IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO.

1, JANUARY 2021 161

A Survey on Intelligent Control for

Multiagent Systems
Peng Shi , Fellow, IEEE, and Bing Yan

Abstract—In practice, the dual constraints of limited

interaction capabilities and system uncertainties make it difficult
for large-scale multiagent systems (MASs) to achieve intelligent
collaboration with incomplete local relative information. In this
article, a review is conducted on the recent development of MASs
intended for intelligent control, including consensus problem,
formation control, and flocking control. Based on the limita-
tions of the interaction level and the constraints of the individual
system level, the published results on intelligent control are cat-
egorized into limited sensing-based control, event-based control,
pinning-based control, resilient control, and collaborative con-
trol under system constraints. Also, the applications of intelligent
control for MASs are presented, especially for robotics, complex
networks, and transportation. Finally, a discussion is given about Fig. 1. Agent and its features.
the challenges and future directions of research in this field.
Index Terms—Consensus problem, flocking control, formation
control, intelligent control, multiagent systems (MASs), multilevel
constraints.

I. I NTRODUCTION
ULTIAGENT collaborative intelligence technology has
M led to revolutionary changes for the practical appli-
cations of robotics, complex networks, and transportation in
recent years [1]–[3]. In the meantime, however, it also presents
challenges as distributed and reliable intelligence technol- Fig. 2. Intelligent control for MASs.
ogy is required to perform cooperation tasks for large-scale
multiagent systems (MASs) in the context of incomplete features commonly used to describe an agent, including auton-
local relative information. Inspired by the collaborative behav- omy, reactivity, sociality, and proactiveness. Autonomy means
iors observed in nature, such as bird migration in groups that an agent is capable to operate without the direct inter-
and flocking behaviors of fish schools, collaborative aware- vention of other entities and exercising control over its own
ness, task assignment, and intelligent control of MASs have actions. Reactivity is defined as the capability of an agent to
attracted much attention from various fields over the past make a response to the changes in the environment and con-
decades [4]–[6]. As the fundamental way to ensure the suc- vert its sensory inputs to actions. Sociality means that an agent
cessful collaborative missions for MASs, advanced intelligent has the capability to make communications with others, while
control strategies are the focus in this survey article. proactiveness stands that an agent can do more than acting
From the perspective of computer science, an agent refers in response to the environment. However, the capability of
to a computing system operating in an environment with cer- a single agent is limited, especially in dealing with complex
tain levels of autonomy and capability of sensing, decision tasks.
making, and acting [1]. As shown in Fig. 1, there are four A MAS refers to a group of agents. It is capable to inter-
act, coordinate their behavior, and cooperate to achieve some
Manuscript received November 19, 2020; revised November 25, 2020;
accepted December 1, 2020. Date of publication December 24, 2020; date of common goals [1]. In comparison with single-agent systems,
current version January 12, 2021. This work was supported by the Australian MASs provide a more effective and robust way to solve
Research Council under Grant DP170102644. This article was recommended various complex problems by means of collaborative intelli-
by Associate Editor C. K. Ahn. (Corresponding author: Peng Shi.)
The authors are with the School of Electrical and Electronic gence. As shown in Fig. 2, a MAS consists of an information
Engineering, University of Adelaide, Adelaide, SA 5005, Australia (e-mail: interaction level and a physical system level. At the interaction
[email protected]; [email protected]). level, information is exchanged not only between individual
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TSMC.2020.3042823. agents but also between agents and their ambient environment
Digital Object Identifier 10.1109/TSMC.2020.3042823 through either communication networks or sensor perception.
2168-2216 
c 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
162 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

In practice, the range of communication capabilities and per-

ception is limited for a single agent. At the physical system
level, constraints, such as uncertainties and complex heteroge-
neous dynamics, can have a severe impact on the performance
of MASs.
As for the work on intelligent control of MASs, it focuses on
three major problems, including consensus problem, formation
control problem, and flocking/swarming problem. As the foun-
dation of research on collaborative intelligence, the consensus
problem of MASs has now been extensively investigated for
all agents to achieve a common goal [3], [7]–[11]. As an exten-
sion to consensus problems, formation control aims at driving
intractable agents to maintain and move as desired geometric
shapes to perform predefined tasks, such as effective search,
patrol, and exploration [6], [12], [13]. As a self-organizing
behavior, flocking or swarming is derived from small-size ani-
mals with lower intelligence [14]–[16], for example, bees, fish Fig. 3. Structure of the article.
school, and bird swarms, in the process of migrating, cruising,
or avoiding enemies. Swarming is also extended to describe
the behavior of lifeless agents, such as robots. Swarm intelli- different capabilities in cooperative operations. As one of the
gence not only expands individual capability but also improves basic heterogeneous systems, hybrid-order MASs consist of
the overall level of survivability. different order integrator systems were described in [29]–[32].
Depending on different control structures, the approaches However, the less restrictive heterogeneous systems with dif-
taken for the intelligent control for MASs can be classified ferent orders and different dynamics are more commonly
into centralized control [17] and distributed control [18]. In applied in practice. In recent years, the robust output regu-
respect of centralized control, there is a control center or host lation control exercised by general heterogeneous MASs has
in place to coordinate the information transmission and the attracted a great number of attention [33]–[36]. Subsequently,
final process of task completion. However, the failure of the it was extended to solve the formation control problem and the
control center will hinder the entire system from functioning flocking problem encountered by heterogeneous systems. The
as normal. To improve the robustness of the whole MASs, dis- challenge still arises from the systems due to the limitations
tributed control approaches have been widely studied, where of uncertainties and heterogeneous dynamics.
all agents determine their behavior separately based on local In this article, a survey is conducted on the recent study
information. The design of distributed intelligent control that of MASs in intelligent control considering the constraints of
only relies on incomplete local information has been a study information interaction level and system level. We try our best
focus in recent years [2], [7], [18]. to summarize the relevant research work in recent years, and
Information interaction and system dynamics play a vital apologize for missing some contributions on the topic, if any.
role in the intelligent control for the entire MASs. At the The overall structure of the article is shown in Fig. 3. The
interaction level, the information interaction between agents background and basic concepts are introduced in Section I.
relies on the capabilities of each agent to carry out sens- Preliminaries and three major problems with intelligent con-
ing and communication. In general, the information level of trol for MASs are described in Section II, including consensus,
practical MASs is subject to a limited range of perception formation, and flocking problems. In Section III, the recent
without communication, limited bandwidth with communica- advancement of the methodologists to tackle the underlying
tion, limited network resources, and other network-induced problems is described with the constraints of the interaction
issues. According to the different limitations on the interaction level and system level. Then, a review is presented on the
level, the recently proposed methods to solve MAS collabo- main applications of robotics, complex networks, and trans-
rative control can be categorized into limited sensing-based portation in Section IV. The conclusion and challenges ahead
control [19], event-driven control [20], pinning control [21], are discussed in Section V.
and resilient control [22].
At the system level, an overview of theoretic advancements
in consensus, formation control, and flocking control has II. P RELIMINARIES AND K EY P ROBLEMS OF I NTELLIGENT
been presented for MASs with single-integrator and double- C ONTROL FOR MAS S
integrator dynamics under fixed and switching topologies In this section, preliminaries about graph theory and three
in [11] and [23]–[25]. In addition to linear MASs, a large key issues of intelligent control are recalled, including con-
number of works on reliable intelligent control have been sensus problem, formation control, and flocking or swarming.
intensively investigated for homogeneous nonlinear MASs In general, the information exchange among agents is mod-
under systems uncertainties and disturbances [7], [26]–[28]. eled by a directed or undirected graph [37]. In a graph
Compared with homogeneous agents, heterogeneous agents G = (V, E), V represents a finite nonempty set of nodes and
show greater flexibility in task allocation depending on E is an edge set, which contains ordered pairs of nodes in

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
SHI AND YAN: A SURVEY ON INTELLIGENT CONTROL FOR MASs 163

TABLE I
R ECENT W ORKS ON C ONSENSUS P ROBLEMS

Fig. 4. Graphs with different features. (a) Directed graph. (b) Undirected
graph. (c) Spanning tree. (d) Connected graph.

a directed graph and unordered pairs of nodes in an undi-

rected graph. For example, (i1 , i2 ) ∈ E indicates that i2 obtains
the information from i1 in a directed graph [Fig. 4(a)], and where zi (t) ∈ Rm represents the state or output of agent i.
i1 and i2 can obtain the information from each other in an z0 (t) ∈ Rm is a common desired trajectory for all agents to
undirected graph [Fig. 4(b)]. The neighbor set of agent i is track asymptotically.
Ni = {j ∈ V|(j, i) ∈ E}. A graph contains a spanning tree Remark 1: It is worth noting that the leader can be a real
if there is a path between one node and all other nodes physical system, or a virtual reference system designed accord-
[Fig. 4(c)]; a graph is connected if there is a path between ing to the tasks. The final consensus of all agents without
every pair of distinct nodes [Fig. 4(d)]. For a graph G, an adja- reference to their initial conditions in the leader–follower con-
cency matrix A = [aij ] specifies the interconnection topology sensus problem. In terms of the leaderless consensus problem,
of MASs, where all agents finally reach a consensus, which is related to the
 initial state of the system and the information interaction
0 i = j, or (j, i) ∈
/E
aij = (1) topology.
1 (j, i) ∈ E. A summary of recent works on Problems 1 and 2 is outlined
The Laplacian matrix L of graph G is L = D − A where in Table I, corresponding to different features and constraints
D = diag(d1 , d
2 , . . . , dn ) is the degree matrix with diagonal from system level and interaction level.
elements di = j aij .
B. Formation Control Problem
A. Consensus Problem Formation control is designed to drive the moving
As one of the research foundations for intelligent con- interacting agents to achieve or maintain a specified geom-
trol for MASs, consensus refers to all systems reaching an etry for a coordinated goal. The formation control problem
agreement on certain interests regarding to their states and can be uniformly summarized into a consensus-based struc-
the concept comes from distributed computing systems and ture [70], after considering the reference formation dynamics
management science. A typical consensus theoretical frame- and the motion characteristics of the agent.
work was presented in [11], and the framework established Most results on formation control have focused on two
some communication rules among the agent and their neighbor main problems: 1) leaderless formation control problem and
agents in the networks in order to achieve a common goal. The 2) formation tracking problem.
study also emphasized that graph theory and Laplace matrix Problem 3: A general leaderless formation control problem
were the core means to solve the consensus problem. Another is to design a controller for a MAS to meet
work linked the minimum spanning tree theory in the graph    
with the information consensus framework [23] and proposed lim  zi (t) − zj (t) − fi − fj  = 0, j ∈ Ni (4)
t→∞
the minimum necessary and sufficient condition of information
where (fi − fj ) represents the reference formation deviation
consensus for MASs under changing topologies, which laid the
between agent i and agent j.
foundation for the research of dynamic topologies.
Problem 4: A general formation tracking problem is to
According to the theoretical frameworks, consensus prob-
design a controller for a MAS to meet
lems can be divided into leaderless consensus problem and
leader–follower consensus problem. lim zi (t) − fi0 − z0 (t) = 0, i = 1, 2, . . . , n (5)
Problem 1: A general leaderless consensus problem is to t→∞

design a controller for a MAS to meet where fi0 ∈ Rm is a referent formation deviation regards to
  a desired trajectory z0 (t) ∈ Rm . If fi0 is a dynamic formation
lim zi (t) − zj (t) = 0, j ∈ Ni (2)
t→∞ variable, the problem is extended to a time-varying formation
where zi (t) ∈ Rm and zj (t) ∈ Rm represent the state or output problem.
of ith agent and jth agent, respectively. Ni is the neighbor set Remark 2: Note that formation control Problems 3 and 4
of agent i. can be viewed as extensions of consensus Problems 1 and 2,
Problem 2: A general leader–follower consensus problem respectively, with respect to a reference formation deviation.
is to design a controller for a MAS to meet For example, as shown in Fig. 5, the agent i needs to maintain a
diamond formation deviation (fi − fj ) with its neighbor agent j,
lim zi (t) − z0 (t) = 0, i = 1, 2, . . . , n (3) in leaderless formation control Problem 3. Therefore, the main
t→∞

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
164 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

Fig. 5. Leaderless formation problem and formation tracking problem. Fig. 6. Flocking or swarming control problem.

TABLE II
R ECENT W ORKS ON F ORMATION C ONTROL the Vicsek model. Its advantage is to quantitatively analyze
the flocking behavior by changing the population density and
noise intensity, from the point of view of statistical mechanics.
These models provide a solid foundation for the research of
the flocking problem.
According to the model and rules, the flocking control
problem can be described as follows.
Problem 5: A general flocking/swarming control problem
is to design a controller for a MAS to meet the three following
rules.
1) Cohesion: Remain close to neighbors

n 
n
 
focus of Problem 3 is to form a formation. As for formation lim pi (t) − pj (t) < R, j ∈ Ni (6)
tracking Problem 4, there are a moving target z0 (t) to track t→∞
i=1 j=1
and a formation deviation fi0 to keep for agent i. In Fig. 5, a
MAS aims to track a desired trajectory z0 (t) from a rectangle where pi (t) ∈ Rm and pj (t) ∈ Rm represent the positions
to a parallelogram. of ith agent and jth agent, respectively. R is the max-
The approaches for formation control reported in the imum value of the sum of relative distances between
literature include the leader–follower control [12], [71], agents.
the behavior-based control [72], and the virtual structure 2) Separation: Avoid collision with neighbors
 
approach [73]. In the leader–follower approach, the controller lim pi (t) − pj (t) ≥ dij , j ∈ Ni (7)
relies heavily on a single leader state. For the behavior-based t→∞

formation method, several basic control behaviors of the agent where dij is the minimum safety distance between agent
are defined and weighted to obtain the final formation control i and agent j.
inputs for the group. However, group behaviors are difficult 3) Alignment: Match velocity with neighbors
to define. In the virtual structure approach, the formation of  
lim vi (t) − vj (t) = 0, j ∈ Ni (8)
agents is regarded as a single object in the virtual structure, t→∞
which limits the application domain as it only controls the where vi (t) ∈ Rmand vj (t) ∈ Rm denote the velocities of agent
motion of one object. The existing results can also be divided i and agent j, respectively.
into position-based, displacement-based, and distance-based To solve the problem according to the three rules shown in
control, according to the sensing capability and the interaction Fig. 6, a theoretical framework was proposed for the design
topology of MASs [6]. Focusing on the constraints and fea- of distributed flocking algorithms of second-order MASs [25].
tures from the system level and the interaction level of agents, The work in [99] analyzed the stability properties of flocking
we list some related works on formation control in Table II. algorithms for second-order MASs under switching networks.
We summarize some recent works on flocking and swarming
C. Flocking or Swarming Control Problem problems in Table III, corresponding to different features and
constraints from the system level and the interaction level.
The flocking or swarming control problem of MASs is to
perform a macroscopic overall synchronization, such as aggre-
gate together and maintain the same direction, by using local III. M ETHODOLOGIES OF I NTELLIGENT C ONTROL FOR
interaction and behavioral rules between agents. The Boid MAS S
model was first proposed based on computer simulation tech- In this section, we review the results reported for the collab-
nology to describe bird swarms, which follows three rules: orative intelligence of MASs and outline advanced methodolo-
1) cohesion: remain close to neighbors; 2) separation: avoid gies based on the limitations of information interaction level
collision with neighbors; and 3) alignment: match velocity and the constraints of system level, respectively.
with neighbors [97]. The work in [98] simplified the Boid Passive sensing and active communication are the two
model and described the behavior of birds as a discrete model, important means of information interactions. For example, if

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
SHI AND YAN: A SURVEY ON INTELLIGENT CONTROL FOR MASs 165

TABLE III
R ECENT W ORKS ON F LOCKING O R S WARMING C ONTROL

Fig. 7. Limited sensing-based intelligent control.

each agent is equipped with advanced sensors, which make
them able to detect the location of neighbors, then commu-
nication is not required in formation control. However, the where ri is the sensing range of the ith agent, which can be
detection range of a single agent is usually limited. It becomes different for heterogeneous agents. The control force f aims
relatively easy for collaboration if agents can directly interact to form a reference formation F, which is a set of scalable
with key information through the network, such as locations displacement vectors F = {. . . , f−1 , f0 , f1 , . . .}. The element
and velocities. While communication networks are restricted fk ∈ Rm is based on m linearly independent basis vectors,
by limited bandwidth, limited resources, and other network- where m = 2 or m = 3 indicates the reference formation
induced issues. From the limitations of information interaction, is in two dimensions or three dimensions. For instance, a
we focus on reviewing the following popular and impor- 2-dimensional (2-D) scalable formation F is shown in Fig. 7,
tant topics: limited sensing-based control, event-based control, where dx and dy are two linearly independent vectors. For any
pinning-based control, and resilient control. fk ∈ F, fk = adx + bdy , a, b ∈ Z always holds. Note that the
basis vectors of F are known for all agents. The local mapping
A. Limited Sensing-Based Control function g is conducted on each agent to choose which vectors
Under certain situations such as an electrostatic shielding in F are the suitable displacement according to the neighbor
environment, the communication between agents is unavail- set Ni . Based on local optimization methods, multiobject map-
able or very limited. Sensor-based perception is another ping protocols without conflict and range control strategies
alternative to achieve multiagent collaboration. In fact, it is have been investigated in [19], [75], and [84]. In addition to
beneficial to realize distributed control if agents can com- collision-free matching, obstacle avoidance algorithms have
pletely perceive neighbors and the environment independently. also been extended to the formation control in Problems 3
Moreover, the system can be immune to any network prob- and 4 without communication [121] and with limited com-
lems due to the independence of the communication network. munication only on identities and mapping decisions [76]. In
However, for a single agent, the detection range of sen- order to reduce the computation burden and avoid the infinite
sors, such as cameras and infrared sensors, is locally limited. trajectory loop, the probability was introduced in mapping and
Note that range-only sensing agents usually refer to robotics distributed control strategies [85], [86].
rather than network agents. Therefore, most of the exist- As for the flocking control in Problem 5, most of the works
ing results on limited sensing-based control have focused have focused on the implementation of three rules for agents
on solving formation control problems and flocking control subject to a limited range-based perception or interaction. A
problems [25], [75]. As shown in Fig. 7, formation reference theoretical flocking framework was proposed of a second-order
denotes static or dynamic predefined displacement references MASs [25]
regard to different formation shapes, while flocking reference ṗi (t) = vi (t), v̇i (t) = ui (t) i = 1, 2, . . . , n (10)
refers to a moving rendezvous point as the group objec-
tive [25]. There are two key issues: 1) how does the agent under the control force
determine the most suitable position relative to its neighbors? g
and 2) how to achieve collaboration actions without collisions? ui (t) = fi + fid + fir
      
In order to solve the two problems, a novel scalable forma- = ψ pj − pi σ nij + vj − vi
tion control strategy was proposed to solve Problem 3, when j∈Ni j∈Ni
the communication network is completely unreachable [75]. + c1 ((p0 (t) − pi (t))) + c2 ((v0 (t) − vi (t)) (11)
Consider a first-order MASs under a control input ui (t)
where the system state xi (t) = [pi (t), vi (t)]T consists of posi-
ṗi (t) = ui (t), i = 1, 2, . . . , n tion pi ∈ Rm and velocity vi ∈ Rm of the ith agent. To
ui (t) = f (Ni , g(Ni , F)) (9) implement the three rules, control input consists of three items:
g
1) gradient-based term fi ; 2) velocity consensus item fid ; and
where pi ∈ Rm and ui ∈ Rm present the position and control
3) navigational feedback term fir , where ψ is a potential func-
input of the ith agent, respectively. Suppose the sensing range
tion based on σ -norm of the relative distance for collision
is limited, the neighbor set is defined as
   avoidance. (nij ) represents a vector connecting agent i and
Ni = (j, i) ∈ E| pi (t) − pj (t) < ri agent j. Parameters c1 and c2 denote the position control law

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
166 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

1) Trigger Mechanism: It is designed to determine the trig-

i
ger time in next step tk+1 . The recent works mainly focus
on two common mechanisms, event-triggered mechanism, and
self-triggered mechanism, as follows:

(1) tk+1
i
= inf t > tki |f (ei ) > g(ei , φi )
(2) tk+1
i
= tki + σi (12)

Fig. 8. Event-based intelligent control. where f (ei ) and g(ei , φi )) are the trigger functions based on
measurement error ei and collaboration error φi . The next time
is triggered if the condition is met. Note that the trigger func-
tions are only based on local information without the prior
and velocity control law of the navigational feedback, respec- information of the topology matrix by adding adaptive laws in
tively. They are designed to meet c1 , c2 > 0. The flocking f (ei ) or g(ei , φi ) [47]–[50]. The second mechanism is the self-
reference position p0 ∈ Rm and velocity v0 ∈ Rm indicate a triggered strategy, where the time interval of next broadcast σi
rendezvous point, which can be viewed as a group objective. can be calculated based on the information of the current trig-
The connectivity preservation of formation control was also ger time without the need to continuously monitor the changes
been discussed in [25]. of events [104]. Although event monitoring costs are reduced,
Based on the basic flocking framework, flocking control additional computational costs are added.
solutions to Problem 5 were provided based on potential There are two common designs of measurement error
field [14] and learning-based approach [15], [16]. Without a  
communication channel, learning vision-based flocking algo- (1) ei (t) = x̂i (t) − xi (t) = xi tki − xi (t)
 i  
rithms was proposed for multidrone swarms [15]. Flocking
(2) ei (t) = x̂i (t) − xi (t) = eA t−tk xi tki − xi (t) (13)
control was developed for a first-order MAS subject to limited
heterogeneous interaction range [100]. However, connectivity where tki is the kth trigger time of agent i, and x̂i (t) repre-
preservation is still a challenge for high-order nonlinear MASs, sents the estimation of xi in the time period t ∈ [tki , tk+1 i ). The
when the sensing capability of the agent is insufficient. state value at the time of the last trigger is used in the first
strategy [48], [51]–[54]. The second approximation scheme of
xi (t) is based on the system state matrix A. System model-
B. Event-Based Intelligent Control based estimation more accurately approximates the state of
The communication-based MASs make up for the lack of the system during the time period [49], [50].
perception of some agents and effectively realize collaboration The design of the collaboration error is based on differ-
by directly interacting with the information of interest through ent intelligent control problems. For leader–follower consensus
the network. However, limited network resources and commu- Problem 2, formation control in Problem 4, and flocking track-
nication bandwidth have largely restricted information trans- ing control in Problem 5, the collaborative errors are generally
mission between agents. Event-based distributed interaction formed as
mechanisms were proposed to alleviate this limitation [9],     
(1) φi (t) = aij x̂j (t) − x̂i (t) + ai0 x̂0 (t) − x̂i (t) (14)
[47]–[57]. Without collecting the state information of all
j∈Ni
agents at every moment, the distributed event-triggered strat-   
egy can better save communication resources and effectively (2) φi (t) = aij x̂j (t) − x̂i (t) − f̂j (t) − f̂i (t)
reduce the frequency of information transmission between j∈Ni
agents and update of agent control protocols. + ai0 x̂0 (t) − x̂i (t) − f̂i0 (t) − x̂i (t) (15)
The structure of event-based distributed intelligent control   
is shown in Fig. 8, where each agent independently determines (3) φi (t) = aij v̂j (t) − v̂i (t)
its own behaviors. The trigger determines the interaction time j∈N
interval of each agent according to the measurement error from i   
+ c1 p̂0 (t) − p̂i (t) + c2 v̂0 (t) − v̂i (t) (16)
the sampler and the collaboration error from the controller. The
controller updates the local information and possible reference where x̂0 (t) is the state estimation of a real leader or a vir-
information at each trigger moment. Note that the possible ref- tual leader indexed by number zero, which can be regarded as
erence information here refers to the consensus reference, or consensus reference. If the agent i is informed by the leader,
the formation reference, or the flocking reference information. ai0 = 1, otherwise, ai0 = 0. f̂i (t) is the estimation of formation
For the leaderless consensus problem in Problem 1, there is reference at time t ∈ [tki , tk+1 i ) [83]. As for leaderless consen-
no external reference information. The event-based distributed sus Problem 1 and formation control without a reference leader
trigger strategy mainly involves three key issues: 1) how to in Problem 3, the terms related to ai0 are zero [51], [52]. For
determine the trigger time; 2) how to design the distributed flocking control in (16), the collaborative error composes of
control laws; and 3) how to exclude the unlimited trigger phe- the velocity consensus error and flocking tracking error, where
nomenon, Zeno phenomenon [122]. We take the event-based x̂0 (t) = [p̂0 (t), v̂0 (t)] is the estimation of a flocking reference
control for second-order MASs (10) as an example. trajectory with position and velocity [103].

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
SHI AND YAN: A SURVEY ON INTELLIGENT CONTROL FOR MASs 167

2) Control Strategies: are also designed in terms of intel-

ligent control Problems 1–5. For consensus and formation
control problems, the controller aims to eliminate collaborative
errors by
ui (t) = Ki φi (t) (17)
where Ki is the control law matrix based on ARE (alge-
braic Riccati equation) [9], [47], [49] and LMI (linear matrix
inequality) [13], [48], [83]. Quantized event-triggered con-
trol is another improvement direction to save more network
resources under limited bandwidth by the quantized control
law Fig. 9. Pinning-based intelligent control.

ui (t) = Ki qu (φi (t)) (18)

where qu is a quantized function to convert the collaborative the communication topology [23]. However, this assumption is
error into a discrete form [48], [53]. difficult to be satisfied especially for MASs under time-varying
For flocking control in Problem 5, one form of control topologies. Moreover, for real-world large-scale MASs, such
protocol is as complex power grids and multidrone light show systems,
     there are generally a large number of nodes or control points
ui (t) = ψ p̂j − p̂i σ nij + φi (t) (19) in their communication networks. It is usually difficult and
j∈Ni expensive to put controllers to all nodes to control the whole
where ψ is a potential function [101]–[103] to remain system. Pinning-based control is one of the effective solutions
close to neighbors without collisions. Another control strat- for the issues. The basic idea of pinning control is by adding a
egy is based on the distributed model predictive control pinner to control a small fraction of agents so that the whole
by solving optimization problems to satisfy the flocking MASs can achieve collaborative performance. As shown in
rules [104], [106]. Fig. 9, a pinner (or virtual leader) is added to the system and
3) Proof That the Zeno Phenomenon Is Excluded: If an defines its desired trajectory, according to different intelligent
event is triggered infinitely within a finite time, the phe- control Problems 1–5. The pinner only controls some pinned
nomenon is called Zeno phenomenon [122]. In the study of agents (agent 2, agent 3, and agent 7) from different groups
an event-based mechanism, one of the key tasks is to exclude (G1 , G2 , and G3 ) to ensure global collaboration, and the choice
Zeno phenomenon. One widely used method is to prove that of pinned agent to inform the reference information is via the
there must be a positive lower bound on the interval length pinning strategies. Therefore, the realization of global collab-
between any two trigger moments [57], [105], [106] oration and the design of pinning strategies are two key issues
of the pinning-based intelligent control.
i
tk+1 − tki ≥ τ > 0 (20) 1) Global Collaboration: Pinning control is widely used to
solve the synchronization of complex dynamic networks [58],
where τ is a positive constant that ensures the Zeno phe-
[59], which is also a special case of leader–follower consensus
nomenon is excluded.
problems defined in Problem 2 [53], [60]. The general first-
The work in [123] also looked at the issue, and proposed
order MASs composed of n agents under pinning control are
another method. That is, if Zeno’s behavior is assumed to exist,
modeled as
then there is at least a gathering point for the time-triggered
sequence. By verifying that this assumption contradicts the 
n
 
existing attributes of the system, it is proved that Zeno’s ẋi (t) = f (xi (t), t) + c aij  xj (t) − xi (t)
j=1
behavior can be excluded.
In addition to deal with the limited bandwidth of commu- + cdi (s(t) − xi (t)) (21)
nication networks, event-based strategies are used to solve where xi (t) ∈ Rm and f (xi (t), t) are the state and nonlinear
other problems subject to time delay [55], [56], [123], dynamic function of the ith agent, respectively. Coefficients
network attacks [22], [57], [74], switching topologies [9], and c and di are the coupling strength and pinning control gain,
multiplicative faults [47]. However, to the best of our knowl- respectively.  ∈ Rn×n represents the inner coupling matrix.
edge, there is a lack of fully distributed trigger strategies and s(t) ∈ Rm denotes the state of a pinner modeled as
methods to effectively exclude Zeno behavior, which does
not involve any global information (such as the number of ṡ(t) = f (s(t), t)
agents, Laplace matrix, etc.), especially for generalized linear
where f (s(t), t) is a nonlinear continuously differentiable func-
and complex nonlinear MASs.
tion related to dynamic characteristics of the pinner [21].
One way to ensure consensus for whole MASs is to design
C. Pinning-Based Intelligent Control suitable coupling strength, pinning control gain, and inner
The realization of information collaboration for MASs usu- coupling matrix based on the Laplacian matrix, including
ally requires the assumption of the original connectivity of the extension of the Laplacian matrix and the submatrix of

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
168 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

the Laplacian matrix (see [21] for details). However, this which can severely and adversely affect system performance.
approach depends on the properties of the global commu- For example, in a denial-of-service attack (DoS), the attacker
nication matrix. The global information is difficult to obtain intends to deny access to the data by making it unavailable to
when the Laplacian matrix is time varying or stochastic due systems. Recently, considerable efforts have been made based
to the random selection of pinned agents. In order to over- on resilient control to detect and defend against attacks for
come this drawback, adaptive pinning consensus strategies MASs in terms of consensus, formation control, and flocking
have been proposed for the first-order system without relying control defined in Problems 1–5.
on any global information. At the same time, by introducing A novel event-based resilient control was proposed in [61],
the adaptive rate in the coupling strength and control gain, which controlled the input signal rather than the state measure-
the conservativeness of the Laplacian-based method has been ment error to solve leaderless consensus in Problem 1 under
reduced [21], [59]. DoS attacks. For leader–follower consensus in Problem 2
In addition to the consensus problem, pinning-based control under DoS attacks, the work in [62] provided a distributed
is also extended to solve formation tracking in Problem 4 [88] fixed-time observer and an improved resilient observer to accu-
and flocking control in Problem 5 [107], [108]. A pinner or rately estimate the leader’s information, thus eliminating or
a virtual leader is used to provide the reference path for the weakening the influence of DoS. Reliable formation track-
agents to perform a collaborative task, such as the task of ing control for MAS under quantized communication and
multidrone formation to reach a designated location with the false data injection (FDI) attacks was investigated in [89]
desired trajectory. The relative formation reference respect for based on a distributed filter with adaptive attack compensator.
the virtual leader is also considered. The work in [88] provided Both the system reliability in the attacked case and original
a pinning-based control for nonlinear multidrone formation. performance in a no-attack case can be guaranteed with the
Although a pinner describes the desired path of MASs, agents developed filter. It can also achieve cooperative output regu-
usually do not strictly follow the pinner when they encounter lation of MAS when the communication is not quantized but
obstacles to avoid in the environment. In terms of the flocking with potential attacks. For the unbounded malicious attacks, a
problems defined in Problem 5, the pinner can be regarded fully distributed attack-resilient control protocol was proposed
as the flocking center with the desired trajectory. Pinning- in [90] to solve the time-varying formation tracking problem
based strategies for the flocking motion of a MAS have been defined in Problem 4. The bounded system stability and uni-
developed under switching topologies [107] and sampled-data formly ultimately bounded synchronization performance have
frameworks [108] to minimize the total cost considering pinner been guaranteed. Considering the presence of noncooperative
tracking, velocity consensus, and obstacle avoidance functions. robots, Saulnier et al. [109] developed a resilient flocking con-
2) Pinning Strategies: They are investigated to determine trol approach for Problem 5. The proposed dynamic connec-
the minimum number of nodes to be controlled and the spe- tivity management and switching control strategies restricted
cific pinned agents. Existing pinning strategies mainly include the communication topology within the resilient threshold and
the random pinning strategy and the specific pinning strategy. allowed the mobile robots to achieve consensus along with the
The strategy of random selection first searches for strongly motion.
connected components [21] to assign groups of agents, and To the best of our knowledge, there are still a lack of effec-
then randomly selects one agent in the group for control. tive detection and deference theoretical frameworks for MASs
Another method is to first arrange the nodes in descending under multiple attacks.
order according to the difference between the out-degree and
the in-degree, and select the first l nodes for control [60]. Both
methods need to verify that the pinner is the root node, and E. Intelligent Control for Homogeneous MASs
there is at least one directed spanning tree in the entire MASs. MASs can be divided into homogeneous and heterogeneous
It is difficult to ensure the connectivity of the original network, systems, depending on whether the system dynamics are the
especially for arbitrary or changing topologies. Through a rea- same or not. In addition to the limitations of the interaction
sonable selection of pinned agents, the connectivity under the level, the constraints of the system level include nonlin-
dynamic topology can be guaranteed at each moment. Many ear dynamics, heterogeneous dynamics, system uncertainties,
works have pointed out that the specific pinning strategy is external interference, and actuator and sensor failures.
more effective than the random pinning strategy in reducing For linear homogeneous MASs, there are a large number
the number of pinned agents [58], [60], [107]. of works on consensus Problems 1 and 2 [38], [39], [63],
Pinning-based control has been also introduced into some formation control Problems 3 and 4 [91], [92], and flock-
methodologies in control theory, such as the impulsive con- ing control Problem 5 [110], [111]. However, in practical
trol [124], robust H∞ control [125] and finite-time con- applications, an agent is always subject to nonlinear dynam-
trol [126] to improve the system performance under time delay ics, such as the fight control for multidrone formation [88]
and disturbances. and flocking control for robots [102]. Various nonlinear con-
trol approaches have been developed for nonlinear MASs
under uncertainties and bounded external disturbances, includ-
D. Resilient Control ing adaptive control [64], backstepping scheme [88], sliding
MASs are likely to suffer from malicious attacks and cor- mode control [7], [93], neural network [27], [93], and fuzzy
ruption of sensory data or manipulation of actuators inputs, control [115]. Sliding mode control has been widely used to

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
SHI AND YAN: A SURVEY ON INTELLIGENT CONTROL FOR MASs 169

control nonlinear systems with uncertainties and unknown dis-

turbances, as the controllers can be designed to compensate
for the uncertainties and disturbances.
Unforeseen threats may occur in system components, such
as sensors, actuators, and controllers. Passive fault-tolerant
control is one of the popular methods for managing phys-
ical faults or damages, with adaptive strategies. The work
in [127] investigated a H∞ consensus protocol together with
an adaptive compensator to tolerate sensor and actuator faults Fig. 10. Output regulation-based intelligent control for heterogeneous MASs.
for Problem 1. In [128], a cooperative adaptive fault-tolerant
fuzzy control was proposed to solve leader–follower consen-
sus Problem 2 of networked MASs with time-varying actuator the output consensus of heterogeneous MASs is solvable.
faults. Adaptive formation control laws for Problem 4 were Subsequently, the approach was promoted to solve the prob-
designed for unmanned aerial vehicles (UAVs) to tolerate actu- lems of formation control Problems 3 and 4 [79], [82], [133]
ator faults [129]. As a typical method of fault-tolerant control, and flocking control Problem 5 [118].
adaptive controllers are used for compensating physical faults A general intelligent control framework of heterogeneous
and passively tolerate system failures without changing the MASs is shown in Fig. 10, where the output of exosystem
controller structure. However, the fixed structure has conser- v includes reference information that needs to be tracked and
vative problems and cannot optimize the system performance. disturbances need to be rejected. Since only a part of the agents
To overcome the shortcomings, active fault-tolerant control can be informed by the exosystem, an observer is designed
strategies have been proposed for MASs. The active strate- to estimate the output of the exosystem v̂ through collabora-
gies usually include two functions: 1) fault detection and tive error φi . The intelligent control problem of heterogeneous
2) control reconfiguration [128], [130], [131]. However, most MASs is transformed into the stabilization problem of dis-
of the existing research results assume that the system will tributed augmented systems by a regulator. The regulator is
not diverge during the time period of fault detection. This composed of a stabilizer and p-copy internal model to stabi-
assumption is limited to multiple failures and severe physical lize the system and compensate for uncertainties, receptively.
damage. Augmented systems consist of a real agent and an internal
model compensator. This method fundamentally uncouples the
dynamics of heterogeneous agents and realizes the regulation
F. Intelligent Control for Heterogeneous MASs of exosystem in a distributed manner.
During the past decades, research on intelligent control has A unified uncertain heterogeneous MAS with n agents of
gradually shifted from homogeneous MASs to heterogeneous different orders can be modeled as
MASs. One reason is that it is difficult to equip the same
agents with all the necessary sensing and computing equip- ẋi = Awi xi + Bwi ui + Ewi v
ment. Even for the same agents with the same equipment, it yi = Cwi xi + Dwi ui + Fwi v, i = 1, 2, . . . , n (22)
is not truly homogeneous MASs due to asynchronous clocks
where xi ∈ Rni , ui ∈ Rmi , and yi ∈ Rpi represent the
and uncertainties. Most important of all, heterogeneous MASs
state, input, and output variables of the ith agent, respectively.
have more advantages in terms of formation flexibility and
Matrices wi = i +i are the corresponding system matrices
complex task decomposition because of the different functions
with uncertainties i , where  = A, B, C, D, E, F.
of individuals.
The dynamic of an exosystem can be described as follows:
The primary research on heterogeneous MASs is focused on
hybrid-order MASs, such as consensus Problem 1 of first-order v̇ = Sv (23)
and second-order hybrid MASs [66] and formation Problem 3
of first-order and fourth-order hybrid MASs [77]. Graph where v ∈ Rn0 represent the state/output variables of the
theory-based matrix methods and Lyapunov theory are used exosystem, which contains the reference information accord-
to solve intelligent problems of hybrid-order MASs, but the ing to different intelligent Problems 1–5. For example, the
global communication matrix is required especially for high- dynamic of a time-varying formation [82] is considered as ḟi =
order hybrid MASs. Another approach for hybrid-order MASs Aif fi , and the signal of exosystem v includes the information
aims to convert the mixed-order MASs into the same-order of virtual leader x0 , time-varying formation dynamics fi , and
MASs and achieve the state consensus of the corresponding disturbance ω.
order by adding virtual zero states [30]. However, expanding In order to achieve intelligent collaboration, the general
the dimensions of agents may increase the computational load. stabilizer and internal model compensator are designed as
State consensus is almost impossible for the general heteroge- follows:
neous systems with different orders and different dynamics. ui = Kix xi + Kiz zi
Therefore, the output regulation for a single system is the
żi = 1 zi + 2 ei (24)
focus, which can be extended to solve the output consensus
problem of heterogeneous MASs [65]. Under the assumption where zi is the state of the dynamic compensator. ( 1 , 2 ) is
that there exists a solution to the regulation equation [132], the p-copy internal model pair of S, based on the internal

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
170 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

TABLE IV
A PPLICATIONS OF I NTELLIGENT C ONTROL FOR MAS S

collaboration system for search and rescue missions. The

framework has been verified and evaluated by outdoor exper-
Fig. 11. Applications of intelligent control for MASs. iments of four prototype UAVs. A multiUAV collaborative
approach in disaster management and civil security appli-
cations has been validated with real UAVs and wireless
model principle [132]. The control law Ki = (Kix , Kiz ) is sensor networks [136]. Formation protocols and consensus
obtained by the Lyapunov stability theory [82], [132]. The approaches were used to achieve time-varying formations,
controller aims to eliminate regulation error, which is defined which were tested and verified by distributed outdoor exper-
according to different intelligent control Problems 1–5. For iments with five quadrotors [138]. Relative information of
example, for the leader–follower consensus in Problem 2, the neighboring UAVs can be used to construct a time-varying
error is defined as ei = yi − v̂; for time-varying formation formation control protocol for swarm systems. An outdoor
tracking in Problem 4 [79], [82], ei = yi − v̂ − fi . target enclosing experiment was carried out for three follower
It should be noted that the observer converts the collabora- quadrotor UAVs to enclose a leader quadrotor UAV by time-
tive error φi to regulation error ei based on the observation v̂. varying formations [139]. In [140], a life support robot system
Therefore, the design of observers is the core of decoupling was developed to perform domestic services that are useful
heterogeneous dynamics and realizing distributed architec- to the well being of the elderly with walking disabilities.
ture. In [82], the adaptive distributed observer was proposed A frequent task during life support is the fetching of daily
under the assumption that the exosystem dynamics S were containers, such as serving drinks and food [140].
globally known. Note that the design of the p-copy internal Furthermore, a large number of results on the forma-
model system also required a known exosystem dynamic tion control for hardware multivehicle platforms have been
matrix. Model-based adaptive controls still involve global verified under laboratory conditions. A distributed forma-
information, until an off-policy reinforcement learning strategy tion control approach for multirotor UAVs was proposed and
was investigated to make up the disadvantages [134]. embedded into the onboard computational units to make them
The intelligent control for linear heterogeneous systems able to keep a balanced formation in 2-D and three dimen-
can also be extended to nonlinear systems to deal with sional environments [137]. The proposed formation control
consensus problems [30], [67], [68], formation control approach was proved to be feasible for arbitrary formation
problems [79], [80], [94]–[96], and flocking control prob- by both simulations and real-system experiment. In order to
lems [120]. It is worth noting that this method is only lead the UGVs moving into the desired formation quickly, a
applicable to leader-involved Problems 2, 4, and 5. For the cooperative coevolutionary algorithm-based distributed model
leaderless problems as defined in Problems 1 and 3, it will predictive control was proposed in [144], which can greatly
raise a big challenge to discover the common internal model improve the performance of formation control as performed
in heterogeneous systems in the absence of the exosystem. by three mobile robots. An adaptive self-organizing map neu-
Learning-based approaches may be a possible way to address ral network was applied to keep the formation when agents
the problems. move along the desired path [145]. Both simulations and real
AUV systems demonstrated the fault-tolerant characteristic in
IV. A PPLICATIONS OF I NTELLIGENT C ONTROL FOR MAS S obstacle avoidance and the benefit of balancing the workload
In this section, we present the achievements of intelligent and energy. A modified constrained adaptive controller was
control for MAS applications in the following directions: proposed to resolve the communication delay and actuator sat-
robotics, computer networks, transportation, and others. A uration [146]. Simulation and experimental validation showed
summary of these applications is outlined in Fig. 11 and that the method can effectively compensate for the effects of
Table IV. state delay in 2 and 5 s, respectively. Follower AUVs were
able to follow the desired path within the accuracy of 5 cm.
A. Robotics
The application of MASs on robotics has received exten- B. Complex Networks
sive attention, especially on UAVs, unmanned ground vehicles The resource-aware consensus theory of the MASs provides
(UGVs), and autonomous underwater vehicles (AUVs). The a theoretical reference for smart grid applications. In [147],
work in [135] proposed a modular architecture of multiUAV the MAS-based algorithm has been applied to control voltage

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
SHI AND YAN: A SURVEY ON INTELLIGENT CONTROL FOR MASs 171

and capacitor to optimally set the system. An event-triggered V. C ONCLUSION AND F UTURE R ESEARCH C HALLENGES
strategy is proved to be an effective method to reduce the In this article, we presented a survey of distributed intel-
communication burden of a network. The two-level reinforce- ligent control for MASs. Focusing on the constraints from
ment learning-based controller was proposed in [148], where the interaction level and system level, the recent results have
the parameters were optimized by particle swarm. The results been reviewed in terms of consensus problem, formation con-
verified the feasibility of the proposed method. Within the trol problem, and flocking control problem. However, this is
decentralized system integrity protection set up, data-driven far from an exhaustive literature review and some important
anomaly detection, and adaptive load rejection were studied results might be missed due to the limitation of our knowl-
in [149]. Anomaly detection has been converted to a multiclas- edge. Furthermore, there still exist several challenges in this
sification problem and can be performed by individual agents, area deserving further study.
but all the interconnected agents devoted to the final decision. 1) Security is highly challenging for MASs. Most existing
Meanwhile, the proposed adaptive load rejection strategy can works design resilient and robust strategies separately
reduce the DoS attacks. on interaction level and system level. For instance,
The application of MASs can be extended to the Internet distributed resilient control under attacks and commu-
of Things (IoT) where objects range from sensors to wearable nication problems tends to use network-level design,
devices. Agent-based resilient control plays a vital role in IoT where the individual agent with high-fidelity dynam-
networks. It is essential to effectively identify the malicious ics is usually simplified, and fault-tolerance control
node and prevent further damage. A combined multiagent mainly focuses on homogeneous system-level robust-
and multilayered game formulation was proposed in [150], ness. However, the separated security control design on
which incorporated a trust model to assess the node/object. two levels fails to realize quick stability and recover
The proposed model can significantly improve the accuracy of to optimal performance, which poses a threat to the
intrusion detection by experimental test. IoT inevitably intro- survival of MASs under multiple threats and unknown
duces a vast amount of real-time data. A multiagent-based real- environments. Therefore, high-reliability intelligent con-
time scheduling architecture was presented to optimally assign trol under both two-level threats is still an open problem
tasks according to the real-time status of machines [151]. to be solved in the future.
2) The design of fully distributed intelligent control and
C. Transportation its optimization is still considered as open issues.
The large-scale intelligent transportation system is one of Although many studies have focused on distributed con-
the typical applications of MASs. Taking the dimension, trol approaches, some global information, such as the
complicated dynamics, and uncertainties into consideration, total number of the agents and the Laplace matrix of
Lin et al. [152] proposed a centralized multiagent control the communication topology are still being involved
method with a serial framework. Agents communicate with their for high-order MASs in intelligent control designs.
neighbors through a model-based predictive control method. Verification on global stability, connectivity preserva-
Using the traffic data provided by the city of Toronto, an adap- tion under dynamic topology, proof of nonZeno phe-
tive reinforcement learning-based traffic signal controller was nomenon, and optimization of task assignments are
proposed in [153], which can work in two modes: 1) decentral- usually not designed with a fully distributed frame-
ized and 2) centralized. However, the dynamic and complex work. Adaptive control strategies and learning-based
traffic conditions make it difficult for the model-based and techniques are used to resolve this imperfect. However,
reinforcement learning-based models to make good decisions. they inevitably increase the computational load. In real-
In [154], a multiagent recurrent deep deterministic policy gra- world applications, agents subject to limited computing
dient algorithm was proposed to control the traffic light in land capability need to perceive, make decisions, and take
traffic. Decisions were made independently by each agent, thus actions independently, which raises higher requirements
avoiding the poor performance caused by an unstable environ- for fully distributed algorithms and optimization tech-
ment. Autonomous driving is another application in intelligent niques.
control, among the key complex problems, the formation will be 3) The research on intelligent control for heterogeneous
outstanding. Due to the formation changes with the traffic flow MASs should be enhanced, especially in both theoretic
and conditions, a dynamic coordination graph was proposed research and applications of heterogeneous multive-
to model the constantly changing topology to coordinate the hicle systems. Although the works on heterogeneous
maneuvers of grouped vehicles in [155], which was proved to MASs have received extensive attention in the past
be effective than some expert rules. decades, most of them focused on the fundamental
consensus problems and theoretic research. In fact,
it is difficult to build truly homogeneous MASs in
D. Others practical applications. The intelligent control of hetero-
The applications of MASs are not limited to the above- geneous vehicles, such as UAVs, UGVs, and AUVs,
mentioned fields. They have also been widely applied to is more promising in practical applications to achieve
aerospace, agriculture, industrial production, and medical multidimensional collaboration under complementary
treatment, to name but a few. capabilities.

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
172 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

4) Verifications for distributed intelligent control strate- [22] D. Zhang, Y. Tang, Z. Ding, and F. Qian, “Event-based resilient forma-
gies in large-scale practical application scenarios are tion control of multiagent systems,” IEEE Trans. Cybern., early access,
Apr. 26, 2019, doi: 10.1109/TCYB.2019.2910614.
urgently needed. Most existing results are obtained [23] W. Ren and R. W. Beard, “Consensus seeking in multiagent systems
under laboratory conditions with centralized structures. under dynamically changing interaction topologies,” IEEE Trans.
For large-scale MASs, very few studies are carried out Autom. Control, vol. 50, no. 5, pp. 655–661, May 2005.
[24] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and coop-
in the actual application environment, which leads to eration in networked multi-agent systems,” Proc. IEEE, vol. 95, no. 1,
the urgent requirements of verifications in the actual pp. 215–233, Jan. 2007.
application environment. [25] R. Olfati-Saber, “Flocking for multi-agent dynamic systems:
Algorithms and theory,” IEEE Trans. Autom. Control, vol. 51, no. 3,
pp. 401–420, Mar. 2006.
R EFERENCES [26] G. Song, P. Shi, S. Wang, and J.-S. Pan, “A new finite-time cooperative
control algorithm for uncertain multi-agent systems,” Int. J. Syst. Sci.,
[1] M. Wooldridge, An Introduction to Multiagent Systems. Hoboken, NJ, vol. 50, no. 5, pp. 1006–1016, 2019.
USA: Wiley, 2009. [27] J. Ni and P. Shi, “Adaptive neural network fixed-time leader–
[2] Y. Cao, W. Yu, W. Ren, and G. Chen, “An overview of recent progress follower consensus for multiagent systems with constraints and
in the study of distributed multi-agent coordination,” IEEE Trans. Ind. disturbances,” IEEE Trans. Cybern., early access, Feb. 24, 2020,
Informat., vol. 9, no. 1, pp. 427–438, Feb. 2013. doi: 10.1109/TCYB.2020.2967995.
[3] J. Qin, Q. Ma, Y. Shi, and L. Wang, “Recent advances in consensus [28] P. Shi and J. Yu, “Dissipativity-based consensus for fuzzy multi-agent
of multi-agent systems: A brief survey,” IEEE Trans. Ind. Electron., systems under switching directed topologies,” IEEE Trans. Fuzzy Syst.,
vol. 64, no. 6, pp. 4972–4983, Jun. 2017. early access, Jan. 27, 2020, doi: 10.1109/TFUZZ.2020.2969391.
[4] M. Brambilla, E. Ferrante, M. Birattari, and M. Dorigo, “Swarm [29] L. Ji, T. Gao, and X. Liao, “Couple-group consensus for cooperative-
robotics: A review from the swarm engineering perspective,” Swarm competitive heterogeneous multiagent systems: Hybrid adaptive and
Intell., vol. 7, no. 1, pp. 1–41, 2013. pinning methods,” IEEE Trans. Syst., Man, Cybern., Syst., early access,
[5] J. Kennedy, “Swarm intelligence,” in Handbook of Nature-Inspired and Nov. 15, 2019, doi: 10.1109/TSMC.2019.2951787.
Innovative Computing. Boston, MA, USA: Springer, 2006, pp. 187– [30] X. Li, P. Shi, and Y. Wang, “Distributed cooperative adaptive tracking
219. control for heterogeneous systems with hybrid nonlinear dynamics,”
[6] K.-K. Oh, M.-C. Park, and H.-S. Ahn, “A survey of multi-agent Nonlinear Dyn., vol. 95, no. 3, pp. 2131–2141, 2019.
formation control,” Automatica, vol. 53, pp. 424–440, Mar. 2015. [31] X. Li and P. Shi, “Cooperative fault-tolerant tracking control of hetero-
[7] Q. Shen, P. Shi, and Y. Shi, “Distributed adaptive fuzzy control for geneous hybrid-order mechanical systems with actuator and amplifier
nonlinear multiagent systems via sliding mode observers,” IEEE Trans. faults,” Nonlinear Dyn., vol. 98, no. 1, pp. 447–462, 2019.
Cybern., vol. 46, no. 12, pp. 3086–3097, Dec. 2016. [32] B. Yan, C. Wu, and P. Shi, “Formation consensus for discrete-time het-
[8] P. Shi and Q. Shen, “Observer-based leader-following consensus of erogeneous multi-agent systems with link failures and actuator/sensor
uncertain nonlinear multi-agent systems,” Int. J. Robust Nonlinear faults,” J. Franklin Inst., vol. 356, pp. 6547–6570, Aug. 2019.
Control, vol. 27, no. 17, pp. 3794–3811, 2017. [33] Y. Su and J. Huang, “Cooperative robust output regulation of a class
[9] Z.-G. Wu, Y. Xu, R. Lu, Y. Wu, and T. Huang, “Event-triggered control of heterogeneous linear uncertain multi-agent systems,” Int. J. Robust
for consensus of multiagent systems with fixed/switching topologies,” Nonlinear Control, vol. 24, no. 17, pp. 2819–2839, 2014.
IEEE Trans. Syst., Man, Cybern., Syst., vol. 48, no. 10, pp. 1736–1746,
[34] Y. Yang, D. Yue, and C. Dou, “Distributed adaptive output consensus
Oct. 2018.
control of a class of heterogeneous multi-agent systems under switching
[10] S. Rao and D. Ghose, “Sliding mode control-based autopilots for lead-
directed topologies,” Inf. Sci., vol. 345, pp. 294–312, Jun. 2016.
erless consensus of unmanned aerial vehicles,” IEEE Trans. Control
[35] Y.-W. Wang, X.-K. Liu, J.-W. Xiao, and Y. Shen, “Output formation-
Syst. Technol., vol. 22, no. 5, pp. 1964–1972, Sep. 2014.
containment of interacted heterogeneous linear systems by distributed
[11] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of
hybrid active control,” Automatica, vol. 93, pp. 26–32, Jul. 2018.
agents with switching topology and time-delays,” IEEE Trans. Autom.
Control, vol. 49, no. 9, pp. 1520–1533, Sep. 2004. [36] S. Zheng, P. Shi, S. Wang, and Y. Shi, “Adaptive neu-
[12] Z. Han, K. Guo, L. Xie, and Z. Lin, “Integrated relative localization ral control for a class of nonlinear multiagent systems,” IEEE
and leader–follower formation control,” IEEE Trans. Autom. Control, Trans. Neural Netw. Learn. Syst., early access, Mar. 30, 2020,
vol. 64, no. 1, pp. 20–34, Jan. 2019. doi: 10.1109/TNNLS.2020.2979266.
[13] X. Li, X. Dong, Q. Li, and Z. Ren, “Event-triggered time-varying for- [37] B. Bollobás, Modern Graph Theory, vol. 184. Berlin, Germany:
mation control for general linear multi-agent systems,” J. Franklin Inst., Springer, 2013.
vol. 356, no. 17, pp. 10179–10195, 2019. [38] B. Cheng and Z. Li, “Fully distributed event-triggered protocols for lin-
[14] K. D. Do, “Flocking for multiple elliptical agents with limited com- ear multiagent networks,” IEEE Trans. Autom. Control, vol. 64, no. 4,
munication ranges,” IEEE Trans. Robot., vol. 27, no. 5, pp. 931–942, pp. 1655–1662, Apr. 2019.
Oct. 2011. [39] R. Sakthivel, R. Sakthivel, B. Kaviarasan, H. Lee, and Y. Lim, “Finite-
[15] A. Weinstein, A. Cho, G. Loianno, and V. Kumar, “Visual inertial time leaderless consensus of uncertain multi-agent systems against
odometry swarm: An autonomous swarm of vision-based quadrotors,” time-varying actuator faults,” Neurocomputing, vol. 325, pp. 159–171,
IEEE Robot. Autom. Lett., vol. 3, no. 3, pp. 1801–1807, Jul. 2018. Jan. 2019.
[16] F. Schilling, J. Lecoeur, F. Schiano, and D. Floreano, “Learning vision- [40] J. Huang, W. Wang, C. Wen, J. Zhou, and G. Li, “Distributed adaptive
based flight in drone swarms by imitation,” IEEE Robot. Autom. Lett., leader–follower and leaderless consensus control of a class of strict-
vol. 4, no. 4, pp. 4523–4530, Oct. 2019. feedback nonlinear systems: A unified approach,” Automatica, vol. 118,
[17] L. Consolini, F. Morbidi, D. Prattichizzo, and M. Tosques, “Leader– Aug. 2020, Art. no. 109021.
follower formation control of nonholonomic mobile robots with input [41] Y. Lv, G. Wen, T. Huang, and Z. Duan, “Adaptive attack-free pro-
constraints,” Automatica, vol. 44, no. 5, pp. 1343–1349, 2008. tocol for consensus tracking with pure relative output information,”
[18] L. Ding, Q.-L. Han, and G. Guo, “Network-based leader-following Automatica, vol. 117, Jul. 2020, Art. no. 108998.
consensus for distributed multi-agent systems,” Automatica, vol. 49, [42] M. Lu and J. Huang, “Internal model approach to cooperative robust
no. 7, pp. 2281–2286, 2013. output regulation for linear uncertain time-delay multiagent systems,”
[19] H. Yu, P. Shi, and C. C. Lim, “Robot formation control in stealth mode Int. J. Robust Nonlinear Control, vol. 28, no. 6, pp. 2528–2542, 2018.
with scalable team size,” Int. J. Control, vol. 89, no. 11, pp. 2155–2168, [43] Q. Shen, P. Shi, J. Zhu, S. Wang, and Y. Shi, “Neural networks-
2016. based distributed adaptive control of nonlinear multiagent systems,”
[20] D. V. Dimarogonas, E. Frazzoli, and K. H. Johansson, “Distributed IEEE Trans. Neural Netw. Learn. Syst., vol. 31, no. 3, pp. 1010–1021,
event-triggered control for multi-agent systems,” IEEE Trans. Autom. Mar. 2020.
Control, vol. 57, no. 5, pp. 1291–1297, May 2012. [44] Q. Song, F. Liu, J. Cao, and W. Yu, “M-matrix strategies for pinning-
[21] W. Xing, P. Shi, R. K. Agarwal, and Y. Zhao, “A survey on global pin- controlled leader-following consensus in multiagent systems with non-
ning synchronization of complex networks,” J. Franklin Inst., vol. 356, linear dynamics,” IEEE Trans. Cybern., vol. 43, no. 6, pp. 1688–1697,
no. 6, pp. 3590–3611, 2019. Dec. 2013.

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
SHI AND YAN: A SURVEY ON INTELLIGENT CONTROL FOR MASs 173

[45] Q. Ma, Z. Wang, and G. Miao, “Second-order group consensus for [68] S. Li, M. J. Er, and J. Zhang, “Distributed adaptive fuzzy control
multi-agent systems via pinning leader-following approach,” J. Franklin for output consensus of heterogeneous stochastic nonlinear multiagent
Inst., vol. 351, no. 3, pp. 1288–1300, 2014. systems,” IEEE Trans. Fuzzy Syst., vol. 26, no. 3, pp. 1138–1152,
[46] L. Li, P. Shi, and C. K. Ahn, “Distributed iterative FIR consensus filter Jun. 2018.
for multiagent systems over sensor networks,” IEEE Trans. Cybern., [69] W. Zou, C. K. Ahn, and Z. Xiang, “Fuzzy-approximation-based dis-
early access, Dec. 9, 2020, doi: 10.1109/TCYB.2020.3035866. tributed fault-tolerant consensus for heterogeneous switched nonlinear
[47] D. Ye, M.-M. Chen, and H.-J. Yang, “Distributed adaptive event- multiagent systems,” IEEE Trans. Fuzzy Syst., early access, Jul. 16,
triggered fault-tolerant consensus of multiagent systems with general 2020, doi: 10.1109/TFUZZ.2020.3009730.
linear dynamics,” IEEE Trans. Cybern., vol. 49, no. 3, pp. 757–767, [70] W. Ren, “Consensus strategies for cooperative control of vehicle for-
Mar. 2019. mations,” IET Control Theory Appl., vol. 1, no. 2, pp. 505–512,
[48] C. Gong, G. Zhu, and P. Shi, “Adaptive event-triggered and double- Mar. 2007.
quantized consensus of leader–follower multiagent systems with semi- [71] S. He, M. Wang, S.-L. Dai, and F. Luo, “Leader–follower formation
Markovian jump parameters,” IEEE Trans. Syst., Man, Cybern., Syst., control of USVs with prescribed performance and collision avoidance,”
early access, Dec. 20, 2019, doi: 10.1109/TSMC.2019.2957530. IEEE Trans. Ind. Informat., vol. 15, no. 1, pp. 572–581, Jan. 2019.
[49] R. Yang, H. Zhang, G. Feng, H. Yan, and Z. Wang, “Robust cooperative [72] G. Lee and D. Chwa, “Decentralized behavior-based formation control
output regulation of multi-agent systems via adaptive event-triggered of multiple robots considering obstacle avoidance,” Intell. Serv. Robot.,
control,” Automatica, vol. 102, pp. 129–136, Apr. 2019. vol. 11, no. 1, pp. 127–138, Jan. 2018.
[50] H. Zhang, J. Chen, Z. Wang, C. Fu, and S. Song, “Distributed [73] D. Zhou, Z. Wang, and M. Schwager, “Agile coordination and assis-
event-triggered control for cooperative output regulation of multiagent tive collision avoidance for quadrotor swarms using virtual structures,”
systems with an online estimation algorithm,” IEEE Trans. Cybern., IEEE Trans. Robot., vol. 34, no. 4, pp. 916–923, Aug. 2018.
early access, Jun. 8, 2020, doi: 10.1109/TCYB.2020.2991761. [74] Y. Tang, D. Zhang, P. Shi, W. Zhang, and F. Qian, “Event-
[51] B. Hu, Z.-H. Guan, and M. Fu, “Distributed event-driven control for based formation control for multi-agent systems under DoS
finite-time consensus,” Automatica, vol. 103, pp. 88–95, May 2019. attacks,” IEEE Trans. Autom. Control, early access, Mar. 10, 2020,
[52] W. Zhu, Z.-P. Jiang, and G. Feng, “Event-based consensus of multi- doi: 10.1109/TAC.2020.2979936.
agent systems with general linear models,” Automatica, vol. 50, no. 2, [75] H. Yu, P. Shi, and C. C. Lim, “Scalable formation control in stealth
pp. 552–558, 2014. with limited sensing range,” Int. J. Robust Nonlinear Control, vol. 27,
[53] Z. Wu, Y. Xu, Y. Pan, P. Shi, and Q. Wang, “Event-triggered pin- no. 3, pp. 410–433, 2017.
ning control for consensus of multiagent systems with quantized [76] Y. Liu, H. Yu, P. Shi, and C.-C. Lim, “Formation control and colli-
information,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 48, no. 11, sion avoidance for a class of multi-agent systems,” J. Franklin Inst.,
pp. 1929–1938, Nov. 2018. vol. 356, no. 10, pp. 5395–5420, 2019.
[54] W. Zou, P. Shi, Z. Xiang, and Y. Shi, “Consensus tracking control of
[77] B. Yan, P. Shi, C.-C. Lim, C. Wu, and Z. Shi, “Optimally dis-
switched stochastic nonlinear multiagent systems via event-triggered
tributed formation control with obstacle avoidance for mixed-order
strategy,” IEEE Trans. Neural Netw. Learn. Syst., vol. 31, no. 3,
multi-agent systems under switching topologies,” IET Control Theory
pp. 1036–1045, Mar. 2020.
Appl., vol. 12, no. 13, pp. 1853–1863, Sep. 2018.
[55] C. Deng, M. J. Er, G.-H. Yang, and N. Wang, “Event-triggered con-
[78] X. Dong, Y. Hua, Y. Zhou, Z. Ren, and Y. Zhong, “Theory and
sensus of linear multiagent systems with time-varying communication
experiment on formation-containment control of multiple multirotor
delays,” IEEE Trans. Cybern., vol. 50, no. 7, pp. 2916–2925, Jul. 2020.
unmanned aerial vehicle systems,” IEEE Trans. Autom. Sci. Eng.,
[56] X. Zhou, P. Shi, C.-C. Lim, C. Yang, and W. Gui, “Event based guar-
vol. 16, no. 1, pp. 229–240, Jan. 2019.
anteed cost consensus for distributed multi-agent systems,” J. Franklin
Inst., vol. 352, no. 9, pp. 3546–3563, 2015. [79] S. Li, J. Zhang, X. Li, F. Wang, X. Luo, and X. Guan, “Formation
control of heterogeneous discrete-time nonlinear multi-agent systems
[57] Y. Xu, M. Fang, P. Shi, and Z. Wu, “Event-based secure consensus of
with uncertainties,” IEEE Trans. Ind. Electron., vol. 64, no. 6,
mutiagent systems against DoS attacks,” IEEE Trans. Cybern., vol. 50,
pp. 4730–4740, Jun. 2017.
no. 8, pp. 3468–3476, Aug. 2020.
[58] X. F. Wang and G. Chen, “Pinning control of scale-free dynamical [80] C. Yuan, S. Licht, and H. He, “Formation learning control of multiple
networks,” Physica A, Stat. Mech. Appl., vol. 310, nos. 3–4, autonomous underwater vehicles with heterogeneous nonlinear uncer-
pp. 521–531, 2002. tain dynamics,” IEEE Trans. Cybern., vol. 48, no. 10, pp. 2920–2934,
[59] H. Chen, P. Shi, and C. Lim, “Cluster synchronization for neutral Oct. 2018.
stochastic delay networks via intermittent adaptive control,” IEEE [81] C. Deng and W.-W. Che, “Fault-tolerant fuzzy formation control for
Trans. Neural Netw. Learn. Syst., vol. 30, no. 11, pp. 3246–3259, a class of nonlinear multiagent systems under directed and switching
Nov. 2019. topology,” IEEE Trans. Syst., Man, Cybern., Syst., early access, Dec. 5,
[60] Q. Song, J. Cao, and W. Yu, “Second-order leader-following consensus 2019, doi: 10.1109/TSMC.2019.2954870.
of nonlinear multi-agent systems via pinning control,” Syst. Control [82] W. Jiang, G. Wen, Z. Peng, T. Huang, and A. Rahmani, “Fully
Lett., vol. 59, no. 9, pp. 553–562, Sep. 2010. distributed formation-containment control of heterogeneous linear
[61] Y. Xu, M. Fang, Z. Wu, Y. Pan, M. Chadli, and T. Huang, “Input- multiagent systems,” IEEE Trans. Autom. Control, vol. 64, no. 9,
based event-triggering consensus of multiagent systems under denial- pp. 3889–3896, Sep. 2019.
of-service attacks,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 50, [83] X. Ge and Q.-L. Han, “Distributed formation control of networked
no. 4, pp. 1455–1464, Apr. 2020. multi-agent systems using a dynamic event-triggered communica-
[62] H. Yang and D. Ye, “Observer-based fixed-time secure track- tion mechanism,” IEEE Trans. Ind. Electron., vol. 64, no. 10,
ing consensus for networked high-order multiagent systems against pp. 8118–8127, Oct. 2017.
DoS attacks,” IEEE Trans. Cybern., early access, Jul. 22, 2020, [84] Y. Liu, P. Shi, and C. C. Lim, “Collision-free formation control for
doi: 10.1109/TCYB.2020.3005354. multi-agent systems with dynamic mapping,” IEEE Trans. Circuits Syst.
[63] X. Ai, S. Song, and K. You, “Second-order consensus of multi- II, Exp. Briefs, vol. 67, no. 10, pp. 1984–1988, Oct. 2020.
agent systems under limited interaction ranges,” Automatica, vol. 68, [85] H. Yu, P. Shi, C.-C. Lim, and Y. Liu, “Probability-triggered for-
pp. 329–333, Jun. 2016. mation control with adaptive roles,” Int. J. Control, vol. 93, no. 8,
[64] Q. Shen, P. Shi, J. Zhu, and L. Zhang, “Adaptive consensus control pp. 1989–2000, 2020.
of leader-following systems with transmission nonlinearities,” Int. J. [86] Y. Liu, P. Shi, H. Yu, and C. C. Lim, “Event-triggered probability-
Control, vol. 92, no. 2, pp. 317–328, 2019. driven adaptive formation control for multiple elliptical agents,”
[65] Y. Su and J. Huang, “Cooperative output regulation of linear IEEE Trans. Syst., Man, Cybern., Syst., early access, Oct. 16, 2020,
multi-agent systems,” IEEE Trans. Autom. Control, vol. 57, no. 4, doi: 10.1109/TSMC.2020.3026029.
pp. 1062–1066, Apr. 2012. [87] Q. Xiao, F. L. Lewis, and Z. Zeng, “Event-based time-interval pinning
[66] S.-L. Du, W. Xia, X.-M. Sun, and W. Wang, “Sampled-data-based con- control for complex networks on time scales and applications,” IEEE
sensus and L2 -gain analysis for heterogeneous multiagent systems,” Trans. Ind. Electron., vol. 65, no. 11, pp. 8797–8808, Nov. 2018.
IEEE Trans. Cybern., vol. 47, no. 6, pp. 1523–1531, Jun. 2017. [88] Y. Kartal, K. Subbarao, N. R. Gans, A. Dogan, and F. Lewis,
[67] W. Zhang, D. W. Ho, Y. Tang, and Y. Liu, “Quasi-consensus of “Distributed backstepping based control of multiple UAV formation
heterogeneous-switched nonlinear multiagent systems,” IEEE Trans. flight subject to time delays,” IET Control Theory Appl., vol. 14, no. 12,
Cybern., vol. 50, no. 7, pp. 3136–3146, Jul. 2020. pp. 1628–1638, Aug. 2020.

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
174 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 1, JANUARY 2021

[89] X. Huang and J. Dong, “Reliable leader-to-follower formation control [112] W. Zhao, H. Chu, M. Zhang, T. Sun, and L. Guo, “Flocking control
of multiagent systems under communication quantization and attacks,” of fixed-wing UAVs with cooperative obstacle avoidance capability,”
IEEE Trans. Syst., Man, Cybern., Syst., vol. 50, no. 1, pp. 89–99, IEEE Access, vol. 7, pp. 17798–17808, 2019.
Jan. 2020. [113] T. Han and S. S. Ge, “Styled-velocity flocking of autonomous vehicles:
[90] S. Zuo and D. Yue, “Resilient output formation contain- A systematic design,” IEEE Trans. Autom. Control, vol. 60, no. 8,
ment of heterogeneous multigroup systems against unbounded pp. 2015–2030, Aug. 2015.
attacks,” IEEE Trans. Cybern., early access, Jun. 30, 2020, [114] T. Ibuki, S. Wilson, J. Yamauchi, M. Fujita, and M. Egerstedt,
doi: 10.1109/TCYB.2020.2998333. “Optimization-based distributed flocking control for multiple rigid
[91] X. Liu, Z. Ji, T. Hou, and H. Yu, “Decentralized stabilizability and for- bodies,” IEEE Robot. Autom. Lett., vol. 5, no. 2, pp. 1891–1898,
mation control of multi-agent systems with antagonistic interactions,” Apr. 2020.
ISA Trans., vol. 89, pp. 58–66, Jun. 2019. [115] B. K. Sahu and B. Subudhi, “Flocking control of multiple AUVs based
[92] M. Yu, H. Wang, G. Xie, and K. Jin, “Event-triggered circle forma- on fuzzy potential functions,” IEEE Trans. Fuzzy Syst., vol. 26, no. 5,
tion control for second-order-agent system,” Neurocomputing, vol. 275, pp. 2539–2551, Oct. 2018.
pp. 462–469, Jan. 2018. [116] D. Gu and Z. Wang, “Leader–follower flocking: Algorithms and
[93] Y. Fei, P. Shi, and C.-C. Lim, “Neural network adaptive dynamic slid- experiments,” IEEE Trans. Control Syst. Technol., vol. 17, no. 5,
ing mode formation control of multi-agent systems,” Int. J. Syst. Sci., pp. 1211–1219, Sep. 2009.
vol. 51, no. 11, pp. 2025–2040, 2020. [117] Y. Jia and L. Wang, “Leader–follower flocking of multiple robotic
[94] R. Rahimi, F. Abdollahi, and K. Naqshi, “Time-varying formation con- fish,” IEEE/ASME Trans. Mechatronics, vol. 20, no. 3, pp. 1372–1383,
trol of a collaborative heterogeneous multi agent system,” Robot. Auton. Jun. 2015.
Syst., vol. 62, no. 12, pp. 1799–1805, Dec. 2014. [118] A. Prorok, M. A. Hsieh, and V. Kumar, “The impact of diversity on
[95] J. Yu, X. Dong, Q. Li, and Z. Ren, “Practical time-varying formation optimal control policies for heterogeneous robot swarms,” IEEE Trans.
tracking for second-order nonlinear multiagent systems with multiple Robot., vol. 33, no. 2, pp. 346–358, Apr. 2017.
leaders using adaptive neural networks,” IEEE Trans. Neural Netw. [119] S. Chen, H. Pei, Q. Lai, and H. Yan, “Multitarget tracking control
Learn. Syst., vol. 29, no. 12, pp. 6015–6025, Dec. 2018. for coupled heterogeneous inertial agents systems based on flocking
[96] A. R. Mehrabian and K. Khorasani, “Distributed formation recovery behavior,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 49, no. 12,
control of heterogeneous multiagent Euler–Lagrange systems subject to pp. 2605–2611, Dec. 2019.
network switching and diagnostic imperfections,” IEEE Trans. Control [120] Q. Zhang, Y. Hao, Z. Yang, and Z. Chen, “Adaptive flocking
Syst. Technol., vol. 24, no. 6, pp. 2158–2166, Nov. 2016. of heterogeneous multi-agents systems with nonlinear dynamics,”
[97] C. W. Reynolds, “Flocks, herds and schools: A distributed behavioral Neurocomputing, vol. 216, pp. 72–77, Dec. 2016.
model,” in Proc. ACM SIGGRAPH Conf. Comput. Graph., Jul. 1987, [121] H. Yu, P. Shi, C.-C. Lim, and D. Wang, “Formation control for multi-
pp. 25–34. robot systems with collision avoidance,” Int. J. Control, vol. 92, no. 10,
[98] T. Vicsek, A. Czirók, E. Ben-Jacob, I. Cohen, and O. Shochet, “Novel pp. 2223–2234, 2018.
type of phase transition in a system of self-driven particles,” Phys. Rev. [122] A. D. Ames, A. Abate, and S. Sastry, “Sufficient conditions for the
Lett., vol. 75, no. 6, p. 1226, Aug. 1995. existence of zeno behavior,” in Proc. IEEE 44th Conf. Decis. Control,
Seville, Spain, Dec. 2005, pp. 696–701.
[99] H. G. Tanner, A. Jadbabaie, and G. J. Pappas, “Flocking in fixed
and switching networks,” IEEE Trans. Autom. Control, vol. 52, no. 5, [123] W. Zhu and Z. Jiang, “Event-based leader-following consensus of
pp. 863–868, May 2007. multi-agent systems with input time delay,” IEEE Trans. Autom.
Control, vol. 60, no. 5, pp. 1362–1367, May 2015.
[100] H. Yu, C.-C. Lim, R. Hunjet, and P. Shi, “Flocking and topology manip-
[124] H. Chen, P. Shi, and C. Lim, “Synchronization control for neu-
ulation based on space partitioning,” Robot. Auton. Syst., vol. 124,
tral stochastic delay Markov networks via single pinning impulsive
Feb. 2020, Art. no. 103328, doi: 10.1016/j.robot.2019.103328.
strategy,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 50, no. 12,
[101] F. Sun, R. Wang, W. Zhu, and Y. Li, “Flocking in nonlinear multi-
pp. 5406–5419, Dec. 2020, doi: 10.1109/TSMC.2018.2882836.
agent systems with time-varying delay via event-triggered control,”
[125] W. Xing, P. Shi, R. K. Agarwal, and L. Li, “Robust H∞ pinning syn-
Appl. Math. Comput., vol. 350, pp. 66–77, Jun. 2019.
chronization for complex networks with event-triggered communication
[102] Y. Shen, Z. Kong, and L. Ding, “Flocking of multi-agent system with scheme,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 67, no. 12,
nonlinear dynamics via distributed event-triggered control,” Appl. Sci., pp. 5233–5245, Dec. 2020, doi: 10.1109/TCSI.2020.3004170.
vol. 9, no. 7, p. 1336, Jan. 2019.
[126] Z. Tang, J. H. Park, and H. Shen, “Finite-time cluster synchronization
[103] P. Yu, L. Ding, Z.-W. Liu, and Z.-H. Guan, “Leader–follower flock- of Lur’e networks: A nonsmooth approach,” IEEE Trans. Syst., Man,
ing based on distributed event-triggered hybrid control,” Int. J. Robust Cybern., Syst., vol. 48, no. 8, pp. 1213–1224, Aug. 2018.
Nonlinear Control, vol. 26, no. 1, pp. 143–153, 2016. [127] H. Modares, B. Kiumarsi, F. L. Lewis, F. Ferrese, and A. Davoudi,
[104] Y. Hu, J. Zhan, and X. Li, “Self-triggered distributed model predictive “Resilient and robust synchronization of multiagent systems under
control for flocking of multi-agent systems,” IET Control Theory Appl., attacks on sensors and actuators,” IEEE Trans. Cybern., vol. 50, no. 3,
vol. 12, no. 18, pp. 2441–2448, Dec. 2018. pp. 1240–1250, Mar. 2020.
[105] Y. Xu, M. Fang, P. Shi, Y. Pan, and C. K. Ahn, “Multileader [128] Q. Shen, B. Jiang, P. Shi, and J. Zhao, “Cooperative adaptive fuzzy
multiagent systems containment control with event-triggering,” IEEE tracking control for networked unknown nonlinear multiagent systems
Trans. Syst., Man, Cybern., Syst., early access, Mar. 12, 2019, with time-varying actuator faults,” IEEE Trans. Fuzzy Syst., vol. 22,
doi: 10.1109/TSMC.2019.2899967. no. 3, pp. 494–504, Jun. 2014.
[106] L. Li, P. Shi, R. K. Agarwal, C. K. Ahn, and W. Xing, “Event-triggered [129] X. Yu, Z. Liu, and Y. Zhang, “Fault-tolerant formation control of
model predictive control for multiagent systems with communication multiple UAVs in the presence of actuator faults,” Int. J. Robust
constraints,” IEEE Trans. Syst., Man, Cybern., Syst., early access, Nonlinear Control, vol. 26, no. 12, pp. 2668–2685, 2016.
Aug. 21, 2019, doi: 10.1109/TSMC.2019.2932838. [130] F. Liao, R. Teo, J. L. Wang, X. Dong, F. Lin, and K. Peng, “Distributed
[107] J. Gao, X. Xu, N. Ding, and E. Li, “Flocking motion of multi-agent formation and reconfiguration control of VTOL UAVs,” IEEE Trans.
system by dynamic pinning control,” IET Control Theory Appl., vol. 11, Control Syst. Technol., vol. 25, no. 1, pp. 270–277, Jan. 2017.
no. 5, pp. 714–722, Mar. 2017. [131] K. Zhang, B. Jiang, and P. Shi, “Adjustable parameter-based distributed
[108] S. Yazdani, M. Haeri, and H. Su, “Sampled-data leader–follower algo- fault estimation observer design for multiagent systems with directed
rithm for flocking of multi-agent systems,” IET Control Theory Appl., graphs,” IEEE Trans. Cybern., vol. 47, no. 2, pp. 306–314, Feb. 2017.
vol. 13, no. 5, pp. 609–619, Mar. 2019. [132] J. Huang, Nonlinear Output Regulation: Theory and Applications.
[109] K. Saulnier, D. Saldaña, A. Prorok, G. J. Pappas, and V. Kumar, Philadelphia, PA, USA: SIAM, 2004.
“Resilient flocking for mobile robot teams,” IEEE Robot. Autom. Lett., [133] Y. Hua, X. Dong, G. Hu, Q. Li, and Z. Ren, “Distributed time-varying
vol. 2, no. 2, pp. 1039–1046, Apr. 2017. output formation tracking for heterogeneous linear multiagent systems
[110] S. H. Semnani and O. A. Basir, “Semi-flocking algorithm for motion with a nonautonomous leader of unknown input,” IEEE Trans. Autom.
control of mobile sensors in large-scale surveillance systems,” IEEE Control, vol. 64, no. 10, pp. 4292–4299, Oct. 2019.
Trans. Cybern., vol. 45, no. 1, pp. 129–137, Jan. 2015. [134] S. Zuo, Y. Song, F. L. Lewis, and A. Davoudi, “Optimal robust output
[111] D. Lee and M. W. Spong, “Stable flocking of multiple inertial agents containment of unknown heterogeneous multiagent system using off-
on balanced graphs,” IEEE Trans. Autom. Control, vol. 52, no. 8, policy reinforcement learning,” IEEE Trans. Cybern., vol. 48, no. 11,
pp. 1469–1475, Aug. 2007. pp. 3197–3207, Nov. 2018.

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.
SHI AND YAN: A SURVEY ON INTELLIGENT CONTROL FOR MASs 175

[135] J. Scherer et al., “An autonomous multi-UAV system for search and [152] S. Lin, B. D. Schutter, Z. Zhou, and Y. Xi, “Multi-agent model-
rescue,” in Proc. 1st Workshop Micro Aerial Veh. Netw. Syst. Appl. based predictive control for large-scale urban traffic networks using a
Civil. Use, 2015, pp. 33–38. serial scheme,” IET Control Theory Appl., vol. 9, no. 3, pp. 475–484,
[136] I. Maza, F. Caballero, J. Capitán, J. R. Martínez-de Dios, and A. Ollero, Feb. 2015.
“Experimental results in multi-UAV coordination for disaster manage- [153] S. El-Tantawy, B. Abdulhai, and H. Abdelgawad, “Multiagent rein-
ment and civil security applications,” J. Intell. Robot. Syst., vol. 61, forcement learning for integrated network of adaptive traffic signal
nos. 1–4, pp. 563–585, 2011. controllers (MARLIN-ATSC): Methodology and large-scale applica-
[137] Y. Liu et al., “A distributed control approach to formation balancing tion on downtown Toronto,” IEEE Trans. Intell. Transp. Syst., vol. 14,
and maneuvering of multiple multirotor UAVs,” IEEE Trans. Robot., no. 3, pp. 1140–1150, Sep. 2013.
vol. 34, no. 4, pp. 870–882, Aug. 2018. [154] T. Wu et al., “Multi-agent deep reinforcement learning for urban traffic
[138] X. Dong, B. Yu, Z. Shi, and Y. Zhong, “Time-varying formation control light control in vehicular networks,” IEEE Trans. Veh. Technol., vol. 69,
for unmanned aerial vehicles: Theories and applications,” IEEE Trans. no. 8, pp. 8243–8256, Aug. 2020.
Control Syst. Technol., vol. 23, no. 1, pp. 340–348, Jan. 2015. [155] C. Yu et al., “Distributed multiagent coordinated learning for
[139] D. Xiwang, L. Yangfan, L. Chuang, H. Guoqiang, L. Qingdong, and autonomous driving in highways based on dynamic coordination
Z. Ren “Time-varying formation tracking for UAV swarm systems with graphs,” IEEE Trans. Intell. Transp. Syst, vol. 21, no. 2, pp. 735–748,
switching directed topologies,” IEEE Trans. Neural Netw. Learn. Syst., Feb. 2020.
vol. 30, no. 12, pp. 3674–3685, Dec. 2019.
[140] G. Yang, S. Wang, J. Yang, B. Shen, and P. Shi, “Pose estimation of
daily containers for a life-support robot,” Int. J. Innovat. Comput. Inf.
Control, vol. 14, no. 4, pp. 1545–1552, 2018.
[141] J. Shao, G. Xie, and L. Wang, “Leader-following formation control
of multiple mobile vehicles,” IET Control Theory Appl., vol. 1, no. 2,
pp. 545–552, Mar. 2007.
[142] X. Liang, H. Wang, Y. H. Liu, W. Chen, and T. Liu, “Formation Peng Shi (Fellow, IEEE) received the Ph.D.
control of nonholonomic mobile robots without position and veloc- degree in electrical engineering from the University
ity measurements,” IEEE Trans. Robot., vol. 34, no. 2, pp. 434–446, of Newcastle, Callaghan, NSW, Australia, in
Apr. 2018. 1994, the Doctor of Science degree from the
[143] Z. Li, Y. Yuan, F. Ke, W. He, and C.-Y. Su, “Robust vision-based tube University of Glamorgan, Wales, U.K., in 2006,
model predictive control of multiple mobile robots for leader–follower and the Doctor of Engineering degree from the
formation,” IEEE Trans. Ind. Electron., vol. 67, no. 4, pp. 3096–3106, University of Adelaide, Adelaide, SA, Australia,
Apr. 2020. in 2015.
[144] S. M. Lee, H. Kim, H. Myung, and X. Yao, “Cooperative coevolution- He is currently a Professor with the University
ary algorithm-based model predictive control guaranteeing stability of of Adelaide. His research interests include automa-
multirobot formation,” IEEE Trans. Control Syst. Technol., vol. 23, tion and control systems, autonomous and robotic
no. 1, pp. 37–51, Jan. 2015. systems, cyber-physical systems, and network systems.
[145] X. Li and D. Zhu, “An adaptive SOM neural network method for Prof. Shi has served on the Editorial Board for a number of journals, includ-
distributed formation control of a group of AUVs,” IEEE Trans. Ind. ing Automatica, IEEE T RANSACTIONS ON AUTOMATIC C ONTROL, IEEE
Electron., vol. 65, no. 10, pp. 8260–8270, Oct. 2018. T RANSACTIONS ON C YBERNETICS, IEEE T RANSACTIONS ON C IRCUITS
[146] C. Suryendu and B. Subudhi, “Modified constrained adaptive formation AND S YSTEMS , IEEE T RANSACTIONS ON F UZZY S YSTEMS , and IEEE
control scheme for autonomous underwater vehicles under commu- C ONTROL S YSTEMS L ETTERS. He is a Member of Board of Governors for
nication delays,” IET Cyber Syst. Robot., vol. 2, no. 1, pp. 22–30, IEEE SMC Society, and an IEEE Distinguished Lecturer. He is a Fellow of
Mar. 2020. the Institution of Engineering and Technology, and the Institute of Engineers,
[147] X. Zhang, A. J. Flueck, and C. P. Nguyen, “Agent-based distributed Australia.
volt/var control with distributed power flow solver in smart grid,” IEEE
Trans. Smart Grid, vol. 7, no. 2, pp. 600–607, Mar. 2016.
[148] V. P. Singh, N. Kishor, and P. Samuel, “Distributed multi-agent system-
based load frequency control for multi-area power system in smart
grid,” IEEE Trans. Ind. Electron., vol. 64, no. 6, pp. 5151–5160,
Jun. 2017.
[149] P. Wang and M. Govindarasu, “Multi-agent based attack-resilient
Bing Yan received the B.Sc. degree in automa-
system integrity protection for smart grid,” IEEE Trans. Smart Grid,
tion and the M.Sc. degree in control theory
vol. 11, no. 4, pp. 3447–3456, Jul. 2020.
and control engineering from the School of
[150] B. U. I. Khan, F. Anwar, R. F. Olanrewaju, B. R. Pampori, and
Automation, Northwestern Polytechnical University,
R. N. Mir, “A novel multi-agent and multilayered game formulation for
Xi’an, China, in 2012 and 2015, respectively. She is
intrusion detection in Internet of Things (IoT),” IEEE Access, vol. 8,
currently pursuing the Ph.D. degree with the School
pp. 98481–98490, 2020.
of Electrical and Electronic Engineering, University
[151] J. Wang, Y. Zhang, Y. Liu, and N. Wu, “Multiagent and bargaining-
of Adelaide, Adelaide, SA, Australia.
game-based real-time scheduling for Internet of Things-enabled flexible
Her research interests include flight control, for-
job shop,” IEEE Internet Things J., vol. 6, no. 2, pp. 2518–2531,
mation control, and multiagent systems.
Apr. 2019.

Authorized licensed use limited to: Northeastern University. Downloaded on October 17,2024 at 01:46:55 UTC from IEEE Xplore. Restrictions apply.