IET Control Theory & Applications

Research Article

ISSN 1751-8644
Consensus for multi-agent systems with Received on 29th November 2015
Revised on 2nd March 2016
distributed adaptive control and an Accepted on 16th March 2016
doi: 10.1049/iet-cta.2015.1221
event-triggered communication strategy www.ietdl.org

Duosi Xie1 , Shengyuan Xu1, , Baoyong Zhang1 ,Yongmin Li2 ,Yuming Chu2
1
School of Automation, Nanjing University of Science andTechnology, Nanjing 210094, People’s Republic of China
2
School of Science, HuzhouTeachers College, Huzhou 313000, People’s Republic of China
E-mail: [email protected]

Abstract: This study addresses the consensus problem of multi-agent systems by combining adaptive control protocol
and event-triggered communication strategy. Two adaptive protocols are designed to guarantee the consensus. The event-
triggered communication strategy is utilised to save communication and reduce channel occupation. Each agent in the
system sends its state to neighbours only at some irregular event instants determined by an appropriately designed
event-triggering condition. At first, the distributed event-triggering condition of an agent only employs the state of its own
and the state of the neighbours at their latest event instants. Later, another event-triggering condition is given such that
no global parameter is required for agents, such as the maximum eigenvalue of the Laplacian matrix. It is proven that
the system can reach consensus by utilising the proposed control method. Illustrative examples are given to show the
effectiveness of the control strategy.

1 Introduction points and uses the sampled state to calculate the event-triggering
condition. If triggered, each agent sends the sampled state to its
Consensus problem is an attractive issue in the research of collec- neighbours, otherwise the agent samples its own state at the next
tive behaviour of multi-agent systems [1, 2]. Many researchers have sampling point.
investigated this problem and gave numerous contributions [3–9]. Some other works in which the event-triggered strategy has been
Multi-agent systems involve a large quantity of agents that are adopted to solve the consensus problem for the case of switching
usually geographically distributed in a wide range. Meanwhile, the network topologies are also considered [29, 30]. It is shown in [29]
agents communicate with each other through a wireless network. that a sufficient condition to guarantee the event-based consensus
Hence considering the physical realities that agents only have the is to have the topology being balanced and strongly connected.
access to their neighbours, distributed consensus control methods The latter part of this sufficient condition is further relaxed in
are required. An efficient approach under such situations is the [30] such that the topology requires to be only jointly connected.
adaptive control strategy, which can adjust the control gain through Overall, it is widely accepted that the event-triggered strategy can
neighbours’ information [10–13]. Moreover, the channel occupa- be implemented into solving the consensus problem, while reduc-
tion might become another issue as the communication channels ing communication times by trading the convergence rate of the
usually have bandwidth limitations. The countermeasures for these system.
limitations are considered in the study of sampled-data control In this paper, an innovative event-triggered adaptive control
strategy, see [14–17] and the reference therein. protocol is proposed to solve the consensus problem in multi-
The event-triggered control strategy, as a special case of the agent systems. The proposed control strategy combines dynamical
sampled-data strategy, has varying sampling rates. The idea of control parameter and the event-triggered communication scheme.
this aperiodic scheme is that the communication among agents The dynamical adaptive parameters are designed compatible to
happens only at irregular event instants [18–22]. The event instants the consensus convergence of the system to provide agents more
are calculated by a predefined event-triggering condition, i.e. flexible control gains. Other than using continuous states of neigh-
communication takes place when this condition is violated. The bours’, the adaptive parameter and the event-triggered control law
event-triggering condition is first proposed in a centralised way of an agent only involve the irregular states of neighbours that
and then extended to decentralised models [23, 24]. In a general are sampled and sent at their latest event instants. Hence the com-
event-triggered communication protocol, each agent samples and munication times can be reduced by applying this event-triggered
sends its own state to its neighbours at its event instants. Mean- scheme. The irregular event instants are generated by a distributed
while, each agent receives states that are sampled and sent by its event-triggering condition, which is state dependent. The state-
neighbours at their event instants. dependent strategy makes the condition relate to the convergence
Based on the aforementioned general scenario, lots of related states of the agents.
works have been done. Different models such as the second-order Moreover, the event-triggering condition does not require con-
system [25] and the general linear model [26] are considered tinuous states of neighbours either, instead it utilises the agent’s
under such communication strategy. Beside the models, various state at the current time and its latest event time as well as its
event-triggering protocols had also been developed. In [26], an neighbours’ states at their latest event instants. Hence no additional
event-triggered communication protocol, where each agent sam- communication is generated in the event calculating process.
ples both its own and neighbours’ information at its own triggering In addition, a distributed event-triggering condition is further
instants, is proposed. A different event-triggering strategy, where proposed such that each agent does not need to have the global
the agents have synchronous sampling points with a fixed time information of the system like the maximum eigenvalue of the
interval, is first discussed in [27] and then extended in [28]. In their Laplacian matrix. Instead, the agents only have the knowledge
work, each agent samples its state at the synchronous sampling of its own in-degree. This distributed event-triggering condition

IET Control Theory Appl., 2016, Vol. 10, Iss. 13, pp. 1547–1555
© The Institution of Engineering and Technology 2016 1547
 17518652, 2016, 13, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-cta.2015.1221 by CAPES, Wiley Online Library on [09/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
makes the proposed strategy fully localised and hence it is more where δij are positive constants, ∀ i = j, i, j ∈ {1, . . . , N }. This
compatible with the widely distributed multi-agent systems. Last adaptive protocol is able to adjust the weights between connected
but not least, the inter-event intervals are analysed in this work to agents and therefore quickly drive the system to achieve consen-
discuss the Zeno behaviour (the accumulation of events). sus. The proposed protocol requires the data exchanges among the
The paper is organised as follows: Section 2 introduces some agents continuously. In this paper, we consider a strategy which
preliminaries that are used in this paper. The consensus prob- reduces the communication times while guaranteing the consensus.
lem and general adaptive control protocol are also stated in
Section 2. Section 3 addresses the main results of this paper, i.e. the
event-triggered communication strategy with two different adap- 3 Main results
tive protocols for multi-agent systems. Section 4 illustrates the
effectiveness of the proposed control strategy by given simulation In this section, the main results of the consensus control of the
examples. Finally, a conclusion is presented in Section 5. multi-agent system (1) are addressed. Associating with the event-
triggered communication strategy, two different adaptive control
Notation. Throughout this paper, xT and x denote the transpose protocols are proposed. The two adaptive protocols are the edge-
and the Euclidean norm of vector x, respectively. For a matrix A, based adaptive protocol and the node-based adaptive protocol. The
AT denotes its transpose. A represents the two-norm of A. |c| former protocol assigns a time-varying weight to each edge in
denotes the absolute value of a scalar c. 1N ∈ RN is a column the system topology. The latter protocol assigns a time-varying
vector with all the entries being 1. weight to each node in the system. Later, a fully distributed
event-triggering condition is proposed to avoid using the maxi-
mum eigenvalue of the Laplacian matrix. Therefore, the agents in
2 Preliminaries
the system do not need to have knowledge of the topology of the
system.
2.1 Graph theory

The network topology of a multi-agent system can be described by 3.1 Consensus with an edge-based event-triggered
a graph G(V, E), where V = {1, 2, . . . , N } and E ⊆ V × V are the adaptive control strategy
sets of nodes and edges, respectively.
An agent j is said to be a neighbour of agent i if it is able We first consider the approach of combining the event-triggering
to send information to agent i. This directed communication link strategy with an edge-based adaptive control protocol for solving
is denoted as (i, j) ∈ E. The set of all neighbours of agent i is the consensus problem in the multi-agent system. Our control input
denoted by Ni = {j|(i, j) ∈ E}. A topology is undirected if (i, j) ∈ E for each agent is designed as
implies (j, i) ∈ E. A path from agent i to agent k in G(V, E) is
a sequence of ordered edges {(i, i + 1), . . . , (k − 1, k)} such that 
N
 
(j, j + 1) ∈ E, ∀ j = 1, . . . , k − 1. An undirected graph is connected ui (t) = − cij (t)aij x̃i (t) − x̃j (t) , (5)
if there exist paths from every agent to every other agents. An adja- j=1
cency matrix A = [aij ] ∈ RN ×N is usually introduced to describe  
the graph topology, where the diagonal entries of A are aii = 0, and ċij (t) = δij aij xi (t) x̃i (t) − x̃j (t) . (6)
the off-diagonal entries are determined by aij = 1, if (j, i) ∈ E and
aij = 0 otherwise. The in-degree matrix  of the system is defined as Here x̃i (t)  xi (tki ), ∀t ∈ [tki , tk+1
i ), where tki denotes kth event
D = diag{d1 , . . . , dN }, where di = j∈Ni aij . Then the Laplacian instant of agent i. Every time agent j generates an event, x̃j (t) will
matrix L is defined as L = D − A. According to the definition, it is be updated and transmitted to agent i. It is observed that agent i’s
straightforward to find out L1N = 0. Moreover, if the topology is control input ui (t) depends on both x̃i (t) and x̃j (t), ∀j ∈ Ni . There-
undirected, A and L are both symmetric. It follows that 1TN L = 0. fore ui (t) will be updated at both the triggering instants of its own
The following lemma states some well-known properties of the and its neighbours.
Laplacian matrix that will be used later on in this paper. The event-triggered communication pattern is described as fol-
lows: each agent will be triggered if its own event-triggering
Lemma 1 [31]: If the undirected graph G of a multi-agent sys- condition is violated. At the event instant of each agent, say agent
tem is connected, then the Laplacian matrix of the graph has one i, its own state xi (tki ), i.e. x̃i (t) will be used to update its control
zero eigenvalue and all the other eigenvalues are positive. By input ui (t), it will also be sent to its neighbours. Meanwhile, agent
using λ1 and λN to denote the smallest and the largest eigenval- i will also receive the state information of all its neighbours (agent
ues of L, respectively, ∀i = 1, . . . , N , then one has 0 = λ1 < λN . j, ∀j ∈ Ni ) whenever any of them is triggered. The received infor-
Furthermore, for any vector x ∈ RN , one has xT Lx ≤ λN xT x. mation from its neighbours will also be used to update its control
input ui (t). Generally speaking, the event-triggering condition of
2.2 Adaptive control for consensus problems each agent (say agent i) involves the state of both its own and its
neighbours at their latest event instant. Before stating the event-
The dynamic of each agent in a distributed multi-agent system is triggering condition for each agent, we first define a measurement
described by error ei (t) for each agent as ei (t)  x̃i (t) − xi (t), ∀i = 1, . . . , N ,
ẋi (t) = ui (t), i = 1, . . . , N , (1) i.e. the state error between the current time t and the latest event
time tki . It is clear that by definition e(tki ) = 0. Therefore, with the
where xi (t) ∈ R and ui (t) ∈ R are the state and control inputs
ei (t) defined, the system dynamics can be written as
of agent i, respectively. The consensus of the multi-agent system
described in (1) is achieved if for any initial condition   
ẋi (t) = − aij cij (t) xi (t) − xj (t) + ei (t) − ej (t) , (7)
lim xi (t) − xj (t) = 0, ∀ i, j ∈ {1, . . . , N }. (2) j∈Ni
t→∞
 
ċij (t) = δij aij xi (t) xi (t) − xj (t) + ei (t) − ej (t) . (8)
In [11, 32], a distributed edge-based adaptive control proto-
col was proposed to solve the consensus problem. The proposed
controller of each agent is given by The distributed event-triggering condition fi (t) for each agent is
 described as
 
ui (t) = − cij (t) xi (t) − xj (t) , (3)  
 
j∈Ni γ    
,
fi (t) : |ei (t)| ≤ x̃ (t) − x̃ (t) (9)
βi  
i j
ċij (t) = δij (xi (t) − xj (t))2 , (4) j=Ni 

IET Control Theory Appl., 2016, Vol. 10, Iss. 13, pp. 1547–1555
1548 © The Institution of Engineering and Technology 2016
 17518652, 2016, 13, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-cta.2015.1221 by CAPES, Wiley Online Library on [09/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License

where γ and βi are both positive constants. The event-triggering Since γ /λN j∈Ni (x̃i (t) − x̃j (t)) is the ith element of vector
condition fi (t) determines the event instants of agent i through the (γ /λN )Lx̃(t). By using 2xT (t)LLe(t) ≤ xT (t)LLx(t) + eT (t)LLe(t),
following procedure: at event instant tki , the event-triggering con- one has
dition fi (t) holds since ei (tki ) = 0 by definition. After tki , |ei (t)|
will fluctuate
 until agent i is triggered when |ei (t)| exceeds γ2 γ2
(γ /βi )| j∈Ni (x̃i (t) − x̃j (t))|. The current time instant is defined as eT (t)e(t) ≤ 2
Lx̃(t)2 = 2 (x(t) + e(t))T LL (x(t) + e(t))
λN λN
the new event instant tk+1 i i
. Since the measurement error ei (tk+1 )=
0, the event-triggering condition holds again until the next event 2γ 2 T
≤ x (t)LLx(t) + eT (t)LLe(t)
instant. Note that for a certain agent i, during the time intervals λ2N
between its two consecutive event instants, if any of its neigh-
bours (j, ∀j ∈ Ni ) is triggered, its event-triggering condition will 2γ 2 T
≤ x (t)LLx(t) + 2γ 2 eT (t)e(t). (16)
be updated with the newly received neighbour state x̃j (t), ∀j ∈ Ni . λ2N
In other words, for agent i, the right-hand side of (9) may be
time-varying during time intervals between every two of its adja- Since the positive parameter γ satisfies the condition (11), it
cent event instants. The potential time-varying property may affect follows that 1 − 2γ 2 > 0. Hence one has
the length of the time intervals between every pair of consecutive
event instants. More detailed analysis will be discussed later in this
2γ 2
paper. eT (t)e(t) ≤ xT (t)LLx(t). (17)
Before addressing the main theorem, we specify some notations λ2N (1 − 2γ 2 )
used later on. x(t) = [x1 (t), . . . , xN (t)]T and e(t) = [e1 (t), . . . ,
eN (t)]T are used to denote the collection of the states and measure- As a result of that, the derivative of the Lyapunov function has
ments of all agents, respectively. It follows that e(t) = x̃(t) − x(t),
where x̃(t) = [x̃1 (t), . . . , x̃N (t)]T . 1 γ2
V̇ ≤ θ − + xT LLx. (18)
2λN λN (1 − 2γ 2 )
Theorem 1: Consider the multi-agent system descried in (1) under
a connected topology. Suppose that the system is controlled by
the adaptive protocols (5), (6) and the actuator of agent i is trig- The condition (11) makes (−1/2λN + γ 2 /λN (1 − 2γ 2 )) < 0.
gered when the event-triggering condition (9) is violated. Then the Hence V̇ (t) ≤ 0 and it equals to 0 only when the consensus is
consensus can be achieved if the following conditions are satisfied achieved. From the LaSalle invariance principle, the consensus of
the system (1) can be achieved under such control strategy. 
βi ≥ λN , (10)
It is worth mentioning that the Lyapunov function (12) is dif-
1
0<γ < . (11) ferent from the commonly used Lyapunov  function
N [10], which
2 is defined as V (t) = 1/2xT (t)Lx(t) + N j=1 (cij − θ ) /2δij .
2
i=1
Proof: We consider the following Lyapunov candidate for system By using the common Lyapunov function, the consensus of the
(1) associated with (5) and (6) system is reached if one can ensure that the derivative of the
Lyapunovfunction is negative definite. The Lyapunov function
N  2 (12) uses N 2 T
1 2 
N 
i=1 1/2xi (t) instead of 1/2x (t)Lx(t). Under this case,
N
cij − θ
V (t) = xi (t) + , (12) the LaSalle invariance principle is introduced to analyse the con-
2 2δij vergence of the agents. It is proven that under the connected
i=1 i=1 j=1
topology, the agents will converge to the consensus space, which
where θ is a positive constant. It is clear that V (t) is positive is x1 (t) = x2 (t) = · · · = xN (t) exactly.
definite. The right derivative of V (t) along the trajectory of (1),
(5) and (6) with respect to time is Remark 1: Adaptive control is a very useful method when dealing
with complex dynamics, such as non-linear dynamics. It assigns

N 
N
  dynamical control gains compatible to the convergence of the sys-
V̇ (t) = − xi (t) cij (t)aij x̃i (t) − x̃j (t) tem. When agents diverge, the adaptive parameters will become
i=1 j=1 larger and when agents converge, the adaptive parameters will be

N 
N
    smaller. Hence it can adjust the convergence speed of the system.
+ cij (t) − θ xi (t)aij x̃i (t) − x̃j (t) The event-triggered strategy can reduce communication times, but
i=1 j=1 it will sacrifice some performance, such as the convergence speed.
In this paper, we consider the adaptive control together with

N 
N
  the event-triggered strategy under a basic multi-agent system as
= −θ xi (t)aij x̃i (t) − x̃j (t) . (13) the first step of the research of the event-triggered adaptive con-
i=1 j=1 trol method. By combing these two methods, the control strategy
can save communication while dynamically adjust the control gain
From Lemma 1, it is obtained that xT (t)LLx(t) = (L1/2 x(t))T L(L1/2 according to the convergence.
x(t)) ≤ λN xT (t)Lx(t). Furthermore, it is clear that xT (t)Le(t) ≤
(1/(2λN ))xT (t)LLx(t) + (λN /2)eT (t)e(t). Therefore, the right- It is well known that the existence of Zeno behaviour must be
hand side of (13) turns into excluded in the work of event-triggered communication strategies
as the initial motivation of such strategies is to reduce the frequency
V̇ (t) = −θxT (t)Lx(t) − θ xT (t)Le(t) of communications. In other words, for each agent i, ∀i = 1, . . . , N
1 λN T in this system, the time interval between every two consecutive
≤θ − xT (t)LLx(t) + e (t)e(t) . (14) event instants tki and tk+1
i must be strictly positive. We propose
2λN 2
the following theorem to exclude Zeno behaviour in our work.
From condition (10) and the event-triggering condition (9)
holding at the right-hand side of any instants, we can obtain Theorem 2: Consider the multi-agent system descried in (1) under
  a connected topology. Suppose that the system is controlled by the
 
γ    
.
adaptive protocol (5), (6) and the actuator of agent i is triggered
|ei (t)| ≤ x̃ (t) − x̃ (t) (15) when the event-triggering condition (9) is violated. If the consen-
λN  
i j
j∈Ni  sus is not achieved yet and the agent is not located at the average

IET Control Theory Appl., 2016, Vol. 10, Iss. 13, pp. 1547–1555
© The Institution of Engineering and Technology 2016 1549
 17518652, 2016, 13, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-cta.2015.1221 by CAPES, Wiley Online Library on [09/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
position of its neighbours, then the event intervals are strictly 3.2 Consensus with a node-based event-triggered
positive. adaptive control strategy

In this part, a node-based adaptive control protocol inspired by

Proof: The event intervals can be obtained by analysing the motion
Li et al. [11] is proposed. Being different from the edge-based
of ei (t). For each agent, say agent i, at each event instant, say tki ,
adaptive control protocol, the node-based adaptive control protocol
one has ei (tki ) = 0. From the event-triggering condition, the (k + adopts a time-varying weight ci (t) for each agent of the system, as
 instant is defined as the first time when |ei (t)| exceeds
1)th time described in the following equation
γ /βi | j∈Ni (x̃i (t) − x̃j (t))|. Thus, the time interval between the
kth and the (k + 1)th eventsis no less than the time for |ei (t)|
to increase from 0 to γ /βi | j∈Ni (x̃i (t) − x̃j (t))|. Two cases are 
N
 
ċi (t) = δi xi (t) aij x̃i (t) − x̃j (t) , (22)
considered separately during each interval.
j=1
One case is the case in which none of the neighbours of agent i
is triggered during the interval between the kth and the (k + 1)th
events. In this case, the right-hand side of the triggering condition where δi , i = 1, . . . , N are positive constants. The control input of
(9) is time invariant during this period. Suppose that agent i is agent i, whose dynamic is described in (1), under the node-based
triggered after time tki + τi . Then one can obtain that case is given by

  
N
 
 i   tki +τ  ui (t) = −ci (t) aij x̃i (t) − x̃j (t) .
ei (t + τi ) =  
ėi (t)dt 
(23)
k  tki  j=1
 
 t i +τ  
 k i    From (22), one can see that the coupling weight ci (t) of agent
=  cij (t) x̃i (t) − x̃j (t) dt  . (19) i is depending on the states of all its neighbours. The consen-
 tki j∈Ni  sus achievement under this node-based control protocol can be
analysed by the following theorem.
 N N √
j=1 (cij − θ ) /2δij ≤ V (0), |cij | ≤ 2δM V (0) + θ,
Since 2
i=1
where δM = max{δ Theorem 3: Consider the multi-agent system as descried in (1)
√ ij |i = j, i, j ∈ V}. In the rest of this paper, we under a connected topology. Suppose that the adaptive protocols
use η to denote 2δM V (0) + θ for simplicity. It follows that
(22) and (23) are adopted. Meanwhile, the events occur when the
  event-triggering condition (9) is violated. Then the consensus is
 t i +τ  
 i   k i    achieved if the conditions (10) and (11) are satisfied. Moreover,
ei (t + τi ) ≤ η  x̃i (t) − x̃j (t) dt 
k  i Zeno behaviour is excluded if the agent is not located at the average
 tk j∈Ni  position among its neighbours.
  
≤ ητi  x̃i (t) − x̃j (t)  . (20)
Proof: Consider the following function as a Lyapunov candidate:
j∈Ni

1 2  (ci (t) − θ)2

N N
By the event-triggering condition (9), the agent will be triggered V (t) = xi (t) + , (24)
after the following equation holds 2 2δi
i=1 i=1
 
    where θ is a positive constant as defined in previous sections. It is
 γ   
ητi  x̃i (t) − x̃j (t)  =  x̃ (t) − x̃ (t) . (21) straightforward to obtain that V (t) is positive definite. It follows
βi  
i j
j∈Ni j∈Ni  that

The right-hand side of (21) is the sum of the state difference among 
N 
N
 
V̇ (t) = − xi (t)ci (t) aij x̃i (t) − x̃j (t)
agent i and its neighbours. There are two cases that this term equals
i=1 j=1
to zero. The first one is that x̃i (t) = x̃j (t), ∀j ∈ Nj . If the consensus
is not reached, then this scenario will not occur. The second one is 
N 
N
 
that di x̃i (t) = j∈Ni x̃j (t). In other words, agent i is located at the + (ci (t) − θ) xi (t) aij x̃i (t) − x̃j (t)
average position of its neighbours. Therefore, if the consensus is i=1 j=1
not achieved and agent i is not specifically  located at the average
position of all its neighbours, one has | j∈Ni (x̃i (t) − x̃j (t))| > 0. It 
N 
N
 
= −θ xi (t)aij x̃i (t) − x̃j (t) . (25)
follows that τi > 0. Hence the interval between every two events
i=1 j=1
of agent i is positive.
The other case is when there is at least one neighbour of agent
i’s being triggered during the event interval of agent i. Suppose Therefore, via a similar approach as that of Theorem 1, it is easily
j shown that V̇ (t) ≤ 0 and V̇ (t) → 0 as t → ∞. Since the graph is
that agent j, as a neighbour of agent i, is triggered after tki , say tk ∗ .
j connected, it follows that xi (t) = xj (t) ∀i, j ∈ {1, . . . , N } as t → ∞.
Then it is guaranteed that tk ∗ − tki > 0. Hence the interval between This implies that the consensus of the system can be achieved
the kth and the (k + 1)th events of agent i must be greater than under the proposed event-triggered strategy. The exclusion of Zeno
j
tk ∗ − tki , which is a positive number. It follows that Zeno behaviour behaviour can be shown via the same approach as Theorem 2. We
is excluded since there are no infinite events in a finite time period omit it here for the interest of space. 
during the consensus achievement. This concludes the proof. 
It is worth mentioning that, compared to the edge-based adaptive
Theorem 1 proves that the event-triggered adaptive control strat- control protocol, the node-based adaptive control protocol requires
egy can guarantee the consensus of the multi-agent system. In less storage spaces for agents. For a system that consists of N
addition, Theorem 2 states that the communication times and (N > 2) agents, there are at least N − 1 edges if the topology is
channel occupation during the convergence period can be reduced. connected. On one hand, the node-based adaptive protocol requires
Therefore, the proposed strategy is very suitable for multi-agent N units of storage spaces for the coupling weights ci , i = 1, . . . , N .
systems, especially for those which have number of agents and On the other hand, since the topology is connected and cij = cji ,
wide distribution ranges. the edge-based adaptive protocol requires at least 2(N − 1) units

IET Control Theory Appl., 2016, Vol. 10, Iss. 13, pp. 1547–1555
1550 © The Institution of Engineering and Technology 2016
 17518652, 2016, 13, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-cta.2015.1221 by CAPES, Wiley Online Library on [09/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
of storage spaces to record cij . Therefore, the number of edges 1  
N N
2   e2
N N
is greater than the number of agents especially for large-scale = − (1 − k1 ) aij x̃i − x̃j + aij i
network. Hence the node-based protocol requires less spaces for 2 2k1
i=1 j=1 i=1 j=1
coupling weight storage compared to the edge-based protocol.
1  
N N
2  di 2 N
= − (1 − k1 ) aij x̃i − x̃j + e .
2 2k1 i
i=1 j=1 i=1
Remark 2: In the dynamic of the adaptive parameters (6) and (22),
both xi (t) and x̃i (t) are utilised for agent i. On one hand, acquiring
agent i’s own state does not need communication through chan-
nel. Therefore the assumption that agent i always has access to its Since an event of agent i occurs when the event-triggering condi-
own state will not increase the communication among agents. On tion (26) is violated, one can obtain that the time derivative of the
the other hand, the adaptive parameters cij (t) and ci (t) are updat- Lyapunov function is negative when the consensus is not achieved
ing continuously anyway. It follows that using xi (t) in the control and it equals to zero when the consensus is achieved, i.e. V̇ (t) ≤ 0
input will not increase the communication nor updating costs. In and V̇ (t) → 0 as t → ∞. Therefore, similar to the former results, it
this paper, the main focus is to reduce the communication among follows that xi (t) = xj (t) ∀i, j ∈ {1, . . . , N } as t → ∞. This implies
agents by introducing the event-triggered adaptive control method. that the consensus of the system can be achieved under the control
Therefore using xi (t) is compatible with this purpose. of strategies (5), (6) and the event-triggering condition (26).
Now we will show that the inter-event interval is positive when
the consensus is not reached. As the dynamic of the measurement
3.3 Consensus with fully distributed event-triggering error ei (t) remains the same, we can borrow the dynamic analysis
condition of ei (t) in Theorem 2. As shown in (21), by the fully distributed
event-triggering condition (26), the agent will be triggered after the
In the previous section, both the edge-based and the node-based following equation holds
event-triggered adaptive control strategies were proposed. Although
  
 
 
these two strategies have distributed property, they still requires
   k1 − k12   2 
some global information such as the maximum eigenvalue of ητi  
x̃i (t) − x̃j (t) =   x̃ (t) − x̃ (t) . (27)
 i j 
the Laplacian matrix. Furthermore, when the agent is located at j∈Ni
di j∈Ni 
the average position among its neighbours, the inter-event inter-
val cannot be guaranteed to be positive even when consensus is Different from the former results, both the left-hand side and
not achieved. In this section, an fully distributed event-triggering the right-hand side of the above equation are positive when the
condition is designed. By using this improved event-triggering con- consensus is not reached. It follows that τi > 0 if the consensus is
dition, each agent no longer needs to know the eigenvalue of not reached.
the Laplacian matrix, instead, the in-degree of the agent is used. This newly designed event-triggering condition can also be
Second, as long as the consensus is not achieved, the inter-event applied by combing the node-based control inputs (22) and (23).
interval is positive. The proof of the convergence and positive inter-event interval are
The newly designed event-triggering condition is that the same with this scenario. Hence we omit the proof to save the
space here. 
  
  
 k1 − k12   2  Remark 3: The irregular communications among agents make the
fi (t) : |ei (t)| ≤   x̃ (t) − x̃ (t) , (26)
 i j  agents has less knowledge about the most current state of their
di j∈Ni  neighbours and therefore will scarify the convergence performance.
However, it is very hard to give theoretical result to prove it. In
the following section, numerical simulations are given to show that
where k1 is a positive constant such that 0 < k1 < 1 and di is the the event-triggered communication strategy can drive the system to
in-degree of agent i. reach consensus while only a few communications occur.

4 Simulations
Theorem 4: Consider the multi-agent system as descried in (1)
under a connected topology. Suppose that the adaptive protocols In this section, simulations are shown to illustrate the efficiency
(5) and (6) are adopted. Meanwhile, the events occur when the of the proposed adaptive protocols and the event-triggered com-
event-triggering condition (26) is violated. Then the consensus of munication strategy. We first give simulation for the edge-based
the system can be achieved and the inter-event intervals are positive adaptive scenario and then give another simulation for the node-
when the consensus is not achieved. based adaptive scenario. After that, a simulation using the fully
distributed event-triggering condition is also given. At last, the
event times under these event-triggered adaptive protocols are
Proof: The same Lyapunov candidate is considered in this scenario. listed among three different topologies. All of the simulations are
From (13) one has implemented to a six-agent multi-agent system whose dynamics
are described in (1).

V̇ ≤ −xT Lx̃ = − (x̃ − e)T Lx̃ = −x̃T Lx̃ + eT Lx̃ Example 1: In this example, the edge-based adaptive control
protocols (5) and (6) are adopted. The communication instants are
1 
N 
N
 2 
N 
N
  determined by the event-triggering condition (9). We set δij = 1,
=− aij x̃i − x̃j + aij ei x̃i − x̃j
2 for i = j, where i, j ∈ {1, . . . , 6}. The used Laplacian matrix for the
i=1 j=1 i=1 j=1 system is
1 
N 
N
 2 ⎡ ⎤
≤− aij x̃i − x̃j 1 0 −1 0 0 0
2 ⎢0 3 −1 0 −1 −1⎥
i=1 j=1 ⎢ ⎥
⎢−1 −1 3 −1 0 0⎥
L1 = ⎢
0 −1⎥
.

N 
N
1 2 k1  2 ⎢0 0 −1 2 ⎥
+ aij ei + x̃i − x̃j ⎣ 0 −1 0 0 1 0⎦
2k1 2 0 −1 0 −1 0 2
i=1 j=1

IET Control Theory Appl., 2016, Vol. 10, Iss. 13, pp. 1547–1555
© The Institution of Engineering and Technology 2016 1551
 17518652, 2016, 13, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-cta.2015.1221 by CAPES, Wiley Online Library on [09/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
As the maximum eigenvalue of the given Laplacian matrix is the adaptive parameters will be smaller. Hence the parameters do
λN = 4.8136, we choose βi = λN , ∀i = 1, . . . , 6 and γ = 0.45 so not have monotonical property. This can be seen in Fig. 1d.
that the conditions (10) and (11) are satisfied. Fig. 1a shows the
states trajectories of agents in the system. It is observed that the Example 2: Now we use the node-based adaptive protocols (22)
system reaches consensus successfully. Fig. 1b illustrates the event and (23) to control the multi-agent system while the event-
instants for each of the agents. It is shown that each agent trig- triggering condition (9) is also adopted to determine the events. We
gers asynchronously. Moreover, the event-triggered communication use the same Laplacian matrix L1 and set δi , βi , ∀i = 1, . . . , N , γ
significantly saves communication times. In order to inspect the as the same value as the previous example to satisfy conditions
motion the measurement error, the evolution of |e1 (t)| is shown (10) and (11)
in Fig. 1c. In Fig. 1c, it is observed that |e1 (t)| is zero at event
instants and it grows until the event-triggering condition (9) is Fig. 2 illustrates the dynamics of the agents, the event sequences
violated. Fig. 1d illustrates the evolution of the adaptive parameter of agents and the motion of agent 1, respectively. It is observed
of agent 6, i.e. c62 and c64 . It shows that these parameters are not that the event-triggered communication can also save communica-
monotonically increasing or decreasing. tion times significantly. From Figs. 1b and 2b, the event sequences
that generated by the node-based strategy appear to have lot more
Remark 4: Owing to the limitation of resolution of the simulation events than the edge-based strategy.
figures, when it seems like the consensus is achieved in Fig. 1a,
actually there still has some slight difference between the states Example 3: Now we give the simulations utilising the fully dis-
of agents (the difference is 0.0001 for instance). Hence, in this tributed event-triggering condition (26) together with control inputs
situation, an event may occur if the event-triggering condition is (5), (6) and control inputs (22), (23), by using the same Laplacian
violated and therefore an event mark appears on the figure. matrix and the set k1 = 0.6. Similar to the previous two cases,
In Fig. 1c, it seems that the evolution |e1 (t)| penetrates the Figs. 3 and 4 illustrate the dynamics of the agents, the event
threshold. This is because the simulation is running on discrete sequences of agents and the motion of agent 6, respectively, under
simulation software to imitate continuous system. Hence |e1 (t)| edge-based fully distributed event-triggered control strategies (5),
might even penetrate the threshold between time step and thereby (6) and node-based fully distributed event-triggered strategies (22),
shows in the figure. In real continuous system, the measurement (23). One can observe that an event occurs when the norm of the
error will not penetrate its threshold. measurement error overpasses its threshold and the system is able
to reach consensus.
Remark 5: The dynamic control parameters fluctuate along with
the convergence situation of the system. When agents diverge, the From these simulations, one can also see that the node-based
adaptive parameters will become larger and when agents converge, strategy generates more events than the edge-based strategy.

a b

c d

Fig. 1 Edge-based adaptive event-triggered strategy

a State of agent i for i = 1, . . . , 6
b Event instant sequence of each agent
c Norm of the measurement error of agent 1
d Motion of the adaptive parameters of agent 6

IET Control Theory Appl., 2016, Vol. 10, Iss. 13, pp. 1547–1555
1552 © The Institution of Engineering and Technology 2016
 17518652, 2016, 13, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-cta.2015.1221 by CAPES, Wiley Online Library on [09/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Fig. 2 Node-based adaptive event-triggered strategy
a State of agent i for i = 1, . . . , 6
b Event instant sequence of each agent
c Norm of the measurement error of agent 1

a b

Fig. 3 Fully distributed edge-based adaptive event-triggered strategy

a State of agent i for i = 1, . . . , 6
b Event instant sequence of each agent
c Norm of the measurement error of agent 6

IET Control Theory Appl., 2016, Vol. 10, Iss. 13, pp. 1547–1555
© The Institution of Engineering and Technology 2016 1553
 17518652, 2016, 13, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-cta.2015.1221 by CAPES, Wiley Online Library on [09/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
a b

Fig. 4 Fully distributed node-based adaptive event-triggered strategy

a State of agent i for i = 1, . . . , 6
b Event instant sequence of each agent
c Norm of the measurement error of agent 6

Example 4: Intuitively, compared to node-based method, the edge- Table 1 Event times of four event-triggered protocols among three
based event-triggered adaptive method is more affected by the graphs in 10 s
distribution of the agents’ in-degree. Meanwhile, from the event-
Laplacian matrix L1 L2 L3
triggering condition (9), it contains the maximum eigenvalue of
the Laplacian matrix. So the maximum eigenvalue of the Laplacian edge based the event times of each 16–67 14–68 26–60
matrix will also affect the performance. Here an example is given agent
to show the event times difference of the four adaptive protocols edge based the average event times 42 31 60
among three different graphs. The Laplacian matrixes of the three node based the event times of each 30–97 24–65 32–70
agent
graphs are L1 , L2 and L3 node based the average event times 66 54 60
⎡ ⎤ fully distributed the event times of each 9–33 8–27 6–48
2 −1 0 0 0 −1 edge based agent
⎢−1 2 −1 0 0 0⎥ fully distributed the average event times 17 14 26
⎢ ⎥
⎢ 0 −1 2 −1 0 0⎥ edge based
L2 = ⎢
0 −1 2 −1 0 ⎥
, fully distributed the event times of each 14–31 20–56 12–37
⎢0 ⎥
⎣0 0 0 −1 2 −1 ⎦ node based agent
fully distributed the average event times 24 29 28
−1 −1 0 0 −1 2 node based
⎡ ⎤
5 −1 −1 −1 −1 −1
⎢−1 1 0 0 0 0⎥
⎢ ⎥
⎢−1 0 1 0 0 0⎥ Comparing the first and the second cases, i.e. edge-based and the
L3 = ⎢ ⎥.
⎢ −1 0 0 1 0 0 ⎥ node-based strategies under all the three topologies, the node-based
⎣−1 0 0 0 1 0⎦ adaptive event-triggered strategy always generates more events.
−1 0 0 0 0 1 This differences between the edge-based and the node-based strate-
gies also can be observed in the latter cases using the fully
The maximum eigenvalue of these three graphs are λN (L1 ) = distributed event-triggering condition.
4.8136, λN (L2 ) = 4, λN (L3 ) = 6. The number of neighbours of Similar to the first case, when using the fully distributed event-
agents under the third topology varies the most, while the sec- triggering condition (26), the event quantities reduced very much
ond topology has the smallest in-degree variance. The event times under the second topology. This may be because the in-degrees
of the adaptive protocols under the same parameters are listed in directly affect the event-triggering condition (26). While using a
Table 1. From the first case, which uses the edge-based adaptive topology that has even in-degrees of each agents, this strategy has
event-triggered control strategies (5), (6), (9) under three topolo- much less average event quantities.
gies, one can see that the events quantity under three topologies When it comes to the fourth case, i.e. combing the node-based
vary more significantly. The edge-based event-triggered adaptive adaptive event-triggered control laws (22), (23) with the fully dis-
method generates the least events under the second topology. tributed event-triggering condition (26), it barely affected by the
variance of the topologies, which cooperates with the situation of
The second case is the node-based event-triggered adaptive the second case.
method (22), (23), (9), which is more affected by the maxi- Furthermore, the fully distributed event-triggering condition (26)
mum eigenvalue of the Laplacian matrix. The larger the maximum generates much less events than the event-triggering condition (9)
eigenvalue is, the more events the strategy generates. does.

IET Control Theory Appl., 2016, Vol. 10, Iss. 13, pp. 1547–1555
1554 © The Institution of Engineering and Technology 2016
 17518652, 2016, 13, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-cta.2015.1221 by CAPES, Wiley Online Library on [09/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Remark 6: The adaptive parameter makes it very difficult to give a 6 Yu, W., Chen, G., Cao, M., Kurths, J.: ‘Second-order consensus for multiagent
positive lower bounded of the interval as the parameter varies along systems with directed topologies and nonlinear dynamics’, IEEE Trans. Syst.
Man Cybern. B, Cybern., 2010, 40, (3), pp. 881–891
with the convergence status. Therefore Theorem 2 only proves that 7 Li, J., Xu, S., Chu, Y., Wang, H.: ‘Distributed average consensus control in
when the consensus is not achieved and the agent is not specifically networks of agents using outdated states’, IET Control Theory Appl., 2010, 4,
located at the average position of all its neighbours, the interval is (5), pp. 746–758
positive rather than strictly lower bounded. In Theorem 4, the inter- 8 Rong, L., Xu, S., Xie, D., Zou, Y.: ‘Forced second-order consensus in directed
networks with time delays’, IMA J. Math. Control Inf., 2013, 30, (3), pp. 329–343
event interval can also proved to be positive when the consensus is 9 Yuan, D., Xu, S., Zhao, H., Chu, Y.: ‘Distributed average consensus via gossip
not reached. However, we still cannot give a positive lower bound algorithm with real-valued and quantized data for 0 < q < 1’, Syst. Control Lett.,
of the interval. In the future, the existence of a lower bound will 2010, 59, (9), pp. 536–542
be sought. After the consensus is achieved, for the worst case, the 10 Li, Z., Ren, W., Liu, X., Xie, L.: ‘Distributed consensus of linear multi-
agent systems with adaptive dynamic protocols’, Automatica, 2013, 49, (7),
event-triggered adaptive protocol will not cause more communi- pp. 1986–1995
cation than continuous protocols. However, from the simulations, 11 Li, Z., Ren, W., Liu, X., Fu, M.: ‘Consensus of multi-agent systems with general
one can see that the event-triggered adaptive protocol is effective linear and lipschitz nonlinear dynamics using distributed adaptive protocols’,
and there is no Zeno behaviour. IEEE Trans. Autom. Control, 2013, 58, (7), pp. 1786–1791
12 Su, H., Chen, G., Wang, X., Lin, Z.: ‘Adaptive second-order consensus of net-
worked mobile agents with nonlinear dynamics’, Automatica, 2011, 47, (2),
pp. 368–375
5 Conclusion 13 Wang, L., Feng, W.-J., Chen, M.Z.Q., Wang, Q.-G.: ‘Consensus of nonlinear
multi-agent systems with adaptive protocols’, IET Control Theory Appl., 2014,
We considered the consensus problem for multi-agent systems in 8, (18), pp. 2245–2252
14 Yu, W., Zheng, W.X., Chen, G., Ren, W., Cao, J.: ‘Second-order consensus in
this paper. We combined the strategy of event-triggered commu- multi-agent dynamical systems with sampled position data’, Automatica, 2011,
nication with the adaptive control protocol. Two adaptive control 47, (7), pp. 1496–1503
protocols, namely, the edge-based and the node-based adaptive pro- 15 Chen, W., Li, X., Jiao, L.C.: ‘Quantized consensus of second-order continuous-
tocols, are considered. It is shown that the proposed controllers time multi-agent systems with a directed topology via sampled data’, Automatica,
2013, 49, (7), pp. 2236–2242
are able to decrease the channel usages and communication fre- 16 Cheng, L., Wang, Y., Hou, Z.-G., Tan, M., Cao, Z.: ‘Sampled-data based average
quencies while achieving consensus for multi-agent systems with consensus of second-order integral multi-agent systems: switching topologies and
only the local information. Furthermore, an event-triggering condi- communication noises’, Automatica, 2013, 49, (5), pp. 1458–1464
tion that only involves local information of the agent is designed. 17 Huang, N., Duan, Z., Chen, G.R.: ‘Some necessary and sufficient conditions
for consensus of second-order multi-agent systems with sampled position data’,
By using this event-triggering condition, the global information of Automatica, 2016, 63, pp. 148–155
the system does not need to be assigned to agents off-line. Zeno 18 Tabuada, P.: ‘Event-triggered real-time scheduling of stabilizing control tasks’,
behaviour is also analytically excluded for both cases. Simulations IEEE Trans. Autom. Control, 2007, 52, (9), pp. 1680–1685
are implemented to show the efficiency of the proposed strategies. 19 Dimarogonas, D.V., Frazzoli, E., Johansson, K.H.: ‘Distributed event-triggered
control for multi-agent systems’, IEEE Trans. Autom. Control, 2012, 57, (5),
Future work includes extending the proposed methods to higher pp. 1291–1297
order multi-agent systems and study of other collective behaviours. 20 Meng, X.: ‘Distributed event triggered control, estimation, and optimization for
cyber-physical systems’. PhD thesis, University of Alberta, 2014
21 Zhang, X.-M., Han, Q.-L.: ‘Event-triggered dynamic output feedback control for
networked control systems’, IET Control Theory Appl., 2014, 8, (4), pp. 226–234
6 Acknowledgments 22 Zhang, X.-M., Han, Q.-L.: ‘Event-based h∞ filtering for sampled-data systems’,
Automatica, 2015, 51, pp. 55–69
This work was supported by the National Natural Science 23 Dimarogonas, D.V., Johansson, K.H.: ‘Event-triggered cooperative control’, Eur.
Foundation of PR China under grants 61203056, 61374086, Control Conf., 2009, pp. 3015–3020
24 Huang, N., Duan, Z., Wen, G., Zhao, Y.: ‘Event-triggered consensus tracking of
61374087, 61473151, the Program for Changjiang Scholars and multi-agent systems with Lur’e nonlinear dynamics’, Int. J. Control, 2016, 89,
Innovative Research Team in University under grant IRT13072, a (5), pp. 1025–1037
project funded by the priority academic program development of 25 Yan, H., Shen, Y., Zhang, H., Shi, H.: ‘Decentralized event-triggered consen-
Jiangsu Higher Education Institutions, Jiangsu Planned Projects for sus control for second-order multi-agent systems’, Neurocomputing, 2014, 133,
pp. 18–24
Postdoctoral Research Funds 1501022C. 26 Zhu, W., Jiang, Z.-P., Feng, G.: ‘Event-based consensus of multi-agent systems
with general linear models’, Automatica, 2013, 50, (2), pp. 552–558
27 Meng, X., Chen, T.: ‘Event based agreement protocols for multi-agent networks’,
7 References Automatica, 2013, 49, (7), pp. 2125–2132
28 Guo, G., Ding, L., Han, Q.-L.: ‘A distributed event-triggered transmission strat-
1 Alexander Fax, J., Murray, R.M.: ‘Information flow and cooperative control of egy for sampled-data consensus of multi-agent systems’, Automatica, 2014, 50,
vehicle formations’, IEEE Trans. Autom. Control, 2004, 49, (9), pp. 1465–1476 (5), pp. 1489–1496
2 Reynolds, C.W.: ‘Flocks, herds, and schools: a distributed behavioral model’, 29 Liu, Z., Chen, Z.: ‘Reaching consensus in networks of agents via event-triggered
Comput. Graph., 1987, 21, (4), pp. 25–34 control’, J. Inf. Comput. Sci., 2011, 8, (3), pp. 393–402
3 Olfati-Saber, R., Alexander, Fax J., Murray, R.M.: ‘Consensus and cooperation 30 Chen, X., Hao, F., Shao, M.: ‘Event-triggered consensus of multi-agent systems
in networked multi-agent systems’, Proc. IEEE, 2007, 95, (1), pp. 215–233 under jointly connected topology’, IMA J. Math. Control Inf., 2015, 32, (3),
4 Ren, W., Beard, R.W.: ‘Consensus algorithms for double-integrator dynamics’, pp. 537–556
Distrib. Consensus Multi-Veh. Coop. Control Theory Appl., 2008, pp. 77–104 31 Godsil, C.D., Royle, G.: ‘Algebraic graph theory’ (Springer, New York, 2001,
5 Ren, W., Atkins, E.: ‘Distributed multi-vehicle coordinated control via local vol. 207)
information exchange’, Int. J. Robust Nonlinear Control, 2007, 17, (10–11), 32 Xiao, F., Meng, X., Chen, T.: ‘Sampled-data consensus in switching networks
pp. 1002–1033 of integrators based on edge events’, Int. J. Control, 2015, 88, (2), pp. 391–402

IET Control Theory Appl., 2016, Vol. 10, Iss. 13, pp. 1547–1555
© The Institution of Engineering and Technology 2016 1555