Академический Документы
Профессиональный Документы
Культура Документы
Sandip Roy
Abstract This article examines whether purely decentralized controllers can be designed to stabilize networks of doubleintegrator agents with general observation topologies and identical non-minimum-phase internal dynamics. A new control
architecture is proposed, that permits stabilization of such
non-minimum-phase double-integrator networks. This design
provides an alternative to solutions that require information
exchange (of controller memory variables) between agents.
I. I NTRODUCTION
The problem of controlling a team of autonomous agents
with networked sensing or communication capabilities has
been intensely studied in the Controls community, with both
motion-control and algorithm-development goals in mind
(see, e.g.,[1], [2], [3], [4], [5]). In recent years, one thrust of
this research has been on developing controllers for increasingly sophisticated but also constrained agents, including
ones with increasingly complicated linear internal dynamics
[4], [6], [7] and those subject to actuator saturation and delay
[8], [9], [10]. Novel control schemes have been developed
that permit networks with such complicated agents to achieve
a range of control tasks, including notably stabilization,
formation-control, and synchronization tasks. There is now
a very wide literature on the control of multi-agent teams,
see e.g. [5] for an overview of recent results.
An interesting dichotomy has arisen in this research, that
is relevant to our development here: a number of studies
seek for purely decentralized solutions in which each agent
uses only its measurements of network dynamics in feedback
[4], [9], [10], [8], while other efforts permit communication of local storage or memory variables in accordance
with the measurement graph topology (i.e., an agent that
measures the dynamics of another may also receive storage
variables from that agent [6], [7]). While there have been
significant advances along both tracks, controller design has
been achieved for a wider class of agent models and network
topologies, and in a more systematic way, when communication of memory variables among the agents is permitted.
That is, communication of memory variables appears to
allow relaxation of several restrictions on both the network
topology and the agents dynamics, that arise when purely
(1)
The first two authors are with the School of Electrical Engineering and
Computer Science at Washington State University, and the last author is
with the Electrical Engineering Department at University of North Texas.
This work has been generously supported by National Science Foundation
Grants ECS-0901137, CNS-1035369, and CNS-1058124. Correspondence
should be sent to sroy@eecs.wsu.edu.
Yan Wan
61
n
X
gij zj ,
(2)
j=1
zj = cT xj ,
(3)
c
the local output matrix c = 1 is a two-element vector
c2
that specifies the (homogeneous) local output statistic, and
the weights gij specify how the local output statistics of
multiple agents are combined in the observation. We find
it convenient to define a matrix G = [gij ], and term this
matrix of weights as the topology matrix since it specifies the
sensing/communication topology among the agents (see e.g.
[12], [4]). We stress that each agent is assumed to only have
available the measurement yi , not the local output zi ; such
models have been termed non-introspective in the literature,
e.g. [7]. We also stress that no restrictions are placed on
the topology matrix: each agents measurements may be
arbitrary linear combinations of the output statistics, and so
the weights are arbitrary (positive or negative) real numbers.
We refer to the above-described model as the generalized
double-integrator network
(GDIN), noting that the special
1
case where c =
has been widely considered in the
0
literature under the heading of double-integrator network. We
also refer to the triple (c, A, b) as the full agent model, since
it specifies the input-to-local-output dynamics of each agent.
A particular focus of this article will be on the case that
the full agent model is non-minimum-phase. In this case,
we will refer to the network model as a non-minimum-phase
double-integrator network (NMPDIN).
We find it convenient to represent the full open-loop
dynamics of the GDIN as a single state-space system.
Specifically, we define xT = [x1,1 , ..., x1,n |x2,1 , ..., x2,n ]T ,
uT = [u1 , ..., un ]T , and yT = [y1 , ..., yn ]T , where xj,i is the
jth entry (j = 1, 2) of xi . In this notation, the open-loop
dynamics is given by
0 In
0
x =
x+
u, y = c1 G c2 G x. (4)
0 0
In
Using Wang and Davisons classical result, we obtain necessary and sufficient conditions on a GDIN, for
stabilization using a decentralized dynamic linear timeinvariant controller. The state-of-the-art in decentralized and
communications-allowing control of the GDIN is then discussed, in the context of this existence result. Here is the
condition for stabilization:
Theorem 1: A decentralized dynamic LTI controller can
be applied to the GDIN to achieve asymptotic stability, if
and only if the topology matrix G has full rank and (c, A)
is observable (which happens if and only if c1 6= 0).
Proof: The result of Wang and Davison [11] states that a
decentralized LTI controller can be used to stabilize a system
if and only if its decentralized fixed modes are all in the open
left-half of the complex plane (OLHP). Since in our case all
eigenvalues of the open-loop system are zero, decentralized
stabilization is possible if and only if 0 is not a decentralized
fixed mode. Per the definition of a decentralized fixed mode
[11], zero is a decentralized fixed mode of the GDIN if and
only if, for any n n diagonal matrix K, the following
determinant is equal to 0:
det(
0
0
In
0
+
K c1 G
0
In
c2 G ) = det(
0
c1 KG
In
).
c2 KG
(5)
For the GDIN, the condition for decentralized stabilization (or absence of closed-right-half-plane fixed modes)
62
is identical to the condition for (centralized) stabilizability and detectability. Hence, decentralized stabilization using dynamic LTI controllers can be achieved in all cases
that centralized control (by any type of controller) can be
achieved. This equivalence also immediately implies that
purely-decentralized LTI controllers can achieve stabilization
in all cases that an information-transmission-allowing controller or a non-linear time-varying decentralized controller
(see e.g. [17]) achieves stabilization.
Although application of Wang and Davisons classical result indicates that decentralized stabilization of the GDIN can
be achieved in full generality, this analysis does not directly
yield a practical controller designit is an existence result.
Several recent research efforts have proposed new controller
architectures for multi-agent networks, that yield practical
decentralized controllers for the GDIN. Two decentralized
control schemes are particularly relevant. First, the multilead-compensator paradigm introduced in [9], [10] permits
stabilization of the GDIN for arbitrary full-rank G, but only
for agents that are minimum-phase. On the other hand, the
observer-based design developed in e.g. [8] can be used
for an arbitrary full agent model, but requires G to have a
special form. Specifically, the authors in [8] (which focuses
on synchronization, not stabilization) assumed a directedLaplacian topology matrix G; similar designs for stabilization can be obtained for grounded-Laplacian matrices G or
(more broadly) for matrices G whose eigenvalues can be
placed in a single open half plane through diagonal scaling,
but the method does not work for arbitrary full-rank G.
Several further controller designs are also available including
a passivity-based approach (see e.g. [16]), however these
methods also only apply to a subset of stabilizable GDINs. In
sum, the existing approaches do not permit design of stabilizing purely-decentralized controllers, for non-minimum-phase
agents and arbitrary full-rank graph matrices.
In comparison with purely decentralized controllers,
schemes that allow communication of memory variables can
be designed to achieve synchronization or stabilization, for
a much broader class of multi-agent networks (including
ones with heterogeneous local agent models [7]). While
most of the communications-allowing control schemes have
been focused on synchronization problems, these results can
be straightforwardly adapted to stabilization problems. In
particular, it is easy to check that a communications-allowing
controller can be designed for the GDIN in full generality.
(6)
where the GDIN inputs ui (t) and the signals vi (t) are
controller memory (state) variables, ki and i are gain
parameters that may be different for each agent, and z, , ,
and are scalar gain parameters. We note that the proposed
control architecture is fully decentralized. We also note that
the proposed control uses derivatives of the output, and so
cannot be directly implemented as a linear state-space system
in the specified architecture. However, we will later check
that its transfer function is proper, and hence verify that a
state-space implementation is possible.
Let us now present and prove the key result of this section,
63
0 I 0
0
Acl1 = 0 0 I 0 (c2 KG + I)1 H, (8)
0 0 0
I
x
v1
x
.
Acl =
0
0
0
zc1 KG
In
0
0
zc2 KG + c1 KG
0
In
0
(c2 + c1 )K(G +
b)
0
,
In
c2 KG + I
(7)
k1
where K = diag( ... ),
b =
1
1
,
c2 +c1 K
and =
kn
1
diag( ... ). Here, we have chosen to phrase the state matrix
n
in terms of
b rather than for convenience in design; the
gains can be computed from
b.
To commence the design, we choose = 1. Also,
we claim that K and
b can be designed so that: 1) the
eigenvalues of c1 K(G+b
) are in the OLHP, 2) the maximum
magnitude among the eigenvalues of K(G+b
) is 1, and 3)
b
is small compared to G in the sense that its largest magnitude
entry is much smaller than the smallest-magnitude eigenvalue
of G (say, less by a scale factor f , where f << 1). To show
that this is possible, consider
b = cI, where c is a small
constant. We claim that there exists c > 0 such that, for
b = G + cI
all c (0, c ], all principal submatrices of G +
are of full rank. To see this, consider Gq + cI, where Gq
is a principal submatrix of G (and the identity matrix I
is of commensurate dimension); this quantity is not of full
rank only if c is an eigenvalue of Gq . However, since each
principal submatrix Gq has a finite number of eigenvalues,
and G only has a finite number of principal submatrices,
all principal submatrices of G + cI have full rank except
for a finite set of values c. It immediately follows that all
principal submatrices of G+cI have full rank for c (0, c ],
for some c . Choosing sufficiently small c in this range, we
immediately obtain a
b that is small compared to G in the
sense that its largest magnitude entry is much smaller than
the smallest-magnitude eigenvalue of G. Further, for this
b,
K can be designed so that the eigenvalues of c1 K(G +
b)
where
H = zc1 KG
zc2 KG + c1 KG (c2 + c1 )K(G +
b) .
(9)
Simplifying the Schur form and eliminating terms of lower
order from each block on the last row, we obtain that the
Schur form (to within a small perturbation) is
0
I
0
.
0
I
Acl1 = 0
(10)
zc1 KG c1 KG c1 K(G +
b)
It remains to characterize the eigenvalues of Acl1 , so as
to characterize all eigenvalues of the closed-loop system.
Let us now focus on the bottom-right block of Acl1 . By
construction, the eigenvalues of this bottom-right block are
in the OLHP, and further they are O(). Again, it is easy to
check that, upon a classical time-scale state transformation,
the remainder of the matrix Acl1 constitutes a small perturbation of the lower-right block. Hence, the eigenvalues
are the union of those of lower-right block and those of the
following Schur form:
0 I
0
Acl2 =
(c1 K(G+b
))1 zc1 KG c1 KG
0 0
I
(11)
To simplify this expression, we recall that
b = cI has been
designed to be small compared to the eigenvalues of G
(by a scale factor of f ), and hence (c1 K(G +
b))1 is a
64
z
I I,
4) This study resolves the gap between purely decentralized controller designs and communication-allowing designs,
for the GDIN. It remains an open question, however, if the
presented design philosophy can be adapted to more complex
multi-agent systems. Much further study is also needed regarding adapting the design to overcome actuator saturation,
delay, and topological variation. We have developed some
preliminary simulations suggesting that a scaling of the controller gains can be used to permit stabilization under actuator
saturation, however further effort is needed to verify the
result. We stress that communications-allowing designs may
have many advantages over the purely-decentralized design
presented here, but our design does provide an alternative
in the case that communication of memory variables in
accordance with the network graph is not possible.
5) Either the decentralized controller design introduced
here or the one given in [10] can be used when the full
agent model is minimum phase. For the NMPDIN, however,
only the design presented here is stabilizing.
6) It is worth stressing that, although time-scale separation
techniques are very common in controller design, highgain feedback of an output derivative equal to the relative
degree of the plant is uncommon. Further study is needed
to evaluate e.g. the external stability and robustness of such
designs, however we believe that the design presented here
is a promising first step. We also stress that such a high
gain solution is needed only if the agent model is nonminimum phase and the graph matrix has eigenvalues with
both negative and positive real parts.
(13)
Example
Let us consider stabilization of a NMPDIN with G =
0 1
, c1 = 1, and c2 = 1. We note that the eigenvalues
1 0
of the topology matrix G in this case cannot be placed in
a single half plane through diagonal scaling, so previouslydeveloped purely-decentralized control schemes cannot be
used. In fact, this NMPDIN is particularly difficult to control,
in that two agents have no local information (in addition
to being non-minimum phase). Using the design method
developed above, we obtain the following stabilizing controller design: k1 = 1, k2 = 1, = 1, = 103 ,
1 = 2 = 3 105 , = 2 109 , and z = 1015 .
For this design, we would expect the 8 eigenvalues of
the closed-loop system to be in the OLHP, at three time
scales: two eigenvalues should be near 1, two eigenvalues
should have magnitudes near 103 (and real parts that are
significantly larger than 106 ), and four eigenvalues should
be near 106 . In fact, the eight eigenvalues of the closed-loop
system are: 1.00 j2.00E 3, 2.60E 5 1.00E 3,
1.12E 6 1.5E 7, and 8.82E 7 9.5E 8 (where
the operator E represents the base-10 exponential). We note
that the two modes corresponding to the middle time scale
are quite oscillatory: this is not surprising given that neither
agent has local measurements.
65
R EFERENCES
[1] W. Ren and R. Beard, Distributed consensus in multi-vehicle cooperative control: theory and applications, Springer: New York, 2008.
[2] S. Roy, A. Saberi, and A. Stoorvogel, eds., International Journal of
Robust and Nonlinear Control, Special Issue on Communicating-Agent
Networks,vol. 17, no. 10-11, Dec. 2006.
[3] L. Xiao and S. Boyd, Fast linear iterations for distributed averaging,
Systems and Control Lettters, vol. 53, no. 1, pp. 65-78, Sep. 2004.
[4] T. Yang, S. Roy, Y. Wan, and A. Saberi, Constructing consensus
controllers for networks with general linear agents, International
Journal of Robust and Nonlinear Control, vol. 21, no. 11, pp. 12371256, Jul. 2011.
[5] Y. Cao, W. Yu, W. Ren, and G. Chen, An overview of recent progress
in the study of distributed multi-agent coordination, arXiv:1207.3231
.
[6] Z. Li, Z. Duan, G. Chen, and L. Huang, Consensus of multi-agent
systems and synchronization of networks: a unified viewpoint, IEEE
Transactions on Circuits and Systems I, vol. 57, no. 1, Jan. 2010.
[7] H. F. Grip, T. Yang, A. Saberi, and A. A. Stoorvogel, Output synchronization for heterogeneous networks of non-introspective agents.
[8] X. Wang, A. Saberi, A. Stoorvogel, H. F. Grip, and T. Yang, Consensus in the network with uniform constant communication delay,
submitted to Automatica.
[9] Y. Wan, S. Roy, A. Saberi, and A. Stoorvogel, The design of
multi-lead-compensators for stabilization and pole placement in
double-integrator networks, IEEE Transactions on Automatic Control,
vol. 55, no.12, pp. 2870-2875.
[10] Y. Wan, S. Roy, A. Saberi, and A. Stoorvogel, A multiple-derivative
and multiple-delay paradigm for decentralized control: uniform rank
systems, Dynamics of Continuous, Discrete, and Impulsive Systems,
Special Issue in Honor of H. Khalils 60th Birthday, vol. 17, no. 6,
pp. 883-907, 2010.
66