem 4. For any fixed 0 <  < 1 and ¯ defined in (10), we 1 −  or max = 1 + , we have
take a constant
(1 − )2 V () (1 + )2 . (18)
1
k> . (11)
4(1 − 2 )¯ Therefore,

If the switching interconnection graph keeps connected, then T Qp  ¯

min
= 2,
T P  max
lim (t)C, (12)
t→∞ where  = ¯ /2(1 + ) > 0 with ¯ defined in (17).
¯ Moreover, if a(t) is known Due to (18),
for some constant C depending on .
¯
(i.e., a(t) = a0 (t) or  = 0), 1 
  √ V ().
1−
lim (t) = 0. (13)
t→∞ Therefore, from (15),

Proof. Take a Lyapunov function V () = T (t)P (t) with sym- (1 + )2 V () ¯
metric positive definite matrix V̇ ()|(9)  − 2V () + 2 
1−
 
In −In (1 + )2 ¯
2
P= ⊗ I2 . (14)  − V () +
−In In (1 − )
The interconnection graph is time-varying, but the interconnec- or equivalently,
tion graph associated with Fp for some p ∈ P is connected on
(1 + )2 ¯
2
an interval [ti , ti+1 ) with its topology unchanged. Consider the
V ((t)) V ((ti ))e−(t−ti ) + (1 − e−(t−ti ) ),
derivative of V (): (1 − )2
V̇ ()|(9) = T (FpT P + P F p ) + 2T P g t ∈ [ti , ti+1 ).

¯ Thus, with t0 = 0,
 − T Qp  + 2(1 + ), (15)
(1 + )2 ¯
2
−t
where V ((t)) V ((0))e + (1 − e−t ), (19)
  (1 − )2
2k(1 − 2 )M p −In
Qp = −(FpT P + P F p ) = ⊗ I2 (16) ¯ Fur-
−In 2In which implies (12) with taking C = ((1 + )/(1 − )).
¯
thermore, if  = 0, then (13) is obtained. 
is a positive definite matrix because 2I − (1/2k(1 − 2 ))Mp−1
and Mp are positive definite (by virtue of (11), Lemmas 1 Next, we consider an extended case: the interconnection
and 3). graph is not always connected. Let T > 0 be a (sufficient
Let i,j , i=1, . . . , n, j =1, 2 denote the (at most) 2n different large) constant, and then we have a sequence of interval
eigenvalues of Qp though Qp ∈ R 4n×4n defined in (16). Based [Tj , Tj +1 ), j = 0, 1, . . . with T0 = t0 , Tj +1 = Tj + T . Each
on i (Mp ), the eigenvalues of Mp , we have the 2n eigenvalues interval [Tj , Tj +1 ) consists of a number of intervals (still
 Y. Hong et al. / Automatica 42 (2006) 1177 – 1182 1181

If Tc >(T −T c ) or Tc >T /(+), then =e−T

c +(T −T c )
expressed in the form of [ti , ti+1 ), during which the intercon- <1.
nection graph is time-invariant), including the intervals during Thus,
which the graphs are connected and those during which the
+ · · · + 1)¯ ¯
j +1 2
graphs are not. We assume that there is a constant > 0, often V ((Tj +1 ))  V ((T0 )) + ( j

called dwell time, with ti+1 − ti
, ∀i. j +1 1 − j +1 ¯ 2

Denote the total length of the intervals associated with the  V ((T0 )) + ¯ .
1−
connected graphs as Tjc in [Tj , Tj +1 ) and the total length of the
intervals with the unconnected graphs as Tjd in [Tj , Tj +1 ). In For any t > 0, there is j such that Tj < t < Tj +1 with
what follows, we denote an upper bound of Tjd (j = 0, 1, . . .) V ((t)) eT V ((Tj )) + ¯ ¯ .
d 2

as T d (< T ), and a lower bound of Tjc (j = 0, 1, . . .) as

T c (=T − T d ). Thus, (12) is obtained with taking

(eT + 1 − )¯ ¯
d
Theorem 5. During each time interval [Tj , Tj +1 ), if the total
C= .
period that the interconnection graph is connected (i.e., T c ) (1 − )(1 − )
is sufficient large, then (12) still holds with k given in (11).
Moreover, (13) holds if ¯ = 0 (or equivalently a(t) = a0 (t)). Furthermore, if ¯ = 0, then C = 0, which implies (13), or  → 0
as t → ∞. 
Proof. Still take a Lyapunov function V () = T P  with P
defined in (14), and then we have (15). If the graph associated In fact, the proposed estimation idea can be extended to the
with Fp for some p ∈ P is connected during [ti , ti+1 ), then, case of an active leader with the following dynamics:
according to Theorem 4, we have ⎧ ẋ 1 = x 2 ,
⎪
⎪ 0 0
⎪
(1 + )2 ¯ 2 ⎨ ẋ0 = x0 ,
2 3
⎪
V ((ti+1 ))e−(ti+1 −ti ) V ((ti )) +  . ..
(1 − )2 . (22)
⎪
⎪
⎪
⎪
If the graph associated with Fq for some q ∈ P is not con- ⎩ ẋ0 = a(t) = a0 (t) + (t),
y = x0 = x01 ∈ R 2 ,
 during [tl , tl+1 ). The minimum eigenvalue of Qq is
nected
 − 1 + 2 (< 0) and, by (18), we have where y(t) is the measured output variable of the leader and
  a(t) is its input variable. The dynamics of each agent is still

− Qq  
T
1 +  −  T   V (),
2
taken in the form of (4). Then we will construct an observer as
2
 we did for system (5). Here, for the space limitations, we only
where  = (2 1 + 2 − 2)/(1 − ). give the corresponding error system, which can be expressed as
Similarly, with (18), ⎛ −kM ⎞
⎛ ˙1⎞ p In ⎛ 1⎞ ⎛ ⎞
x̄ x̄ 0
2(1 + )2 ¯ 2 ⎜ −1 kM p 0 In ⎟
V̇ ((t))|(9) V ((t)) +  , t ∈ [tl , tl+1 ), ⎜ x̄˙ ⎟ ⎜
2
⎟ ⎜ x̄ 2 ⎟ ⎜ 0 ⎟
(1 − ) ⎜ . ⎟=⎜ .. ⎟ ⊗I2 ⎜ . ⎟ + ⎜ . ⎟
⎝ .. ⎠ ⎜ . ⎟ ⎝ .. ⎠ ⎝ .. ⎠
⎝ ⎠
and therefore, ˙x̄ − −2 kM p 0 In x̄ −1⊗
− −1 kM p 0
2(1 + )2 T d
V ((tl+1 ))e(tl+1 −tl ) V ((tl )) + − 1)¯ . (20)
2
(e or equivalently in a compact form:
2 (1 − )
Denote ˙ = Fp  + g ∈ R 2n ,
 
(1 + )2 2(1 + )2 T d where k > 0 and 0 < j < 1, (j = 1, . . . , − 1) are suitable
= max , (e − 1) .
2 (1 − ) 2 (1 − ) real numbers and x̄ i = x i − 1 ⊗ x0i ∈ R 2n with x 1 = x and
x i (2 i  ) as the vector whose components are the respective
It is not hard to see that there are at most md = [T d / ] + 1 estimated values of x0i by n agents.
intervals (in [Tj , Tj +1 )) associated with unconnected graphs. To obtain the results similar to Theorems 4 or 5, we need to
Therefore, we have find a suitable quadratic Lyapunov function; that is, to construct
V ((Tj +1 ))e−Tj +Tj V ((Tj )) + (1 + eT + e2T
c d d d an appropriate positive definite matrix P such that FpT P +P F p
is negative definite then the corresponding graph Ḡp is con-
+ · · · + emd T ) ¯
d 2
nected. For example, when = 3, we can choose
e−T V ((Tj )) + ¯ ¯
c +(T −T c ) 2
(21) ⎛ −kM ⎞
p In 0 ⎛ ⎞
In − 23 I 0
with ⎜ 8k ⎟
Fp =⎜ − Mp 0 In ⎟ ⊗I2 , P =⎝− 2 I In − 21 In⎠⊗I2 .
e(md +1)T − 1
d ⎝ 9 ⎠ 3 n
¯= > 0. 4k
− Mp 0 0 0 − 21 In In
eT − 1
d
9
1182 Y. Hong et al. / Automatica 42 (2006) 1177 – 1182

4. Conclusions Olfati-Saber, R., & Murray, R. M. (2004). Consensus problems in networks

of agents with switching topology and time-delays. IEEE Transactions on
Automatic Control, 49(9), 1520–1533.
This paper studied the consensus problem of a group of au- Savkin, A. (2004). Coordinated collective motion of groups of autonomous
tonomous agents with an active leader, whose velocity can- mobile robots: Analysis of Vicsek’s model. IEEE Transactions on
not be measured. To solve the problem, a distributed feedback Automatic Control, 49(6), 981–983.
(i.e., (6)) along with a distributed state-estimation rule (i.e.,
(7)) was proposed for each continuous-time dynamical agent, Yiguang Hong received his B.S. and M.S. de-
and Lyapunov-based convergence analysis was given for the grees from Department of Mechanics, Peking
University, China, and his Ph.D. degree from
considered multi-agent system with a varying interconnection Chinese Academy of Sciences (CAS). He
topology. is currently a professor in Institute of Sys-
tems Science, Academy of Mathematics and
Systems Science, CAS. He is a recipient of
Acknowledgements Guangzhaozhi Award of Chinese Control Con-
ference (in 1997), Young Author Prize of IFAC
World Congress (in 1999), and Young Scientist
The authors wish to thank the reviewers for their construc- Award of CAS (in 2001). His research interests
tive suggestions, especially the reviewer who introduced us a include nonlinear dynamics and control, robot control, multi-agent systems,
and reliability of software systems.
framework of distributed observer design for multi-agent sys-
tems. This work was supported by the NNSF of China under Jiangping Hu received his B.S. degree in ap-
Grants 60425307, 50595411 and 60221301. plied mathematics and his M.S. degree in com-
putational mathematics from Lanzhou Univer-
sity, China, in 2000 and 2004, respectively. He
References is currently a doctoral student in Institute of
Systems Science, Chinese Academy of Sciences.
His research interests include multi-agent sys-
Fax, A., & Murray, R. M. (2004). Information flow and cooperative control tems and numerical analysis.
of vehicle formations. IEEE Transactions on Automatic Control, 49(9),
1453–1464.
Godsil, C., & Royle, G. (2001). Algebraic graph theory. New York: Springer.
Horn, R., & Johnson, C. (1985). Matrix analysis. New York: Cambridge Lixin Gao received his M.S. degree in math-
University Press. ematics from East China Normal University in
Jadbabaie, A., Lin, J., & Morse, A. S. (2003). Coordination of groups of 1994, and his Ph.D. degree in Automatic Control
mobile autonomous agents using nearest neighbor rules. IEEE Transactions from Zhejiang University of China in 2003. He
has been with School of Mathematics and Infor-
on Automatic Control, 48(6), 998–1001.
mation Science, Wenzhou University of China
Lin, Z., Broucke, M., & Francis, B. (2004). Local control strategies for groups since 1994. He is also a postdoctoral fellow in
of mobile autonomous agents. IEEE Transactions on Automatic Control, Institute of Systems Science, Chinese Academy
49(4), 622–629. of Sciences from 2004. His research interests
Merris, R. (1994). Laplacian matrices of graphs: A survey. Linear Algebra include robust control, multi-agent systems, and
and Applications, 197, 143–176. numerical analysis.