Distributed Optimization Design Based on Passivity Technique
∗
Yutao Tang
1
, Yiguang Hong
2
, and Peng Yi
2
Abstract— In this paper, we study distributed optimization
based on passivity techniques. At first, we revisit the concept
of passivity respect to non-zero equilibrium point and convert
the stabilization problem into another problem that to find
a function with relative monotonicity. Then, a passivity-based
interpretation is proposed on existing centralized optimization
algorithms by adopting a continuous-time dynamic system view.
We utilize this passivity technique to solve a resource allocation
problem via distribute algorithms. Its effectiveness is verified
by a simulation example.
I. INTRODUCTION
Interest in distributed optimization of multi-agent systems
has been intense due to the emergence of the smart grid,
smart buildings, intelligent transportation systems, and social
networks ([1], [2], [3]). Roughly speaking, a distributed
optimization problem is to seek the feasible solution that
minimizes certain global objective functions of multi-agent
systems in a distributed manner.
While most distributed optimization algorithms are imple-
mented in discrete time (e.g., [4]), recent works ([5], [6])
have introduced continuous-time dynamic solvers, whose
convergence properties can be analyzed via classical stability
analysis. This has the advantage of facilitating the charac-
terization of properties such as basin of attraction, speed
of convergence, and robustness to uncertainty. For example,
a novel class of distributed continuous-time coordination
algorithms was proposed in [6] to solve distributed optimiza-
tion problems whose cost function is a sum of local cost
functions associated to the individual agents. Exponential
convergence of the proposed algorithm was established under
some connectivity assumption of the communication graph
and convexity of the objective functions. A similar opti-
mization problem for multi-agent systems subject to external
disturbances was studied in [7], where an internal-model
based protocol was proposed to reject those disturbances
while reaching the optimal point.
Among those results, there is a tendency to provide a
unified viewpoint for control, optimization, and coordination
of multi-agent systems. In fact, this idea can be traced back
to early work on optimization done by economists [8], which
*This work was supported by Fundamental Research Funds for the
Central Universities (24820152015RC36), NSFC (61503033, 61333001),
Beijing Natural Science Foundation (4152057), and Program 973
(2014CB845301/2/3)
1
Yutao Tang is with the School of Automation, Beijing
University of Posts and Telecommunications, Beijing, China
yttang@bupt.edu.cn
2
Yiguang Hong and Peng Yi are with the Key Laboratory of Systems
and Control, Institute of Systems Science, Chinese Academy of Sciences,
Beijing, China
yghong@iss.ac.cn, yipeng@amss.ac.cn
shows that there is a very natural continuous-time optimiza-
tion system associated with the Lagrangian function of the
convex optimization problem at hand. Such dynamic system
can be studied as a feedback control system. The connections
between duality in control theory and convex optimization
was also studied in [9]. Insight into the system-theoretic
meaning of the optimization problem can be very helpful in
developing efficient algorithms and proving its stability and
convergence rate. In [10], the stability of primal-dual gradient
dynamics was investigated by a Krasovskii’s method to find
Lyapunov functions. This method was proved to be useful
for these primal-dual laws in different scenarios. In [11], a PI
distributed optimization method was introduced to overcome
the limitation of diminishing step size in gradient searching
and allow fast asymptotic convergence. Distributed PID
control was also proposed for network congestion control
problems at the level of fluid flow models to improve the
transient performance and delay robustness of the overall
system by tuning the control gains [12].
On the other hand, a well-known concept in the control
society—passivity—has been extensively investigated due to
its explicit physical meaning and simplicity to manipulate,
for either a single plant or multi-agent systems (e.g.,[13],
[14]). For example, a passivity-based design was proposed
in [14] for coordination control of affine and passive agents,
which strictly includes the single integrator in [15], where
synchronization of Kuramoto oscillators and multiple La-
grangian systems were also studied. A unifying passivity
framework for network flow control was provided in [16],
which encompasses the existing stability results as its special
cases. In addition, the new approach significantly expands
the current classes of stable flow controllers by augmenting
the source and link update laws with passive dynamic
systems. Following this, the authors in [ 17] introduced a
new definition of passivity that serves better the need of
convergence analysis in dynamic networks without comput-
ing the convergence point a priori. It extends the traditional
notion of passivity to a forced system whose equilibrium is
dependent on the control input. However, some important
system classes are excluded, such as integrators, which was
remended in [18] by refining it as maximal equilibrium
independent passivity, where the duality between passivity-
based cooperative control and network optimization was
explicitly investigated and illustrated on a nonlinear traffic
dynamics model.
In this paper, we take a control-theoretic perspective to
build connections between the well-known passivity property
and (centralized/distributed) optimization problem. Different
from the existing publications on passivity and optimization
12th IEEE International Conference on Control & Automation (ICCA)
Kathmandu, Nepal, June 1-3, 2016
978-1-5090-1738-6/16/$31.00 ©2016 IEEE 732