A Lyapunov Treatment of Swarm Coordination Under Conflict
-
Upload
independent -
Category
Documents
-
view
3 -
download
0
Transcript of A Lyapunov Treatment of Swarm Coordination Under Conflict
A Lyapunov treatment of swarmcoordination under conflict
Paul McCullough1, Mark Bacon1, Nejat Olgac1,Daniel A. Sierra2,3 and Rudy Cepeda-Gomez1
AbstractWe consider hostile conflicts between two multi-agent swarms, called pursuers and evaders. A Newtonian dynamics-based double integrator model is taken into account, as well as a control strategy using the relative positions and velocitiesof opposing swarm members. This control is introduced to achieve stability and the capture of the evaders by the pur-suers. The present document considers only swarms with equal membership strengths and equal mass for simplicity. Thiseffort begins with a set of suggested interaction force profiles, which are functions of local vectors. To formulate a robustcontrol law, a Lyapunov-based stability analysis is used. The group pursuit is conceived in two phases: the approach phase,during which the two swarms act like two individual agents, and the assigned pursuit phase, where each pursuer has anassigned evader. We show that the uncontrolled dynamics, which are marginally stable, are stabilized by the newcontroller.
KeywordsLyapunov stability, multi-agent modeling, stability of nonlinear systems
Date received: 13 July 2009; accepted: 16 November 2009
1. Introduction
This study addresses the modeling, analysis and control
of multi-agent swarms starting from Newton’s equations.
When each agent influences every other agent, the rules
of interaction and stability of group dynamics become
quite complicated. Much of the inspiration stems from
the study of biological swarms (Warburton and Lazarus,
1991; Camazine et al., 2001). Most of the earlier inves-
tigations that focus on this general theme consider homo-
genous swarms, i.e., those composed of alike members,
with a single integrator model and momenta profiles
(Chu et al., 2003, 2006; Gazi and Passino, 2003,
2004a, 2004b; Yao et al., 2007). A difference in treat-
ment in this paper is to start from Newton’s second law
instead.
Gazi and Passino (2004a) expand an earlier pioneering
work on swarm coordination by generalizing the stability
analysis for a class of attraction/repulsion functions for
homogenous swarms. As a consequence of the homoge-
nous membership and symmetric characteristics of the
momenta, stationary and stable swarm behavior is
achieved. They also propose modifications to account for
finite body size of the swarm members. In these studies, all
members are required to know (or sense) the position of all
the others.
Chu et al. (2003, 2006) address the stability analysis of
anisotropic (asymmetric behavior), but nonhostile,
swarms. They propose some aggregation rules for
swarms with reciprocal and nonreciprocal interactions
between agents. They also point out that the swarming
behavior results from interplay between long-range
attraction and short-range repulsion among individuals.
For nonreciprocal interaction, a condition of weighted
momenta is assumed (Chu et al., 2003). In the present
paper, we extend the application of asymmetric
1 Mechanical Engineering Department, University of Connecticut, Storrs,
CT, USA2 Biomedical Engineering Program at the University of Connecticut,
Storrs, CT, USA3 Electrical and Electronic Engineering School at Universidad Industrial de
Santander, Colombia
Corresponding Author:
Nejat Olgac, Mechanical Engineering Department, University of
Connecticut, Storrs, CT, USA
Email: [email protected]
Journal of Vibration and Control
ª The Author(s) 2010Reprints and permission:sagepub.co.uk/journalsPermissions.navDOI: 10.1177/1077546309360047jvc.sagepub.com
17(5) 641–650
interaction forces to hostile swarm dynamics, which is
poorly studied in the literature.
Gazi and coworkers have also incorporated sliding
mode control and artificial potential functions to ensure
swarm aggregation even under uncertainties (Gazi,
2005; Yao et al., 2007). In Yao et al. (2007), the swarm
in a formation is guided to track a target. In these investi-
gations, the principal aim is the definition of a decentra-
lized formation control to increase robustness in the
task, taking into account the limited range and angle of
vision permitted by sensors.
Kumar and coworkers investigated the dynamic coordi-
nation of multiple robots to perform cooperative tasks
(Belta and Kumar, 2004; Chaimowicz et al., 2004). They
use a hybrid systems framework to model the cooperative
tasks and dynamic role assignment among multiple robots.
The approach coordinates the cooperative execution of the
task by considering the individual characteristics of each
one of the robots in the team. The strategy is also tested
using an experimental platform. They propose an abstrac-
tion of the configuration space to define the controllers in
a lower dimensional space. This results in improved com-
putational efficiency to command a large number of robots
(i.e., agents). This same research group has used artificial
potentials to maintain connectivity and to avoid collisions
(Zavlanos and Pappas, 2007).
Sierra et al. (2009) expand the momenta-based, homo-
genous swarm dynamics to include two hostile swarms of
pursuers and evaders, where the pursuers aim to catch the
evaders and the evaders attempt to escape. The interactions
between the pursuers and evaders, along with interactions
of alike members, make the stability analysis complicated.
To compensate for this, a Lyapunov-based dissipative con-
trol momentum is deployed when needed.
Jin and Gao (2008) expand the single integrator
model of Gazi and Passino (2004a) to a double integrator
model for homogenous swarms. They investigate the sta-
bility and swarm cohesion for several attraction/repul-
sion force functions in their second-order system. It is
shown that the members of the swarm will converge to
a bounded area and their velocities converge to a com-
mon velocity.
The main contribution of this paper is the expansion of
the works done by Sierra et al. (2009) and Jin and Gao
(2008) by addressing the stability and control of two groups
of antagonistic swarms with second-order dynamics. In
particular, we consider the pursuit of a swarm of ‘‘evaders’’
by a swarm of ‘‘pursuers,’’ an operation that includes het-
erogeneous agents and hostile interactions. Owing to the
asymmetric governing dynamics, as well as the pursuer/
evader assignment policies, the controller design and
assessment of stability become quite complex. For simpli-
city, within this study, we consider only cases with an equal
number of pursuers and evaders, all with the same mass,
and a simple assignment policy.
We begin by deploying models of the individual interac-
tions similar to those described in Jin and Gao (2008). The
major novelty, however, is the introduction of heterogene-
ity (i.e., two hostile swarms), which is described in Section
2. This section treats the antagonistic scenario in two
separate phases: approach and assigned pursuit. Control
terms are added to the pursuer dynamics to ensure that cap-
ture takes place. Section 3 outlines the Lyapunov stability
of the system and Section 4 presents some simulation
results. The conclusions and discussions suggest some
future steps in this research.
It should be noted that throughout the paper, boldface
letters represent matrices or vectors, and scalars are indi-
cated by italics.
2. Multi-agent swarm modeling
The dynamic modeling of the multi-agent swarms is pre-
sented here. We consider a two-phase approach for these
interactions. Phase 1 treats the two swarms as single agents
that are conceptually lumped at the respective swarm cen-
ters. The forces on individual agents during this phase are
assumed to be uniformly distributed. This phase brings the
two swarm centers sufficiently close to transition into
phase 2. In phase 2 (the assigned pursuit) the agent-to-
agent interaction logic enables the capture of each evader
by a pursuer. The two phases of the operation are combined
under one seamless scheme as follows:
€xi ¼ gPE �yð Þ þXNp
j¼1j6¼i
gpp xi � xj
� �þ
Trans �yk kð Þ gpe zli � xið Þ þ D _zli � _xiÞ þ K zli � xið Þ;ðð1aÞ
i ¼ 1 : Np;
€zi ¼ gEP �yð Þ þXNe
j¼1j6¼i
gee zi � zj
� �þ
Trans �yk kð ÞXNp
j¼1
gepðzi � xjÞ( )
i ¼ 1 : Ne;
ð1bÞ
where xi 2 <n, zi 2 <n are the local vectors of pursuers
and evaders in n-dimensional space. li denotes the evader
assigned to pursuer i. D _zli � _xiÞð and K zli � xið Þ are the
proposed control forces applied to the pursuer to facilitate
capture. Np and Ne represent the number of pursuers and
evaders, respectively. All agents are taken as point ele-
ments for simplicity. We will use the notation
yi ¼ zli � xi 2 <n to represent the vector connecting the
pursuer (i) to its assigned evader (li).
Journal of Vibration and Control 17(5)642
The notations gpe :ð Þ and gep :ð Þ in equations 1a and 1b
refer to the forces received by a pursuer due to its assigned
evader and the forces received by the evader due to the
surrounding pursuers, respectively. The force profiles
described below and shown in Figure 1 are biologically
inspired (Warburton and Lazarus, 1991; Camazine et al.,
2001):
gpe yð Þ ¼ qpe
Ape
Bpe þ yk kCpe
!2 <n
gep yð Þ ¼ qpe
Aep
Bep þ yk kCep
!2 <n:
ð2Þ
The unit vector qpe ¼ y= yk k gives direction to the forces.
Notice that gpe :ð Þ is an attraction force that vanishes as
yk k ! 1, as does gep :ð Þ, which is a repulsion force. This
force is bounded by the peak force, gpe
�� ��max¼ Ape
�Bpe,
which occurs at ype
�� ��max¼ 0. We define a distance, dpe,
as the influence range of a force profile where the force
is higher than 2% of the maximum value. It can be shown
that dpe ¼ 49Bpe
� � 1=Cpeð Þ. Similar analysis and definitions
are valid for gep :ð Þ and dep.
We use �x and �z for the local vectors of the centers of the
pursuer and the evader swarms, respectively, and �y is the
vector connecting them:
�x ¼ 1
Np
XNp
i¼1
xi; �z ¼ 1
Ne
XNe
i¼1
zi; �y ¼ �z� �x: ð3Þ
The phase 1 forces exerted on the pursuers and evaders are
given as gPE �yð Þ and gEP �yð Þ. These forces are the same as in
equation 2, except for different parameters. Note that the
force gPE �yð Þ is common to all pursuers since the vector �ydoes not change with respect to individual pursuers. The
same can be said for gEP �yð Þ and the evaders.
Notation gpp :ð Þ in equation 1a refers to the force
received by a pursuer due to other pursuers. Likewise,
gee :ð Þ in equation 1b represents the force received by an
evader due to other evaders. These force descriptions, pre-
sented in equations 4a and 4b, are shown in Figure 2. Note
that the parameter selection for both force profiles is iden-
tical, thus the two curves overlap. Using ui;j ¼ xi � xj
� �,
vi;j ¼ zi � zj
� �, we define the force on the ith member by
the jth member as
gppðuÞ ¼ qpp
App � Dpp uk kCpp
Bpp þ uk kCpp
!¼ qpp gpp; ð4aÞ
geeðvÞ ¼ qee
Aee � Dee vk k Cee
Bee þ vk kCee
!¼ qee gee: ð4bÞ
The unit vectors qpp ¼ u= uk k and qee ¼ v= vk k give direc-
tion to the forces. Note that these forces are characterized
as functions with bounded repulsion (in short proximity)
and constant attraction (at long distance). The maximum
repulsion occurs at upp
�� ��max¼ 0 with a value of
gpp
�� ��max¼ App
�Bpp. There is a distance, dpp, at which
gpp switches from attraction to repulsion (i.e., switches sign
or gpp
�� �� ¼ 0). This distance is defined as dpp ¼ App
��DppÞ 1=Cppð Þ. At this distance, the repulsion and attraction
forces are in balance.
The transition number, Trans dð Þ in equations 1a and 1b
is defined as
Trans dð Þ ¼ d0 þ d2d0 þ d
� �1� d � d0
d � d0k k þ d
� : ð5Þ
It represents the transition from phase 1, i.e., �yk k >> d0,
when Trans �yk kð Þ ffi 0, to phase 2, i.e., �yk k << d0, when
Trans �yk kð Þ ffi 1; d0 is the preferred distance for transition
0 5 10 15 20 250
4
8
12
||y||
For
ce
||gpe||
||gep||
Figure 1. Force profile in equation (2) for parameters in theexample section.
0 5 10 15 20 25
−0.2
0.2
0.6
1
0
For
ce
||u|| and ||v||
Attraction
Repulsion
− Dpp
, − Dee
γpp
and γee
Figure 2. Force profiles between like members for parametersin the example section (overlaid).
McCullough et al. 643
from phase 1 to 2, whereas the small positive parameter dcontrols the smoothness of this transition function (see
Figure 3); d0 is the distance where Trans d0ð Þ ffi 1=2.
The earlier formulation given in equations 1a and 1b has
three critical features:
(i) It seamlessly handles the dynamic transition between
phases 1 and 2.
(ii) Each pursuer has a relative exclusive nearest evader
(RENE) assignment, as detailed below.
(iii) There is a control term D _yi þ K yi, where D 2 <þand K 2 <þ, the selection of which is discussed
later. The function of these terms is to guarantee sta-
bility and capture of the evader by the pursuer.
The term gpe yið Þ in equation 1a, in fact, represents the
force on the ith pursuer due to the llth evader, which is
identified as the RENE for the ith pursuer. No other evader
but the RENE affects this pursuer. By contrast, all pursuers
influence all evaders as gep terms are under the summation
in equation 1b. In essence, every pursuer is engaged with
an evader called the RENE, and this assignment has
some subtleties:
� The assignment is performed on the relative positions of
the pursuers and evaders. This relative position is found
by conceptually shifting the two swarms such that the
swarm centers are coincident. RENE assignment is then
done in ascending distances (i.e., the closest relative pur-
suer–evader pair is assigned first and so on). The process
continues until all pairs are identified.
� RENE assignment is performed at all times, and sudden
changes in assignment can occur. However, during the
approach phase, the transition number is approximately
zero; therefore, the effect of the RENE assignment
(excluding the control action) is minuscule.
2.1. Phase 1: For �yk k >> d0, Trans �yk kð Þ ffi 0
If the centers of the swarms are separated by a distance
much greater than d0, �yk k >> d0, the resulting dynamics
leads to the merging of the two swarms. This is achieved
by treating the two swarms as two agents who are at the
center of each swarm. The forces gPE �yð Þ and gEP �yð Þ are
evenly distributed to the pursuers and evaders, respectively.
By taking Trans �yk kð Þ ffi 0, the dynamics in equations 1a
and 1b become
€�x ¼ 1Np
XNp
i¼1
gPE �yð Þ þXNp
j¼1j6¼i
gpp xi � xj
� �þ D _yi þ Kyi
0BB@
1CCA
¼ gPE �yð Þ þ D
Np
XNp
i¼1
_yi þK
Np
XNp
i¼1
yi;
ð6aÞ
€�z ¼ 1
Ne
XNe
i¼1
gEP �yð Þ þXNe
j¼1j 6¼i
gee zi � zj
� �0BB@
1CCA ¼ gEP �yð Þ:
ð6bÞ
The merging of the two swarms (i.e., phase 1) occurs in a
bounded approach time, tapp; furthermore, tapp strongly
depends on the relative initial positioning of the two
swarms with respect to each other. If the initial distance
between the two swarms is more than the maximum
recognition distance or the range of influence of the pur-
suers, which is depicted previously as dPE, the governing
forces are predominately the control forces on the pursuers
and the like member interaction forces gpp :ð Þ and gee :ð Þ,(i.e., gPE :ð Þk k � gEP :ð Þk k � 0). In this phase, if the
approach time is larger than each of the homogeneous
swarm convergence times, the area occupied by the evader
swarm will be dictated only by the nature of the interaction
forces gee :ð Þ. This area and time are evaluated in Jin and
Gao (2008), but only for homogenous swarms.
2.2. Phase 2: For �yk k << d0, Trans �yk kð Þ ffi 1
Once the two swarm centers are closer than the distance d0,
phase 2 (assigned pursuit) becomes dominant since �yk k is
diminishing. The property �yk k � d0 is enforced via the for-
mation of the forces gPE �yð Þ, gEP �yð Þ, and the control forces.
Any departure away from this region (i.e., creating
�yk k > d0) forces Trans �yk kð Þ to be very small in equations
1a and 1b, which, in turn, invites phase 1 dynamics (equa-
tions 6a and 6b) and forces �yk k � d0. Therefore, it is clear
that �yk k � d0 is stable and attainable. In phase 2 (when
Trans �yk kð Þ is closer to 1), the attraction force on a pursuer
by its assigned evader becomes more relevant. Similarly,
the evaders are repelled by all the pursuers in equation lb.
0 2 4 6 80
0.2
0.4
0.6
0.8
1
distance(d)
Tra
ns(d
)
d0
δ=0.1
δ=0.5
Figure 3. Function Trans(d), d0 ¼ 3.
Journal of Vibration and Control 17(5)644
With this pursuer–evader pairing we study the stability of
the dynamics next.
3. Lyapunov-based stability analysis
3.1. Stability analysis
In this section we analyze the stability of the system given
in equation 1 using LaSalle’s theorem (LaSalle, 1960).
Main Lemma: Consider the governing dynamics of a
particular pursuer and assigned evader pair, from equa-
tions 1a and 1b as
€yi ¼ f iðx; zÞ � D _yi � Kyi 2 <n; ð7Þ
where f iðx; zÞ contains all the self-evident interaction
forces from equations (1a) and (1b). It can be shown that
the pursuer–evader swarms approach a bounded state of
aggregation described by εk k � c ; c > 0, where
εi ¼yi
_yi
� �2 <2n; i ¼1 : Np; and ε ¼ col εif g2<2nNp:
Proof: Equation 7 leads to the individual state equations
_εi ¼_yi
€yi
� �¼
0 1
�K �D
�� In
� �yi
_yi
� �þ
0
f iðx; zÞ
� �¼ F� Inð Þ εi þ ~f i ¼ �Fεi þ ~f i 2 <2n i ¼ 1 : Np;
ð8Þ
where �F ¼ F� In 2 <2n�2n, the F matrix is self-evident,
In is an identity matrix with dimension n, and � denotes
Kronecker multiplication. Here, ~f i is the combination of
nonlinear interaction forces between pursuers and evaders
and alike agents, ~f i ¼ 0T f iðx; zÞT�
T 2 <2n with
f i x; zð Þ 2 <n. These forces are all bounded functions as
further analyzed later.
The combined system dynamics can be written in the
following form:
_ε ¼
_y1
€y1
..
.
_yNp
€yNp
0BBBBBB@
1CCCCCCA¼ Fð Þ
y1
_y1
..
.
yNp
_yNp
0BBBBBB@
1CCCCCCAþ
0
f1ðx; zÞ...
0
fNpðx; zÞ
0BBBBBB@
1CCCCCCA
¼ Fεþ f 2 <2nNp ;
ð9Þ
where F ¼ INp � �F 2 <2nNp�2nNp and f ¼ col ~f i
� �2 <2nNp i ¼ 1 : Np .
The Kronecker multiplication of two square matrices
M1 ðn�nÞ and M2 ðm�mÞ, M12 ¼M1 � M2 2 <nm� nm has a
critical property: the mn eigenvalues of M12 are formed
as dual multiplications of those eigenvalues of M1 and
M2, or in mathematical formalism lM12;ij ¼ lM1;i lM2;j
i ¼ 1:: n; j ¼ 1::m (Theorem 13.12 of Laub, 2005).
Furthermore, if one of these matrices, say M1, is an identity
matrix, the eigenvalues of M12 will be n repetition of the m
eigenvalues of M2, i.e., n-tupled lM2;1; :::lM2;m
� �. In short,
Kronecker multiplication Iðn�nÞ �M2 would repeat the
eigenvalues of M2 n times. We utilize this property twice
for F ¼ INp� ðF� In Þ 2 <2nNp � 2nNp in order to
determine the eigenvalues of F as l Fð Þf g ¼ Np-tupled
l �Fð Þf g ¼ nNp-tupled l Fð Þf g where Np-tupled f g implies
that the set f g is repeated Np times. From equations 8
and 9, the following features of the eigenvalues can be
stated:
l Fð Þf g ¼ l : l2 þ Dlþ K ¼ 0� �
2 C2;
l �Fð Þf g 2 C2n; l Fð Þf g 2 C2nNp:
Clearly for D 2 <þ and K 2 <þ, < l Fð Þð Þ < 0, thus they
represent a stable system.
Keeping the above discussions in mind, we propose a
candidate Lyapunov function to be
V ¼XNp
i¼1
Vi ¼XNp
i¼1
εTi P� Inð Þ εi ¼
XNp
i
εTi
�Pεi¼ εTPε:
ð10ÞP and �P are positive definite symmetric (PDS) matrices,
where �P ¼ P� In and P ¼ INp � �P. P is the solution to
the Lyapunov equation
FTPþ PF ¼ �Q 2 <2�2; ð11Þ
where Q is a PDS real matrix. Q and �Q also have similar
definitions to P and P. Using the features of the Kronecker
product, we expand the F;P , and Q matrices to handle
multiple pursuers in n-dimensional space without the added
complication of very large matrices. It can be shown that
the analysis of the dynamics and stability assessment can
be reduced to studying only the P and Q matrices instead
of the higher dimensional P; �P and Q; �Q.
Taking the derivative of equation 10, and knowing that
Q and �Q are positive definite and symmetric,
_V ¼XNp
i¼1
_Vi ¼XNp
i¼1
�_εT
i�Pε
iþ εT
i�P _ε�¼ _εT
Pεþ εTP _ε
¼�
Fεþ f�T
Pεþ εTPðFεþ fÞ
¼ �εTQεþ 2εTPf :
ð12Þ
Using the features of a PDS matrix A,
xTA y�� �� � lmax Að Þ xk k yk k; ð13aÞ
lmin Að Þ xk k2� xTAx � lmax Að Þ xk k2: ð13bÞ
Equation 12 can be reduced to
_V � �lmin Qð Þ εk k2þ2 εk k fk kmax lmax Pð Þ: ð14Þ
McCullough et al. 645
To guarantee _V � 0, the following condition must be met:
εk k � 2lmax Pð Þlmin Qð Þ fk kmax¼ 2lP=Q fk kmax ð15Þ
where
lP=Q ¼lmax Pð Þlmin Qð Þ ¼
lmax Pð Þlmin Qð Þ ¼ lP=Q;
This is true because the eigenvalues for P are the
n Np-tupled eigenvalues of P, which is also true for Q and
Q. This simplification reduces equation 15 to the following
definition of c:
εk k � 2lmax Pð Þlmin Qð Þ fk kmax¼ 2lP=Q fk kmax ¼definition c 2 <;
ð16Þ
which provides the upper bound of εk k as a hypersphere.
The dynamics given by equation 9 are guaranteed to be
entrapped within this hypersphere. This bound, however,
depends on the most pessimistic value of fk kmax, which
is studied later in the paper. The obvious aim in the follow-
ing segment is to minimize this upper bound by properly
selecting the Q matrix.
3.2. Selection of Q matrix
Using P ¼ p11 p12
p12 p22
�and the F matrix defined in equa-
tions 8 and 11 yields
Q ¼2Kp12 �p11 þ Dp12 þ Kp22
�p11 þ Dp12 þ Kp22 �2p12 þ 2Dp22
�
¼q11 q12
q12 q22
�:
The Q matrix is our selection and q11; q22 2 <þ. Back sol-
ving for the elements of P, one finds
P ¼Dq11
2K� q12 þ Kq22 þ q11
2D
q11
2K
q11
2K
Kq22 þ q11
2DK
24
35: ð17Þ
To analyze the effect that the values of the Q matrix have
on the ratio of eigenvalues, lP=Q, we introduce two con-
stants, c1 and c2, in the following manner:
Q ¼ q11 q12
q12 q22
�¼ q11 q11c1
q11c1 q11c2
�¼ q11
1 c1
c1 c2
�:
ð18Þ
For Q to be positive definite and symmetric, c21 < c2 and
c2 > 0 in addition to q11 > 0. Rewriting equation 17 using
the new values from equation 18, we redefine P as
P ¼ q11
1 þ Kc2
2D� c1 þ D
2K1
2K
12K
Kc2 þ 12DK
24
35: ð19Þ
The objective is to minimize c of equation 16, which
brings us to minimize lP=Q. It is clear that lP=Q is a function
of c1, c2, K, and D. It is independent of q11 due to the lin-
earity of equation 11. Therefore, the selection of q11 is
completely arbitrary and has no influence on the dynamics.
To minimize lP=Q we utilize a MATLAB routine called
‘‘fmincon.’’ It performs a Sequential Quadratic Program-
ming (SQP) procedure as detailed in The MathWorks Inc.
(2009). This procedure results in optimum values for
c1 ¼ 0; c2 ¼ 1, and K ¼ D!1. An indefinite increase
of Kand D results in increased control efforts and diminish-
ing benefits in the minimization of c. Therefore, the user
should select Kand D to resolve the trade-off between the
control effort and the bound of c.
Using the above numerical procedure, we demonstrate
that the lowest possible bound for lP=Q is 1. The Q matrix
is defined in order to determine P in equation 19 and the
bound of εk k in equation 16.
3.3. Calculation of maximum force
From the dynamics
f i x; zð Þ ¼ gEP �yð Þ � gPE �yð Þ þXNe
j¼1j6¼li
gee zli � zj
� �
�XNp
j¼1j 6¼i
gpp xi � xj
� �
þ Trans �yk kð ÞXNp
j¼1
gep zli � xj
� �� gpe yið Þ
" #:
Notice that these forces can be categorized into three types:
forces that are dominant in phase 1, phase 2, and like-agent
interactions (friendly forces):
Phase 1 ¼ gEP �yð Þ � gPE �yð Þ;
Friendly ¼XNe
j¼1j 6¼li
gee zli � zj
� ��XNp
j¼1j 6¼i
gpp xi � xj
� �;
Phase 2 ¼ Trans �yk kð ÞXNp
j¼1
gep zli � xj
� �� gpe yið Þ
" #:
We want a pessimistic maximum value for f ik k in order to
determine our lower bound of εk k in equation 16. That is,
f ik kmax � f i x; zð Þk k i ¼ 1 : Np:
Journal of Vibration and Control 17(5)646
To do this, we consider the worst possible scenario of the
vectorial addition of forces. Keeping in mind the general
force profile depicted in Figure 1, and all other pessimistic
compositions, the following bounds are reached:
Phase 1max ¼ gPEk kmax � gEPk kmax;
Friendlymax ¼ Np � 1� �
gpp
�� ��maxþ Ne � 1ð Þ geek kmax;
Phase 2max ¼ Np � 1� �
gep
�� ��max;
which results in a conservative value for the maximum
force, f ik kmax. Note that the gpe
�� ��max
term is assisting cap-
ture, thus is not included in phase 2. Recognizing that the
transition number is also a factor of the phase 2 forces, the
upper bound of this force can be written as follows:
f ik kmax ¼ gPEk kmax� gEPk kmaxþ Np � 1� �
gpp
�� ��max
þ Ne � 1ð Þ geek kmax
þ Trans �yk kð Þ Np � 1� �
gep
�� ��max:
ð20Þ
Since the transition number varies with time, f ik kmax also
varies with time. Knowing the value of f ik kmax, we can cal-
culate fk kmax for use in equation 16 as the Euclidean norm
of f:
fk kmax¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffif 2
1maxþ f 2
2maxþ þ f 2
Npmax
q¼
ffiffiffiffiffiffiNp
pf ik kmax:
ð21ÞThe ultimate objective of the control is the capture of the
evader by the pursuers. In order to measure the success
of the pursuers, capture and the time of capture, tcapture,
must also be defined. Logically, capture is achieved when
the relative positions and velocities of a pursuer–evader
pair approach zero. In order to use a comparable quantity
among different cases when deciding if capture is com-
pleted, a fraction of the value of c in pure phase 1 is set
as the threshold for capture. e.g.,
ε tcapture
� ��� �� < 0:02 c Trans �yk kð Þ¼0
�� ¼ 0:02 c0: ð22Þ
Once εk k is below the value of 0:02 c0, capture is declared.
Another aspect to consider in this pursuit is the amount
of control energy needed to stabilize the pursuers. This
control energy is directly associated with the selection of
K and D, as these parameters will introduce control forces
on the pursuers. In order to establish a scalar comparison
basis, we evaluate the control energy:
Econtrol ¼XNp
i¼1
ðtcapture
0
D _yi þ Kyið Þ � _xið Þ dt
0@
1A ð23Þ
This control energy represents the intensity of the control
effort throughout the pursuit. It is useful in selecting the
values of K and D. Of course, there is always a balance
between performance and minimizing the control effort.
To select the parameters of K and D, several simulations
were executed and the control energy and the performance
were compared.
4. Simulation results and performancetests
In this section we present two case studies that display
interesting properties of the proposed swarm control logic.
The first case is a 2 pursuers–2 evaders (2P–2E) scenario
using the parameters given in Table 1 and random initial
conditions. The values in the table are computed based
on the desired maximum force and influence range, d,
vis-a-vis Figures 1 and 2.
The second case is a 6P–6E scenario with randomly
selected initial swarm settings and the force profiles again
using the parameters listed in Table 1. The matrix Q was
selected for both cases:
Q ¼ 2KD 0
0 2KD
�:
From equation 11, this Q matrix yields the following P
matrix:
P ¼ K2 þ D2 þ K D
D K þ 1
�;
with K ¼ D ¼ 5, lP=Q ¼ 1:11:The selection of Q is consistent with the criteria pre-
sented in Section 3.2 and K and D were selected taking the
performance of several test cases and the associated control
energy into consideration.
Figures 4a and 5a show the Lyapunov function varia-
tions for the 2P–2E scenario and the 6P–6E scenario,
respectively. The black dots in Figure 5a represent
moments when a change in the RENE assignment is made.
The vertical dotted line in Figures 4a–d indicates the time
of capture, tcapture, which was defined previously in equa-
tion 22. Capture for the 2P–2E scenario is declared as
tcapture ¼ 4:13 s and for the 6P–6E scenario as
tcapture ¼ 3:81 s.
Figures 4b and 5b present violation flags to denote sam-
ple times at which violations to the negativity of the Lyapu-
nov candidate’s derivative are found, i.e., _V > 0. Red flags
Table 1. Parametric properties of the case studies.
Aee ¼ 23:4 Ape ¼ 171:3 APE ¼ 3608Bee ¼ 20:7 Bpe ¼ 13:8 BPE ¼ 180:3Cee ¼ 3 Cpe ¼ 3 CPE ¼ 3Dee ¼ 0:03 Aep ¼ 108:7 AEP ¼ 2623App ¼ 23:4 Bep ¼ 18:8 BEP ¼ 252:2Bpp ¼ 20:7 Cep ¼ 3 CEP ¼ 3Cpp ¼ 3 d ¼ 0:75 D ¼ K ¼ 5Dpp ¼ 0:03 d0 ¼ 3
McCullough et al. 647
indicate violations before capture takes place and the blue
dots indicate violations after capture is declared. Notice
that a single time-step violation of _V < 0 occurs at one
of the changes in the RENE assignment at t ¼ 0:05 s. This
violation can be observed by the single violation flag in
Figure 5b at this time. The other changes in the RENE
assignment do not result in _V > 0, although this cannot
always be guaranteed. Figures 4c and 5c present the time
history of the transition number, Trans �yk kð Þ, signaling how
fast the swarms merge in phase 1.
Figures 4d and 5d show the time evolution of the
magnitude of the error vector εk k and the time evolution
of the stability bound c for both sample cases. Notice how
c varies with time as it is dependent on the transition num-
ber. Lyapunov stability is guaranteed as long as the error
vector is larger than that of the stability bound. Once εk kis within the bound of c, _V < 0 can no longer be assured.
It is clear from the violation flags in Figures 4b and 5b that
violations of _V < 0 do not begin until well after εk k < c.
An exception appears as a single time step violation due
to the change in RENE at t ¼ 0:05 s. The violations after
εk k < c, however, are very small in comparison to the
scale of the Lyapunov function. Both scenarios result in suc-
cessful captures of the evaders by the pursuers, as marked by
the dashed line in Figures 4 and 5. One can observe that most
of the _V < 0 violations occur after the capture is declared
and these violations of the _V < 0 condition is caused only
by residual and ignorable oscillatory motion in pursuer and
evaders, which yield only minor increases in V .
Figures 6 and 7 present the trajectories of the pursuit
(black dotted line for pursuers and red for evaders) and the
final locations for the pursuers (black squares) and evaders
(red circles) for the 2P–2E scenario and the 6P–6E sce-
nario, respectively. The solid lines represent the motion
of the centers of each swarm (blue for pursuers and red for
evaders). It is clear that the pursuers capture the evaders
successfully as is evident by their positions being coinci-
dent. What cannot be observed in the figures is that the final
−50 0 50 100
−80
−60
−40
−20
0
20
40
X
Y
Figure 6. Traces of the 2P–2E sample case (red circles; evaders,black squares: pursuers).
0 2 4 60
5
10x 10 5
V
A
0 2 4 60.8
1
Vio
latio
n F
lags
B
0 2 4 60
0.5
1
Tra
ns #
C
0 2 4 60
100
200||ε||
ψ
time (s)
|| ε
|| an
d ψ
D
Figure 5. Simulation results for 6P–6E.
0 2 4 60
1
2x 10 6
V
A
0 2 4 60.8
1
Vio
latio
n F
lags
B
0 2 4 60
0.5
1
Tra
ns #
C
0 2 4 60
100|| ε ||
ψ
time (s)
||ε ||
and
ψ
D
Figure 4. Simulation results for 2P–2E.
Journal of Vibration and Control 17(5)648
velocities of each pursuer match that of its RENE, which is
another criterion for successful capture.
To illustrate the effect of the uncontrolled dynamics
(i.e., K ¼ D ¼ 0), a 2P–2E example case is studied. The
parameters used are identical to the previous 2P–2E case.
Figure 8 shows the variation of the transition number with
time. As is obvious from the figure, the dynamics never
reach phase 2 for more than an instant, thus undermining
the effect of the RENE on the pursuer. Without a dissipa-
tive term, there is nothing to slow the pursuers swarm as
it approaches the evaders. In fact, without the control terms
D and K, the system never reaches a sustained phase 2;
instead, the system oscillates in phase 1 as the dynamics
attempt to bring together the centers of the swarm. Capture
will not occur with these dynamics and the pursuers’ swarm
center will continually oscillate about the evaders’ center.
5. Discussion and conclusions
In this study we present a control strategy involving two
swarms that are in conflict. The membership counts of the
swarms are taken as equal for simplicity. The alike and
non-alike swarm members interact under some nonlinear
forces. The strategy has two phases: phase 1, called the
approach phase, brings the center of the swarms closer, and
phase 2, the assigned pursuit phase, enables a pursuit based
on an assignment strategy.
Control rules are based on the relative position and
velocity of the pursuer and evader. Lyapunov analysis
shows that the system is guaranteed to be stable. The pro-
posed Lyapunov candidate behaves properly until the
error vector enters a bounded region. Despite small
increases in the Lyapunov candidate within this region,
capture occurs due to the control forces. Another source
of violations of the Lyapunov function, which may be
nonnegligible, are those excursions appearing due to the
reassignment decisions of the RENE. This behavior does
not disrupt the deployment of Lyapunov stability in
piece-wise manner.
Current efforts look for other strategies to establish sta-
bility until capture is reached. Another associated area for
future research is the design and deployment of control
strategies for antagonistic swarms with second-order
dynamics and unequal membership strengths.
Acknowledgements
The authors wish to thank Prof. Eldridge Adams of the University
of Connecticut for his insightful suggestions for biologically
inspired force profiles. This work was supported in part by ARO
W911NF-07-1-0557 and DHS 2008-ST-061-TS0002.
References
Belta, C. and Kumar, V., 2004, ‘‘Abstraction and control for
groups of robots,’’ IEEE Transactions on Robotics 20(5),
865–875.
Camazine, S., Deneubourg, J., and Franks, N. R., 2001, Self-
Organization in Biological Systems, Princeton University
Press, Princeton.
Chaimowicz, L., Kumar, V., and Campos, M. F. M., 2004, ‘‘A
paradigm for dynamic coordination of multiple robots,’’
Autonomous Robots 17(1), 7–21.
Chu, T., Wang, L., and Chen, T., 2003, ‘‘Self-organized motion in
anisotropic swarms,’’ Journal of Control Theory and Applica-
tions 1(1), 77–81.
Chu, T., Wang, L., and Chen, T., 2006, ‘‘Complex emergent
dynamics of anisotropic swarms: convergence vs oscillation,’’
Chaos, Solitons & Fractals 30(4), 875–885.
−40 −20 0 20 40 60
−40
−30
−20
−10
0
10
20
30
40
50
60
X
Y
Figure 7. Traces of the 6P–6E sample case (red circles; evaders,black squares: pursuers).
0 50 100 150 2000
0.2
0.4
0.6
0.8
1
Tra
ns #
time(s)
Figure 8. Transition number with K ¼ 0;D ¼ 0.
McCullough et al. 649
Gazi, V., 2005, ‘‘Swarm aggregations using artificial potentials
and sliding-mode control,’’ IEEE Transactions on Robotics
21(6), 1208–1214.
Gazi, V. and Passino, K. M., 2003, ‘‘Stability analysis of swarms,’’
IEEE Transactions on Automatic Control 48(4), 692–697.
Gazi, V. and Passino, K. M., 2004a, ‘‘A class of attractions/repul-
sion functions for stable swarm aggregations,’’ International
Journal of Control 77(18), 1567–1579.
Gazi, V. and Passino, K. M., 2004b, ‘‘Stability analysis of social
foraging swarms,’’ IEEE Transactions on Systems, Man, and
Cybernetics, Part B: Cybernetics 34(1), 539–557.
Jin, D. and Gao, L., 2008, ‘‘Stability analysis of a double integra-
tor swarm model related to position and velocity,’’ Transac-
tions of the Institute of Measurement and Control 30(3-4),
275–293.
LaSalle, J. P., 1960, ‘‘Some extensions of Liapunov’s second
method,’’ IRE Transactions on Circuit Theory 7(4), 520–527.
Laub, A. J., 2005, Matrix Analysis for Scientists and Engineers,
SIAM, Philadelphia, PA.
The MathWorks Inc. ‘‘Optimization Toolbox: fmincon’’ The
MathWorks, Inc. 1994-2009. [Online]. Available: http://
ww.mathworks.com. [Accessed: March 20, 2009].
Sierra, D., McCullough, P., Olgac, N., and Adams, E. 2009,
‘‘Swarm coordination under conflict,’’ in Proceedings of the
American Control Conference, St Louis, MO, pp. 1021–1026.
Warburton, K. and Lazarus, J., 1991, ‘‘Tendency-distance models
of social cohesion in animal groups,’’ Journal of Theoretical
Biology 150(4), 473–488.
Yao, J., Ordonez, R., and Gazi, V., 2007, ‘‘Swarm tracking using
artificial potentials and sliding mode control,’’ ASME Journal of
Dynamic Systems, Measurement and Control, 129(5), 749–754.
Zavlanos, M. M. and Pappas, G. J., 2007, ‘‘Potential fields for
maintaining connectivity of mobile networks,’’ IEEE Transac-
tions on Robotics 23(4), 812–816.
Journal of Vibration and Control 17(5)650