Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

2021 29th Mediterranean Conference on Control and Automation (MED)

June 22-25, 2021. Bari, Puglia, Italy

Dynamic Average Consensus based on Group Information for


Multi-Swarm Systems.
Fabien NIEL and François BATEMAN1

Abstract— This paper deals with distributed protocols to information and the group information to solve the consensus
track the time-varying average of a dynamic Multi-Swarm Sys- in both continuous and discrete time. This work is extended
tem (MSS) using group information. To solve the swarm average to nonlinear group information transformation in [18]. In
consensus problem, the agents of each swarm can exchange
information at the intra-swarm level. However, only one agent these previous works, the group information is supposed to
of each swarm, referred to as the swarm communicating agent, be available to all the agents of the network, which may be
can share its group information with the other swarms. We restricting for some applications.
show that the discrete time protocols asymptotically achieve
consensus on the dynamic swarm average with a bounded error.
The algorithms are developed for both fixed and switching
communicating agents configurations with strongly connected 3 4

and weighted balanced directed graphs. We provide numerical


simulations to illustrate the efficiency of our solutions.
2 1
I 7
I. INTRODUCTION
5
Many systems can be described as large-scale networks
of simple agents. Consensus methods allow the emergence 6
II
of collective behaviors from basic rules, using local infor- 11 8

mation. These methods deliver robust results for their task,


despite a changing environment and non-stationary network 10 9
configurations such as the loss of several agents for instance.
III
Over the last decades, consensus problems for Multi-Agent
Systems (MAS) have been of particular interest [1]–[3] and
Fig. 1. Example of a multi-swarm system with 11 agents divided into 3
have found many applications such as mobile robots and interacting swarms. Plain blue arrows represent the intra-swarm interactions
UAVs cooperative control [4], spacecraft formation [5], smart while the dotted red arrows show the inter-swarm interactions of the
electrical grids [6], distributed optimization methods [7], communicating agents.
sensor fusion [8] and distributed Kalman filter [9].
Dynamic consensus problem has been significantly stud- In this paper, we study the average consensus for a set
ied, particularly since the work of [10] where the reference of swarms. We refer to this problem as the multi-swarm
signals derivative is added to the consensus equation to track average consensus. In this problem, the agents of each swarm
the time-varying average. More recently, the tutorial from calculate the average of their swarm. This local information
[11] analyses this protocol and presents several algorithms represents the group information. Some of the agents of each
to cope with some of the challenges such as the initialization, swarm can communicate their own group information with
the convergence rate or the robustness. In [12] and [13], one or several agents from the other swarms as presented
dynamic consensus algorithms are applied to the containment Fig. 1. The objective is to track the dynamic average of the
control problem. Open multi-agent systems, where agents set of swarms based on group information rather than agent
may join or leave the network at any time, have also been information. In particular, this work proposes a particular
investigated in [14] and [15], employing a modified dynamic protocol to consider the possible switchings between the
consensus. agents communicating at the inter-swarm level. The com-
In [16], the MAS is partitioned into several groups. munication are assumed to be in discrete time. Comparing
The consensus is reached using not only the agents’ states to a multi-agent system, this multi-swarm approach can
but also an intragoup collective information, called group ultimately provide an interesting protocol to reduce the
information. If this value has been previously introduced in number of interactions between the agents.
The paper is organized as follows. Section II presents some
[17], [16] uses a hierarchical structure, combining the agent
basic notations, graph theory fundamentals and a preliminary
1 Fabien NIEL and François BATEMAN are with the Centre de Recherche result, concerning dynamic average consensus. The dynamic
de l’Ecole de l’air at the French Air Force Academy, Salon de Provence swarm average consensus problem is then formulated in Sec-
13300, France tion III. Section IV details the main results for both fixed and
fabien.niel.af@gmail.com or
https://orcid.org/0000-0003-2912-937X switching communicating agents configurations. Numerical
francois.bateman@ecole-air.fr simulations are provided and discussed in Section V. Finally,

978-0-7381-1098-1/21/$31.00 ©2021 IEEE 1222


1
Section VI concludes this paper. sup k(IN − 1N 1TN )(r(tk̄+1 ) − r(tk̄ ))k = γ(tk ) < ∞.
k̄∈Z≥k N
II. MATHEMATICAL PRELIMINARIES
A. Notations III. PROBLEM FORMULATION
The sets of real, positive real, non-negative real, and Consider a MSS composed of n agents partitioned into a
natural numbers are respectively denoted by R, R>0 , R≥0 set of m swarms, G = [G1 , .., Gm ].
and N. Rn and Rn×m represent respectively the set of n Each agent belongs to only one swarm and can only
dimension column vectors and the set of n × m matrices. communicates with the neighbors of its own swarm. We
Given A ∈ Rn×m , AT represents the transpose of this define for each swarm a digraph GI = (VI , EI , AI ), I =
matrix with AT ∈ Rm×n . Col(.) denotes a column vector. {1, ..m}. We denote λI,2 , their second smallest eigenvalues
1n is Col(1, ..., 1) of dimension n × 1. Given x ∈ Rn , the and dmax
G = max dmax (GI ). Using a dynamic consensus
I
Euclidean norm is denoted by kxk. Consider a set of m algorithm, each swarm tracks its own group information. The
swarms containing respectively nI agents for I ∈ {1, .., m}. group information is shared between the swarms using only
Then, xiI represents the ith agent of the I th swarm with a single agent from each swarm in order to compute the
i ∈ VI = {1, .., nI }. The aggregate of the ni agents average of the swarms’ group information. Thus, we can
iT
define a digraph G = (V, E, A) where |V| = m. We denote
h
nI T
is denoted by xI = x1T I , .., x . |V| represents the
I
dmax the maximum degree of G and λ2 , the second smallest
cardinality of set V.
eigenvalue.
B. Graph theory Assumption 1: The digraphs at the intra and inter-swarm
Some necessary results and definitions concerning graph level are strongly connected and weight-balanced.
theory from [11], [19], [20] are provided in the sequel. 1 Pm
We define the swarm average x̄S (tk ) = x̄I (tk )
A weighted digraph G = (V, E, A) is defined by a finite set th
m I=1
of n vertices V , a finite set of edges E ⊂ V 2 and a weighted with x̄I the group information of the I -swarm. We select
adjacency matrix A where aij > 0 if (i, j) ∈ E, aij = 0 the swarm element average, x̄I (tk ) as the group information
1 Pni j
otherwise. The weighted out-degree and weightedP in-degree such as x̄I (tk ) = x (tk ).
Pn n nI j=1 I
are respectively dout (i) = j=1 aji and din (i) = j=1 aij .
Remark 1: Note that in general the swarm average is
We denote by dmax the maximum weighted out-degree. A
different from the average of all the agents, i.e. x̄S (tk ) 6=
digraph is weighted-balanced if for each vertex i ∈ V, the 1 PnI j
weighted in-degree is equal to the weighted out-degree. The x (tk ).
n j=1
out-degree matrix is represented by Dout = [δij ] where δij = These information are then used to solve the swarm
dout (i) for i = j and 0 otherwise. The Laplacian matrix L average consensus problem as defined:
is given by Dout − A. The smallest eigenvalue is 0 and Definition 1: (swarm average consensus) The swarm av-
the corresponding eigenvector is 1n such as L1n = 0. We erage consensus is reached if ∃ e ∈ R≥0
denote by λ2 the second smallest eigenvalue of Sym(L) =
(L + LT )/2 . lim kw̄I − x̄S k ≤ e, I ∈ {1, .., m} (4)
k→∞
A directed path is an ordered sequence of vertices. A
digraph is strongly connected if there exists a directed path with w̄I the swarm average estimation of the I th -swarm.
between every pair of vertices {vi , vj } ∈ V 2 . This configuration aims at taking advantage of the local
Theorem 1 (Dynamic average consensus algorithm [11]): swarm information and reducing the communication between
Let G(V, E) be a strongly connected and weight-balanced the agents of different swarms, as in multi agent consensus.
digraph of N agents. Assume that each agent i ∈ V has This system is also different from [16] since the group
access to a dynamic input ri (k) = ri (tk ) at time tk = kδ, information is computed internally by the swarms and shared
δ ∈ R>0 , k ∈ N. For δ ∈ (0, β −1 (dmax )−1 ), where β ∈ R>0 by only a single agent of each swarm, eventually switching.
if each agent i ∈ V implements
IV. M AIN R ESULTS
N
pi (tk+1 ) = pi (tk ) + δβ
X
aij (xi (tk ) − xj (tk )), (1) A. Dynamic Group Information consensus with fixed com-
j=1
municating agents
In this section, and without loss of generality, we consider
xi (tk ) = ri (tk ) − pi (tk ) (2)
that only the first agent of each swarm can communicate with
starting at pi (0) = 0, then the trajectory k → xi (k) of agent the other swarms and execute the swarm average consensus.
i ∈ V is bounded and satisfies We later refer to these agents as the communicating agents.
Let introduce the group information discrete-time proto-
N
1 X j γ(∞)δ col, based on two hierarchical level proportional dynamic
lim kxi (tk ) − r (tk )k ≤ (3) average consensus algorithms. The first average consensus,
k→∞ N j=1 βλ2
executed by all the agents inside each swarm, aims at estimat-
with ing the group information, i.e. the swarm element average.

1223
The second algorithm, only executed by the communicating This result holds ∀I ∈ {1, .., m} and ∀i ∈ {1, .., nI }, and
agents, estimates the overall swarm average. in particular for i = 1. Considering that the communicating
Using their own position as reference signal, ri = xi , agent is the first agent of each swarm, we have wI1 = r̄I ,
each agent i ∈ VI , ∀I ∈ {1, .., m} executes the following ∀I ∈ {1, .., m} and we can write:
algorithm, for δa ∈ (0, β −1 (dmax
G )
−1
), where β ∈ R>0 : nI
1 X γ(∞)δ
N lim kr̄I (tk ) − rj (tk )k ≤ (10)
X k→∞ nI j=1 I βλI,2
piI (tk+1 ) = piI (tk ) + δa β aIij (wIi (tk ) − wIj (tk )), (5a)
j=1 1 PnI j
With r (tk ) = x̄I (tk )
wIi (tk ) = riI (tk ) − piI (tk ), (5b) nI j=1 I
γ(∞)δ
with piI (0) = 0. lim kr̄I (tk ) − x̄I (tk )k ≤ (11)
Considering wI1 = r̄I , ∀I ∈ {1, .., m}, and for δs ∈ k→∞ βλI,2
(0, β̄ −1 (dmax )−1 ), where β̄ ∈ R>0 each communicating Considering the second part of Eq. (8), we have:
agent implements : m m
1 X 1 X
m k r̄I (tk ) − x̄I (tk )k ≤
X m m
p̄I (tk+1 ) = p̄I (tk ) + δs β̄ āij (w̄I (tk ) − w̄J (tk )), (6a) J=1
m
J=1
(12)
J=1 1 X
kr̄I (tk ) − x̄I (tk )k
w̄I (tk ) = r̄I (tk ) − p̄I (tk ). (6b) m
J=1

where w̄I , ∀I ∈ {1, .., m} is the swarm average estimation and then, using the previous result:
and p̄I (0) = 0. 1 X
m
1 X
m

Theorem 2: (Distributed consensus for multi-swarm prob- lim k r̄I (tk ) − x̄I (tk )k ≤
k→∞ m m
J=1 J=1
lem) m (13)
1 X γ(∞)δ
Under protocol Eq.(5), if assumption 1 holds, then, sys-
tem (6) achieves dynamic swarm average consensus with a m βλJ,2
J=1
bounded error, ē ∈ R>0 , such that: Considering the first part of Eq. (8) and applying The-
lim kw̄I (tk ) − x̄S (tk )k ≤ ē orem 1 to system (6), with r̄I (tk ) finite ∀k ∈ N, we can
k→∞
γ̄(∞)δ 1 Pm γ(∞)δ (7) write:
ē = + <∞ m
β̄λ2 m J=1
βλJ,2 1 X γ̄(∞)δ
lim kw̄I (tk ) − r̄J (tk )k ≤ (14)
k→∞ m β̄λ2
with δ ∈ (0, β −1 (dmax
G )
−1
) ∩ (0, β̄ −1 (dmax )−1 ), J=1
1 with γ̄ defined as:
sup k(In − 1n 1Tn )(r(tk̄+1 ) − r(tk̄ ))k = γ(tk ) < ∞,
n 1
k̄∈Z≥k sup k(In − 1n 1Tn )(r̄I (tk̄+1 )−r̄I (tk̄ ))k = γ̄(tk ) < ∞.
and k̄∈Z≥k n
1
sup k(In − 1n 1Tn )(r̄I (tk̄+1 ) − r̄I (tk̄ ))k = γ̄(tk ) < ∞ Then, for k → ∞, Eq.(8) becomes:
k̄∈Z≥k n
m
γ̄(∞)δ 1 X γ(∞)δ
Proof: lim kw̄I (tk ) − x̄S (tk )k ≤ +
k→∞ β̄λ2 m βλJ,2
Let consider kw̄I (tk ) − x̄S (tk )k with x̄S (tk ) = J=1
1 Pm (15)
x̄I (tk ) γ̄(∞)δ 1 Pm γ(∞)δ
m J=1 Thus, denoting ē = + < ∞, we
β̄λ2 m J=1 βλJ,2
1 Pm have a finite bounded error, which concludes the proof.
kw̄I (tk ) − J=1 x̄I (tk )k =
m
m m m Remark 2: In particular, we observe that the error depends
1 X 1 X 1 X
kw̄I (tk ) − r̄I (tk ) + r̄I (tk ) − x̄I (tk )k ≤ on the connectivity of the different digraphs. Also, if the
m J=1 m J=1 m J=1
m m m reference signals follow the same dynamics (γ(tk ) → 0, for
1 X 1 X 1 X k > k̄ which implies γ̄(tk ) → 0), then, the error tends to
kw̄I (tk ) − r̄I (tk )k + k r̄I (tk ) − x̄I (tk )k
m J=1 m J=1 m J=1 zero as pointed in [13] and lim w̄I (tk ) = x̄S (tk ), ∀I ∈
(8) k→∞
{1, .., m} .
Applying Theorem 1 to system (6), we obtain for each B. Dynamic Group Information consensus with switching
swarm: communicating agents
nI
1 X γ(∞)δ In this section, we consider that the communicating agents
lim kwIi (tk ) − rj (tk )k ≤ (9)
k→∞ nI j=1 I βλI,2 can switch among the agents of their own swarm. If the
communication graph between the swarms is preserved,
1 protocol (5)-(6) introduces an error due to the initialization
with sup k(InI − 1n 1T )(rI (tk̄+1 ) − rI (tk̄ ))k =
k̄∈Z≥k n I I nI of the communicating agent’s state and fails to follow the
γ(tk ) < ∞. swarm average.

1224
Consequently, protocol (5)-(6) is modified, and a third Let consider the piece-wise continuous input signal, ŵIv
leader/follower consensus is introduced and run among the composed of the estimated Ith -swarm average of agents s
agents in order to propagate the group information to every and t such as:
agent of the swarm, similar to what can be done for an open (
ŵIs t < tk0 ,
multi-agent system [14]. ŵIv (t) = (20)
Let consider a switching signal σ(tk ), such as σ(tk ) : ŵIt + ∆st (t
pI k0 ) t ≥ tk0 .
N → P with P the finite set of all possible combinations of Now, let consider a virtual communicating agent for the
communicating agents. In particular, |P| = ΠmI=1 nI I th swarm which corresponds to agent s for t < tk0 and
Assumption 2: Let Nσ (tk ) denotes the number of switch- to agent t for t ≥ tk0 . The objective is to demonstrate that
ings of σ(tk ). Then, there exists N = Nσ(∞) < ∞ ∈ N the agent switching is equivalent to keep the same agent but
such that Nσ → Nσ(∞) with a change in its reference signal.
The estimation of the swarm average by the virtual agent
Thus, for δp ∈ (0, β̂ −1 (dmax
G )
−1
), where β̂ ∈ R>0 and at t = tk0 +1
∀I ∈ {1, .., m} and ∀i ∈ {1, .., nI }\{l} with {l(tk ) = m
lI (tk )} the communicating agent index at time tk , we
X
p̄I (tk0 +1 ) = p̄I (tk0 ) + δ β̄ āij (w̄I (tk0 ) − w̄J (tk0 )),
implement: J=1
(21a)
N
X w̄I (tk0 +1 ) = r̄I (tk0 +1 ) − p̄I (tk0 +1 ). (21b)
p̂iI (tk+1 ) = p̂iI (tk ) + δp β̂ aIij (ŵIi (tk ) − ŵIj (tk )), (16a) Using Eq. (19)
j=1
p̄I (tk0 +1 ) = p̄I (t−
kP ) + ∆st
pI (tk0 )
ŵIi (tk ) = r̂iI (tk ) − p̂iI (tk ), (16b) 0
m
+δβ J=1 āij (w̄I (tk0 ) − w̄J (tk0 )).
(22)
with r̂iI (tk ) = wIi (tk ), p̂iI (0) = 0 and
Let make the change of coordinates p̄∗I = p̄I − ∆st pI k0 )
(t
Then we have
p̂lI (tk+1 ) = p̂lI (tk ), (16c) Pm
p̄∗I (tk0 +1 ) = p̄I (t−
k0 ) + δβ J=1 āij (w̄I (tk0 ) − w̄J (tk0 ))
ŵIl (tk ) = w̄I (tk ) − p̂lI (tk ). (16d) (23a)
w̄I (tk0 +1 ) = r̄I (tk0 +1 ) − p̄∗I (t− st
k0 ) − ∆pI (tk0 )
Theorem 3: (Distributed consensus for multi-swarm prob- ∗ − (23b)
= (r̄I (tk0 +1 ) − ∆stpI (tk0 )) − p̄I (tk0 )
lem with switching communicating agent)
If assumptions 1 and 2 hold, and under protocol defined by We can see that Eq.(23) is equivalent to keep agent s as
Eq.(5) and (16), system (6) achieves dynamic swarm average the communicating agent following the reference signal (20)
consensus with a bounded error, e ∈ R>0 , such that: including the additional constant error, ∆st
pI (tk0 ), after tk0 .
lim kw̄I (tk ) − x̄S (tk )k ≤ e This error being present in the reference signal itself for
k→∞
γ̄(∞)δ 1 Pm γ(∞)δ 1 PN σ(j) tk ≥ tk0 , the dynamic swarm average is also modified as
with e ≤ + J=1 + ∆pI (tkj ) followed :
β̄λ2 m βλJ,2 m j=1
(17)
m
1 X ∆st
pI (tk0 )
x̄∗S (tk ) = x̄I (tk ) + (24)
where δ ∈ (0, β −1 (dmax
G )
−1
) ∩ (0, β̂ −1 (dmax
G )
−1
) ∩ m m
−1 max −1 I=1
(0, β̄ (d
), )
1 Thus, applying Theorem 2 for k ≥ tk0 , we have
sup k(In − 1n 1Tn )(r(tk̄+1 ) − r(tk̄ ))k = γ(tk ) < ∞,
k̄∈Z≥k n
1 lim kw̄I (tk ) − x̄∗S (tk )k ≤ ē (25)
sup k(In − 1n 1Tn )(r̄I (tk̄+1 ) − r̄I (tk̄ ))k = γ̄(tk ) < ∞, k→∞
k̄∈Z≥k n
And we finally have:
σ(j)
and ∆pI (tkj ) < ∞ represents the internal state error due
lim kw̄I (tk ) − x̄∗S (tk )k
to the switching σ(j) at time tkj . k→∞
Proof: ∆st
pI (tk0 )
= lim kw̄I (tk ) − x̄S (tk ) + k (26)
Let consider a switching at time t+ th
k0 of the I -swarm k→∞ m
st
such that the communicating agent s loses its communication ∆pI (tk0 )
≤ kw̄I (tk ) − x̄S (tk )k +
function to the benefit of agent t. m
Thus, we have: Extending this case to the finite number of switchings
w̄I (t− s −
p̄I (t+ t − following assumption 2, the error can be added for each
k0 ) = wI (tk0 ) k0 ) = pI (tk0 )
+ + (18) switching, which guarantees a bounded error and thus con-
w̄I (tk0 ) = ŵIt (tk0 ) p̄I (tk0 ) = p̂tI (tk0 )
cludes the proof.
For this agent, we also define:
It is interesting to notice that ∆pI depends on the parame-

p̄I (t+ st
k0 ) − p̄I (tk0 ) = ∆pI (tk0 ). (19) ters β̂ and δp from Eq. (16), which can affect the convergence

1225
speed as well as the error of the consensus to propagate the
30
group information. Then, reducing this local error directly
decays the effect of the switchings as it will be shown in the
20
next section.

10
V. SIMULATIONS

error averages
To illustrate our theoretical solutions, let consider a MSS 0
of 11 agents partitioned into 3 swarms as presented Fig. 1.
The agents use a ring configuration to communicate at -10
the intra-swarm level, while the communicating agents can
all exchange their group information. We consider that the
-20
communicating agents of swarm II and swarm III change
respectively at t = 2s from agent 5 to agent 6 and t =
-30
20s from agent 6 to agent 5 and agent 8 to agent 11. The 0 5 10 15 20 25 30
communication frequencies at the intra and inter-swarm level time (s)

are equal with δ = δa = δs = δp = 0.1s.


Fig. 3. Estimation error of the communicating agents using protocol (5-6)
The simulated agent positions follow with switchings at t = 2s and t = 20s
the signal u(t) with u(0) = u0 =
[5, 3, 0.5, −0.5, 27, 26.5, 24.5, −24, −25, −26, −29]T and
70
u̇ = 2(sin(0.1t) + H(t − 10)) with H(t) the Heaviside
function. 60

Simulations are computed for both protocols and 2 values


swarm and multi-swarm averages

50
of β̂ for the second protocol i.e. for β̂ = 1 and for β̂ = 10.
40

30
70

60 20

10
swarm and multi-swarm averages

50

40 0

30 -10

20 -20

10 -30
0 5 10 15 20 25 30
0 time (s)

-10
Fig. 4. Swarm average estimation of the communicating agents using
-20 protocol (5-6 and 16) and β̂ = 1 with switchings at t = 2s and t = 20s

-30
0 5 10 15 20 25 30
time (s) 30

Fig. 2. Swarm average estimation of the communicating agents using 20


protocol (5-6) and β̂ = 1. The communicating agents switch at t = 2s and
t = 20s. Plain lines represents the inter-swarm average estimations, dotted
lines show the intra-swarm averages, while the ’.-’ thick line is the actual 10
error averages

swarm average.

0
We can see from Fig. 2 and Fig. 3, that the protocol (5-
6) successfully tracks the multi-swarm average. However,
-10
constant errors are introduced at each switching of the
communicating agents.
-20
Fig. 4 and Fig. 5 illustrate the benefit of introducing
Eq. (16). Indeed, if an error is still present after t = 2s
due to transient time, this error is smaller than the previous -30
0 5 10 15 20 25 30
simulation and no significant error is added after the second time (s)
agent switching at t = 20s (when algorithm (16) has
converged and the agents can successfully track the group Fig. 5. Estimation error of the communicating agents using protocol (5-6
and 16) with switchings at t = 2s and t = 20s
information). Fig. 6 shows the difference signals estimated

1226
70
exchange the group information. The solution is proposed for
60 strongly connected and weighted-balanced graph in discrete
50 time. We show that the communicating agents reach the
swarm average consensus with a bounded error. The case of
40
Average estimations

fixed communicating agents is first explored and extended to


30 switching communicating agents. Finally, numerical simula-
20 tions demonstrate the effectiveness of our solutions and point
10
out the possibilities for tuning the algorithm by a careful
selection of the sampling times and parameters β.
0
R EFERENCES
-10
[1] A. Jadbabaie, J. Lin, and A. S. Morse, “Coordination of groups
-20
of mobile autonomous agents using nearest neighbor rules,” IEEE
-30
Transactions on automatic control, vol. 48, no. 6, pp. 988–1001, 2003.
0 5 10 15 20 25 30 [2] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks
time (s) of agents with switching topology and time-delays,” IEEE Transac-
tions on automatic control, vol. 49, no. 9, pp. 1520–1533, 2004.
Fig. 6. Swarm average estimation using protocol (5-6 and 16) and β̂ = 1 [3] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and
with switchings at t = 2s and t = 20s cooperation in networked multi-agent systems,” Proceedings of the
IEEE, vol. 95, no. 1, pp. 215–233, 2007.
[4] A. Tahir, J. Böling, M.-H. Haghbayan, H. T. Toivonen, and J. Plosila,
70 “Swarms of unmanned aerial vehicles—a survey,” Journal of Indus-
trial Information Integration, vol. 16, p. 100106, 2019.
60
[5] W. Ren, “Formation keeping and attitude alignment for multiple
spacecraft through local interactions,” Journal of Guidance, Control,
swarm and multi-swarm averages

50
and Dynamics, vol. 30, no. 2, pp. 633–638, 2007.
40 [6] F. G. Mendoza, D. Bauso, and T. Namerikawa, “Transient dynamics
of heterogeneous micro grids using second order consensus,” in
30 2017 International Conference on Wireless Networks and Mobile
Communications (WINCOM), 2017, pp. 1–6.
20
[7] R. Carli, G. Notarstefano, L. Schenato, and D. Varagnolo, “Analysis
10 of newton-raphson consensus for multi-agent convex optimization
under asynchronous and lossy communications,” in 2015 54th IEEE
0 Conference on Decision and Control (CDC). IEEE, 2015, pp. 418–
424.
-10
[8] W. Li, Z. Wang, G. Wei, L. Ma, J. Hu, and D. Ding, “A survey
-20
on multisensor fusion and consensus filtering for sensor networks,”
Discrete Dynamics in Nature and Society, vol. 2015, 2015.
-30 [9] R. Olfati-Saber, “Distributed kalman filter with embedded consensus
0 5 10 15 20 25 30 filters,” in Proceedings of the 44th IEEE Conference on Decision and
time (s) Control. IEEE, 2005, pp. 8179–8184.
[10] D. P. Spanos, R. Olfati-Saber, and R. M. Murray, “Dynamic consensus
Fig. 7. Swarm average estimation of the communicating agents using on mobile networks,” in IFAC world congress. Citeseer, 2005, pp.
protocol (5-6 and 16) and β̂ = 10 with switchings at t = 2s and t = 20s 1–6.
[11] S. S. Kia, B. Van Scoy, J. Cortes, R. A. Freeman, K. M. Lynch, and
S. Martinez, “Tutorial on dynamic average consensus: The problem,
its applications, and the algorithms,” IEEE Control Systems Magazine,
by protocol (5-6 and 16) and compare to the tracked swarm vol. 39, no. 3, pp. 40–72, 2019.
average. [12] Y.-F. Chung and S. S. Kia, “Dynamic active average consensus and its
application in containment control,” arXiv preprint arXiv:2008.05722,
As expected, we also observe Fig. 7 that the choice 2020.
of β̂ directly influences the permanent error following a [13] ——, “Distributed dynamic containment control over a strongly con-
switching of communicating agent. In particular, we can nected and weight-balanced digraph,” IFAC-PapersOnLine, vol. 52,
no. 20, pp. 25–30, 2019.
notice that the first switching appeared after the convergence [14] M. Franceschelli and P. Frasca, “Proportional dynamic consensus in
of algorithm (16), and the error is significantly reduced. If open multi-agent systems,” in 2018 IEEE Conference on Decision and
two values of the parameter β̂ have been selected to demon- Control (CDC). IEEE, 2018, pp. 900–905.
[15] ——, “Stability of open multi-agent systems and applications to
strate the influence on the tracking error, similar simulations dynamic consensus,” IEEE Transactions on Automatic Control, 2020.
could have been performed on the sampling times of the 3 [16] J. Hou and R. Zheng, “Hierarchical consensus problem via group
algorithms. This study has been omitted due to the lack of information exchange,” IEEE transactions on cybernetics, vol. 49,
no. 6, pp. 2355–2361, 2018.
space. [17] Z. Lin, J. Hou, G. Yan, and C. B. Yu, “Reach almost sure consensus
with only group information,” Automatica, vol. 52, pp. 283–289, 2015.
VI. CONCLUSION [18] J. Hou, M. Xiang, and Z. Ding, “Group information based nonlinear
consensus for multi-agent systems,” IEEE Access, vol. 7, pp. 26 551–
This paper presents two distributed protocols to solve the 26 557, 2019.
swarm average consensus problem for multi-swarm systems. [19] W. Ren, R. W. Beard, and E. M. Atkins, “Information consensus in
multivehicle cooperative control,” IEEE Control systems magazine,
In our scheme, the agents can communicate with each other vol. 27, no. 2, pp. 71–82, 2007.
inside each swarm. However, the communication between the [20] F. Bullo, J. Cortes, and S. Martinez, “Distributed control of robotics
swarms is limited to a single agent per swarm and can only networks, princeton,” 2009.

1227

You might also like