Institute for Design and Control of Mechatronical Systems
Introduction aux Systèmes Collaboratifs Multi-Agents
UPJV, Département EEA
Fabio MORBIDI
Laboratoire MIS
Équipe Perception Robotique E-mail: [email protected]
Jeudi 13h30-16h30, Salle 8
M1 EEAII - Découverte de la Recherche (ViRob)
Introduction
Agreement is one of the fundamental problems in multi-agent coordination
– A collection of agents are to agree on a joint state value
We will study the dynamics of the so-called consensus protocol over static undirected and directed networks
We will see the interdependency between the convergence properties of the consensus protocol and the structural attributes of the underlying network
Consensus protocol
Undirected networks
Consensus protocol: a motivating example
A group of sensors are to measure the temperature in Picardie
20o 17o
19o 20o
21o 22o
19o
22o
Consensus protocol: a motivating example
Using an information-sharing network, the sensor group is required to agree on a single value (e.g. 20o) which represents the temperature of the region
For this, the sensor group needs a protocol over the network, allowing it to reach consensus on what the common sensor measurement value should be
20o 20o
20o
20o 20o
20o 20o
20o
Consensus protocol
The consensus protocol involves dynamic units, interconnected via relative information-exchange links
The rate of change of the state (e.g. the temperature) of each unit is
assumed to be governed by the sum of the relative states with respect to the neighboring units
Example:
1 2
3 4
Consensus protocol
Denoting the scalar state of node as one has then:
where is the neighborhood of node in the network (i.e. set of nodes adjacent to ).
The overall system can be rewritten in vector form as:
• is the Laplacian of the agents‘ interaction network
•
Consensus dynamics
Consensus protocol
Remark (Case of nonscalar states)
If (e.g. the 2D position of a robot) one can still obtain a
compact description of the consensus dynamics using the Kronecker product The Kronecker product of two matrices and , is the block matrix:
We then have that:
where is the identity matrix and
Consensus protocol: electrical example
Example (Resistor-capacitor circuit)
• Let us suppose that R = 1 Ω, C = 1 F.
• Kirchhoff‘s current and voltage laws lead to:
C
C C
R
R
which describes the dynamics
R
of the capacitors‘ voltages
represents the set of nodes in the circuit that are connected to node via a resistor
Consensus set
From the previous examples we note that the state of each vertex in the network is ‘‘pulled“ toward the states of the neighboring vertices
Will all vertices reach some weighted average of their initial states, which also corresponds to the fixed point of their collective dynamics?
The consensus set is the subspace i.e.
Definition (Consensus set)
Example:
Reaching consensus: undirected networks
The spectrum of the Laplacian for a connected undirected graph is:
and is the eigenvector corresponding to the zero eigenvalue Consider the spectral factorization of the Laplacian ( is a symmetric matrix):
where
is the
n
n
matrix consisting of normalized and mutually orthogonal eigenvectors of andand
The solution of initialized with is
Using the spectral factorization of the Laplacian, we get:
Reaching consensus: undirected networks
Let be a connected graph. Then the protocol converges to the consensus set with a rate of convergence that is dictated by
Proof:
Theorem
The proof follows directy from the equation
by observing that for a connected graph for as always Thus,
and hence As is the smallest positive eigenvalue of the Laplacian, it dictates the slowest mode of convergence in the limit above
(Average Consensus)
Reaching consensus: undirected networks
Example
The Laplacian spectrum of is:
Complete graph
Cycle
The Laplacian spectrum of is:
Then Then
Reaching consensus: undirected networks
Reaching consensus: remark I
Therefore, the quantity
i.e. the centroid of the network states, is a constant of motion for the consensus dynamics
Note that:
The state trajectory generated by the consensus protocol converges to the projection of its initial state onto the consensus subspace, since
Reaching consensus: remark II
A spanning tree of a connected graph is a tree composed of all the vertices and some (or perhaps all) of the edges of
Informally, a spanning tree of is a selection of edges of that form a tree spanning every vertex
Definition (Spanning tree) Definition (Tree)
A tree is connected graph without cycles
Reaching consensus: remark III
Proposition
A necessary and sufficient condition for the consensus protocol
to converge to the consensus set from an arbitrary initial condition , is that the underlying graph containts a spanning tree
Example:
Two possible spanning trees
of (blue)
Reaching consensus: remark III
Example 1: rendezvous problem
A collection of mobile agents with single integrator dynamics
This location is not given in advance and the agents do not have access to their global positions (no GPS)
The agents can only measure the relative displacement w.r.t. their neighbors are to meet at a single location. Here, denotes the position of agent
Consensus protocol ‘‘in action“
Consider the following control input for agent : Solution:
Example 1: rendezvous problem
Rendezvous point
Simulation (Network topology: )
Launch the Matlab file:
“Rendezvous.m”
Consider different network topologies and study the rate of convergence to the rendezvous point
Study the behavior of the system when all the agents initially lie on the same line (see the figure below)
Exercise
a1
a2
a3
a
Example 1: rendezvous problem
Remark Wheel graph Petersen graph
with 10 nodes
Example 2: cyclic pursuit (‘‘ n -bug problem‘‘)
Consider
n
mobile agents with single integrator dynamicsand suppose that we want agent to pursue agent modulo
n
. As before denotes the 2D position of agentWe can then select the following control input
for agent :
1
2
3
4
where is a positive constant
This results in the following system:
where and
Circulant matrix