Consensus algorithms L. Giarr´ e* * DIETT, Universit` a di Palermo joint work with D. Bauso and R. Pesenti SCUOLA OTTIMIZZAZIONE - TEORIA E APPLICAZIONI Palermo 2011
Consensus algorithms
L. Giarre*
* DIETT, Universita di Palermo
joint work with D. Bauso and R. Pesenti
SCUOLA OTTIMIZZAZIONE - TEORIA E APPLICAZIONI
Palermo 2011
DIETT 1
Outline
� Introduction
– Applications
– Why consensus problems
– Graph Theory
– The Consensus Problem
– Reference
• Distributed consensus protocols for coordinating buyers
• Mechanism Design for Optimal Consensus Problems
• Lazy consensus for networks with unknown but bounded
disturbances
• References
SCUOLA DOTTORATO- PALERMO 2011
DIETT 2
Applications:
• Cooperative control for multi-agent systems:
• formation control problems with applications to mobile robots
unmanned air vehicles (UAVs), autonomous underwater vehicles
(AUVs), satellites, aircraft, spacecraft, and automated highway
systems
• nonformation control problems such as task assignment, payload
transport, role assignment, air traffic control, timing, and search,
inventory problems, sensor networks, distributed estimation,
syncronization
SCUOLA DOTTORATO- PALERMO 2011
DIETT 3
Shared Information
• Definition and management of shared information among a group
of agents to facilitate the coordination among agents
• The shared information may take the form of common objectives,
common control algorithms, relative position information, ...
• Examples of information necessary for cooperation:
– Relative position sensors may enable vehicles to construct
state information for other vehicles (UAV’s formations)
– joint knowledge might be pre-programmed into the vehicles
before a mission begins (robotics team)
SCUOLA DOTTORATO- PALERMO 2011
DIETT 4
Why Consensus Problems?
• A team of agents must be able to respond to unanticipated
situations sensed as a cooperative task
• As the environment changes, the agents on the team must be in
agreement as to what changes took place.
• A consequence that shared information is assumed necessary for
coordination → cooperation requires that the group of agents
reach consensus on the coordination data.
• Convergence to a common value is called agreement or consensus
problems.
• Consensus problems have a history in computer science. We focus
on applications to cooperative control of multi-agent systems.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 5
Main aspects
• Challenging aspects:
– decentralized information
– decentralized decision making
– distributed objectives
• Agents may need to be in agreement regarding a certain quantity
of interest
– global state (information sharing)
– future decisions (coordinated decision making)
– common/individual goals
SCUOLA DOTTORATO- PALERMO 2011
DIETT 6
Graph theory
• It is natural to model information exchange between agents in a
cooperative team by directed/undirected graphs
• A digraph (directed graph) consists of a pair (N,E), where N is
a finite nonempty set of nodes and E is a set of ordered pairs of
nodes, called edges.
• The pairs of nodes in an undirected graph are unordered.
• A directed path is a sequence of ordered edges of the form
(vi1, vi2), (vi2, vi3), , where vij ∈ N , in a digraph.
• An undirected path in an undirected graph is defined
analogously, where (vij , vik) implies (vik, vij).
SCUOLA DOTTORATO- PALERMO 2011
DIETT 7
Graph theory
• A digraph is called strongly connected if there is a directed path
from every node to every other node.
• An undirected graph is called connected if there is a path
between any distinct pair of nodes.
• A directed tree is a digraph, where every node, except the root,
has exactly one parent.
• A spanning tree of a digraph is a directed tree formed by graph
edges that connect all the nodes of the graph. We say that a
graph has (or contains) a spanning tree if there exists a spanning
tree that is a subset of the graph.
• Note that the condition that a digraph has a spanning tree is
equivalent to the case that there exists a node having a directed
path to all other nodes.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 8
Graph theory
• The adjacency matrix A = [aij ] of a weighted digraph is defined
as aii = 0 and aij > 0 if(j, i) ∈ E where i = j.
• The Laplacian matrix of the weighted digraph is defined as
L = [ℓij ], where ℓii =∑
j aij and ℓij = −aij where i = j.
• For an undirected graph, the Laplacian matrix is symmetric
positive semi-definite.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 9
Example
• A team of 4 UAVs in longitudinal flight and initially at different
heights.
• Each UAV controls the vertical rate without knowing the relative
position of all UAVs but only of neighbors
• The information flow in a network of 4 agents ( the
communication network topology) − > A =? L =?
SCUOLA DOTTORATO- PALERMO 2011
DIETT 10
Example
v1 v3 v4
v2
SCUOLA DOTTORATO- PALERMO 2011
DIETT 11
Example
• A =
0 1 1 1
1 0 1 0
1 1 0 0
1 0 0 0
• L =
3 −1 −1 −1
−1 2 −1 0
−1 −1 2 0
−1 0 0 1
SCUOLA DOTTORATO- PALERMO 2011
DIETT 12
Graph theory
• The incidence matrix C = [cij ] has a row for each node, an a
column for each arc.
• The ordering of the arc is arbitrary.
• Defining a direction (arbitrary if the digraph is undirected, for
example arc ij, consider the direction from Ni to Nj).
• The Laplacian matrix of the weighted digraph is defined as
L = C ∗ CT .
SCUOLA DOTTORATO- PALERMO 2011
DIETT 13
Example
v2
v3
v1
v4
•C =
1 1 1 0
−1 0 0 1
0 −1 0 −1
0 0 −1 0
e12 e13 e14 e23
• L = C ∗ CT
SCUOLA DOTTORATO- PALERMO 2011
DIETT 14
Graph theory
• For a graph G and its Laplacian matrix L with eigenvalues
λ0 ≤ λ1 ≤ . . . ≤ λn−1
• L is always positive-semidefinite (∀i, λi ≥ 0).
• The second smallest eigenvalue of graph Laplacians λ1 is called
the algebraic connectivity.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 15
Example
• L =
3 −1 −1 −1
−1 2 −1 0
−1 −1 2 0
−1 0 0 1
• Eigenvalues: 0, 1, 3, 4
SCUOLA DOTTORATO- PALERMO 2011
DIETT 16
Networks of Dynamic Agents
),( )(iii xxfx =&
• Consider a system of n dynamic agents Γ = {1, . . . , n}
• Model interaction through connected network (graph) G = (Γ, E)
• let xi be the state of agent i
SCUOLA DOTTORATO- PALERMO 2011
DIETT 17
Consensus Problem
• Distributed and stationary control policy (protocol )
xi = ui(xi, x(i)) ∀i ∈ Γ.
• Let χ : IRn → IR be the agreement function, usually
– χ(ξ) = ave(ξ1, . . . , ξn)
• Consensus reaching
∥xi − χ(x(0))∥ −→ 0 for t −→ ∞.
• System converges to
χ(x(0))1
.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 18
A Consensus Algorithm for Averaging
• A simple consensus algorithm to reach an AVERAGING
agreement regarding the state of n integrator agents with
dynamics xi = ui
• can be expressed as an nth-order linear system on a graph:
xi(t) =∑j∈Ni
(xj(t)− xi(t)), xi(0) = zi
• The collective dynamics of the group of agents following protocol
can be written as
x = −Lx
where L = [lij ] is the graph Laplacian of the network.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 19
Averaging
For a connected network:
• The consensus value is equal to the average of the initial values.
• Irrespective of the initial value of the state of each agent, all
agents reach an asymptotic consensus regarding the value of the
function
f(z) =1
n
∑i
zi
SCUOLA DOTTORATO- PALERMO 2011
DIETT 20
MIMO system
s
s
s
100
0
01
0
001
�
���
�
�
L
b +
-
y=x u
Input bias
Consensus feedback
output
SCUOLA DOTTORATO- PALERMO 2011
DIETT 21
Example
v1 v3 v4
v2
x = −Lx
with L =
3 −1 −1 −1
−1 2 −1 0
−1 −1 2 0
−1 0 0 1
• Eigenvalues: 0, 1, 3, 4. The second smallest eigenvalue is −1
• It quantifies the speed of convergence of consensus algorithms.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 22
Comments on the Linear algorithm
• We recall that the above algorithm is working for linear protocols.
• The linearity can be comprised by many factors: saturations,
disturbances, noisy measurements, etc...
• The convergence is only guaranteed for arithmetic mean.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 23
Comments on the Topology
• Switching topology: some link can be lost (e.g. a dropping in a
transmission channel)
• If the network is still connected, under certain hypothesis on the
switching time, it is possible to prove convergence.
• The topology is not connected at each instant, but it is on
averaging.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 24
Applications
• J. Cortes, S. Martinez, and F. Bullo, “Robust rendezvous for
mobile autonomous agents via proximity graphs in arbitrary
dimensions”, IEEE TAC, 2004.
• R. Olfati-Saber, J. S. Shamma, “Consensus filters for sensors
networks and distributed sensor fusion”, CDC-ECC 2005.
• L. Xiao, S. Boyd, “ Fast linear iterations for distributed
averaging”, Systems and Control Letters 2004.
• A. Papachristodoulou, A. Jadbabaie, “Synchronization in
oscillator networks: switching topologies and non-homogeneous
delays”, TAC 2003.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 25
Literature
• N. A. Lynch, Distributed Algorithms. San Francisco, California:
Morgan Kaufmann Publishers, Inc., 1996.
• L. Moreau, “Stability of multi-agent systems with
time-dependent communication links”, TAC 2005.
• D. Angeli, P.-A. Bliman, “Extension of a result by Moreau on
stability of leaderless multi-agent systems”, CDC 2005.
Surveys
• W. Ren, R. Beard, E. M. Atkins, “Information consensus in
multi-vehicle cooperative control”, IEEE Control Systems
Magazine 2007.
• Olfati-Saber, Fax, Murray, “Consensus and Cooperation in
Networked Multi-Agent Systems”, Proc. IEEE 2007.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 26
Outline
• Introduction
– Applications
– Why consensus problems
– Graph Theory
– The Consensus Problem
– References
� Distributed consensus protocols for coordinating buyers
• Mechanism Design for Optimal Consensus Problems
• Lazy consensus for networks with unknown but bounded
disturbances
• References
SCUOLA DOTTORATO- PALERMO 2011
DIETT 27
Problem Setting
• Consider a group of buyers (agents or Decision makers, DM )
aiming at coordinating their daily ordering decisions.
• Motivation: sharing fixed transportation costs.
• Coordination requirements for each buyer is expressed in terms of
a minimum threshold l on the number of buyers to coordinate
order with.
• Corresponding to a maximum threshold on the fixed
transportation.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 28
Problem Setting
• Need of a protocol for information interchange among buyers such
that the number of active buyers is maximized.
• Active buyers = the buyers that eventually place orders, on the
basis of the available information.
• This is a consensus problem.
• The desired distributed protocol should turn out to be a
consensus protocol that allows all the node states to eventually
converge to the exact percentage of active buyers.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 29
Problem Statement
• Coordination can be achieved in a decentralized setting without
the necessity that each buyer communicates the threshold to a
central Decision Maker.
• The problem is how to reach the maximum coordination without
letting them communicate their threshold to every agents, but
only to the neighbors.
• What to communicate?
• How to find the policies to reach the best equilibrium?
• How to get convergence?
SCUOLA DOTTORATO- PALERMO 2011
DIETT 30
Problem Solution
• At the beginning of the day each buyer exchanges information
with its neighbors regarding its initial local estimate about the
percentage of active buyers.
• Information propagates in a decentralized setting and converges
to a common decision-value on the estimate within a
pre-specified time interval.
• Once convergence is reached, the current active buyers
synchronize their new decision to give up ordering if the
decision-value is lower than their threshold.
• Consensus on the final number of active buyers is reached
asymptotically and coordination is the best achievable for the
assigned thresholds.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 31
Final Remarks
• The presented protocol can be seen as an algorithm to find the
best strategies associated to the best Nash equilibrium.
• It guarantees the BEST cooperative solution in a
NONCOOPERATIVE context where each buyers is maximizing
its cost.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 32
Outline
• Introduction
– Applications
– Why consensus problems
– Graph Theory
– The Consensus Problem
– References
• Distributed consensus protocols for coordinating buyers
� Mechanism Design for Optimal Consensus Problems
• Lazy consensus for networks with unknown but bounded
disturbances
SCUOLA DOTTORATO- PALERMO 2011
DIETT 33
Main Problems:
• The consensus protocol answers to the question of determining
the policies that the agents must implement to reach a given
consensus.
• The mechanism design problem answers to the question of which
policy is implemented, and hence which consensus value is
reached, by selfish agents with given individual objectives.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 34
Networks of Dynamic Agents
),( )(iii xxfx =&
• Consider a system of n dynamic agents Γ = {1, . . . , n}
• Model interaction through connected network (graph) G = (Γ, E)
SCUOLA DOTTORATO- PALERMO 2011
DIETT 35
• let xi be the state of agent i
• let x(i) collect the states of its neighbors
SCUOLA DOTTORATO- PALERMO 2011
DIETT 36
Consensus Problem
• Distributed and stationary control policy
xi = ui(xi, x(i)) ∀i ∈ Γ.
• Let χ : IRn → IR be the agreement function, usually
– χ(ξ) = ave(ξ1, . . . , ξn)
– χ(ξ) = min(ξ1, . . . , ξn)
– χ(ξ) = max(ξ1, . . . , ξn)
• Consensus reaching
∥xi − χ(x(0))∥ −→ 0 for t −→ ∞.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 37
Consensus problem
Problem 1 (Consensus Problem) Consider a network G = (Γ, E) of
dynamic agents with first-order dynamics. For any function χ
determine a (distributed stationary) protocol, whose components have
the feedback form xi = ui(xi, x(i)), that makes the agents
asymptotically reach consensus on χ(x(0)) for any initial state x(0).
SCUOLA DOTTORATO- PALERMO 2011
DIETT 38
Agreement on any mean of order p
• χ is any mean of order p not only ave/min/max.
• The group decision value is permutation invariant function of the
agents’ initial states: the value is independent of the agents
indexes: χ(ξ1, ξ2, . . . , ξn) = χ(ξσ(1), ξσ(2), . . . , ξσ(n)) (Perm. Inv.)
• The agreement function is confinated between:
mini∈Γ{ξi} ≤ χ(ξ) ≤ maxi∈Γ{ξi}, for all ξ ∈ IRn.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 39
Structure
mean χ(x) f(y) g(z)
arithmetic∑
i∈Γ1nxi
1ny z
geometric n√∏
i∈Γ xi e1ny log z
harmonic 1∑i∈Γ
nxi
ny
1z
mean of order p p
√∑i∈Γ
1nx
pi
q
√1ny zp
• Structure χ(ξ) = f(∑
i∈Γ g(ξi)),
• f, g : IR → IR with dg(ξi)dξi
= 0 and for all ξi.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 40
Consensus protocol
• A protocol solving the consensus problem that is stationary and
distributed is such that the system state trajectory enjoys the
property that χ is time-invariant.
• We find a family of protocols guaranteeing this property.
• We prove that such protocols are consensus protocols.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 41
Time Invariancy of χ(x(t))
• Basic result: stationary means
limt−→∞
∥xi − χ(x(0))∥ = 0 =⇒ χ(x) time-invariant
• (Protocol design rule) For any χ(.), the protocol
ui(xi, x(i)) =
1dg(xi)dxi
∑j∈Ni
ϕ(xj , xi), for all i ∈ Γ
lets the value χ(x(t)) be time-invariant if ϕ(xj , xi) = −ϕ(xi, xj).
SCUOLA DOTTORATO- PALERMO 2011
DIETT 42
• Consider the linear function ϕ(xj , xi) = α(xj − xi) and the
different means.
• The arithmetic mean is time-invariant under protocol
u(xi, x(i)) = α
∑j∈Ni
(xj − xi)
• the geometric mean under protocol
u(xi, x(i)) = αxi
∑j∈Ni
(xj − xi)
• the harmonic mean under protocol
u(xi, x(i)) = −αx2
i
∑j∈Ni
(xj − xi)
• the mean of order p under protocol
u(xi, x(i)) = α
x1−pi
p
∑j∈Ni
(xj − xi).
SCUOLA DOTTORATO- PALERMO 2011
DIETT 43
Sufficient conditions for convergence• g(.) strictly increasing
• ϕ(xj , xi) = αϕ(ϑ(xj)− ϑ(xi)),
– α > 0
– ϕ : IR → IR continuous, locally Lipschitz, odd and strictly
increasing,
– ϑ : IR → IR differentiable with dϑ(xi)dxi
locally Lipschitz and
strictly positive.
• For any χ(.), the protocol
ui(xi, x(i)) = α
1dgdxi
∑j∈Ni
ϕ(ϑ(xj)− ϑ(xi)), for all i ∈ Γ.
lets the agents reach consensus on χ(x(0)1 (Liapunov Approach).
SCUOLA DOTTORATO- PALERMO 2011
DIETT 44
Lyapunov Approach
• define a new variables ηi = g(xi)− g(χ(x(0))).
• η = 0 corresponds to x = χ(x(0))1
• candidate Lyapunov function
V (η) =1
2
∑i∈Γ
η2i
• η = 0 is asymptotically stable equilibrium (in the quotient space
IRn/span{1}).
SCUOLA DOTTORATO- PALERMO 2011
DIETT 45
Switching Topology
• E finite set of all possible edgesets connecting the agents in Γ.
• The elements E ∈ E define family of dynamical systems
xi = uiE(xi, x(i)) ∀i ∈ Γ,
• We define a switching function and we assume dwell time to
avoid Zeno behaviors [Jadbabaie et al.2003]
• To extend the results and prove convergence we must prove that
V (η) = 12
∑i∈Γ η
2i is a common Lyapunov function
i) positive definite
ii) continuously differentiable and
V (η) ≤ −W (η), ∀η ∈ IRn, E ∈ E with W (η) independent
from E.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 46
Common Lyapunov Function
• time varying edge costs
cij = α(g(xj)− g(xi))ϕ(ϑ(xj)− ϑ(xi)), ∀(i, j) ∈ E.
• subset Q ⊆ E of edgsets defining (connected) trees (Γ, Q).
• For each Q ∈ Q define continuous and positive definite function
WQ(η) =∑
(i,j)∈Q
cij .
• find minimum spanning tree
W (η) = minQ∈Q
{WQ(η)}
• W (η), continuous, positive definite and
V (η) ≤ −W (η), ∀η ∈ IRn, E ∈ E
SCUOLA DOTTORATO- PALERMO 2011
DIETT 47
Mechanism design
• The distributed/individual optimality of the considered protocols
allows interpreting our consensus problem as a non cooperative
differential game
• A supervisor entails the agents to reach a consensus by imposing
individual objectives.
• The main topic of the mechanism design is the definition of game
rules or incentive schemes that induce self-interested players to
cooperate and reach Pareto optimal solution.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 48
Mechanism Design
• Recast as noncooperative differential game
Ji(xi, x(i), ui) = lim
T−→∞
∫ T
0
(F (xi, x
(i)) + ρu2i
)dt
– ρ > 0 weights control energy
– F : IR× IRn → IR nonnegative penalty function that measures
the deviation of xi from neighbors’ states.
• Mechanism design problem: Design F (.) such that the
agents reach consensus.
• Difficult problem: reformulate as Receding Horizon problem.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 49
No mechanism Design
Consider the case when it is not employed.
• The distributed protocols are generally planned at high level by a
supervisor.
• The supervisor communicates the planned control policies to the
agents that are in charge to implement them at a low level.
• The agents are only said what to do by the supervisor and do not
exhibit any decision capability.
• A drawback in systems with large dimensions and complex
structures (too onerous)
• if agents are not motivated to behave in the desired way, then a
costly monitoring system may be necessary.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 50
Mechanism Design
• The advantage of the mechanism design is that intelligence as
well as implementation capabilities are distributed.
• Reviewing the asymptotically consensus reaching as a team
objective
• The supervisor decomposes this team objective in n individual
objectives.
• the agents are said what to aim at instead of “what to do”, and
are free to find the best solution to their subtasks.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 51
Assumptions
• We need to make some assumptions on the information available
to each agent and its computational capabilities.
• At each time instant, the generic agent can observe the state of
its neighbor agents
• but cannot predict their future behavior
• Hence uses a naive expectation of the state of the neighbor agents
in evaluating its individual objective function.
• naive expectation means that the agents choose their optimal
control policy assuming the neighbors’ state variables as
constants.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 52
Naive Expectation
• The naive expectation is a major assumption, but it is reasonable
when agents:
– cannot keep record of the past behavior of their neighbors
– are so numerous that they can infer little or none on the
group decision value from the punctual observation of their
neighbors’ state
– are organized in a network with a switching topology whose
switching times and pattern are unpredictable.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 53
Why Naive Expectation?
• Reduces the RH problem to 1-dimension
Ji(xi(tk), x(i)(tk), ui(τ, tk)) = lim
T−→∞
∫ T
tk
(F(xi(τ, tk), x
(i)(τ, tk)) + ρu2i (τ, tk)
)dτ
⇓
Ji(xi(tk), ui(τ, tk)) = limT−→∞
∫ T
tk
(F(xi(τ, tk)) + ρu2
i (τ, tk))dτ.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 54
Optimality results
• If the supervisor imposes convex penalty functions, rational
agents have a unique optimal protocol.
• We prove this result through the Pontryagin Minimum Principle.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 55
Main Result:
• If update time interval δ −→ 0 then
i) the penalty function
F(xi(τ, tk)) −→ F (xi, x(i)) = ρ
1dgdxi
∑j∈Ni
(ϑ(xj)− ϑ(xi))
2
ii) the applied receding horizon control law
u⋆iRH
(τ) −→ ui(xi, x(i)) =
1dgdxi
∑j∈Ni
(ϑ(xj)− ϑ(xi)).
• Fact: for any χ and x(0), with the penalty F (xi, x(i)) as above,
selfish agents reach consensus on χ(x(0)).
SCUOLA DOTTORATO- PALERMO 2011
DIETT 56
Vertical alignement Maneuver for UAVs
• A team of 4 UAVs in longitudinal flight and initially at different
heights.
• Each UAV controls the vertical rate without knowing the relative
position of all UAVs but only of neighbors according to the
communication network topology:
v1 v3 v4
v2
SCUOLA DOTTORATO- PALERMO 2011
DIETT 57
Arithmetic Mean
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 55
10
15
20
time
heig
ht
F (xi, x(i)) =
(∑j∈Ni
(xj − xi))2
; u(xi, x(i)) =
∑j∈Ni
(xj − xi).
SCUOLA DOTTORATO- PALERMO 2011
DIETT 58
Geometric Mean
0 0.1 0.2 0.3 0.4 0.5 0.6 0.75
10
15
20
time
heig
ht
F (xi, x(i)) =
(xi
∑j∈Ni
(xj − xi))2
; u(xi, x(i)) = xi
∑j∈Ni
(xj − xi).
SCUOLA DOTTORATO- PALERMO 2011
DIETT 59
Harmonic Mean
0 0.02 0.04 0.06 0.08 0.1 0.125
10
15
20
time
heig
ht
F (xi, x(i)) =
(x2i
∑j∈Ni
(xj − xi))2
; u(xi, x(i)) = −x2
i
∑j∈Ni
(xj −xi)
SCUOLA DOTTORATO- PALERMO 2011
DIETT 60
Mean of order 2
0 20 40 60 80 100 120 140 160 180 2005
10
15
20
time
heig
ht
F (xi, x(i)) =
(1
2xi
∑j∈Ni
(xj − xi))2
; u(xi, x(i)) = 1
2xi
∑j∈Ni
(xj −xi)
SCUOLA DOTTORATO- PALERMO 2011
DIETT 61
Concluding Remarks
• AIM:To make the agents’ states reach consensus on a group
decision value with general structure
• RESULT: the agents can reach consensus with a distributed and
stationary linear or non-linear protocol, provided that the
networks defined by the communication links between agents is
connected though possibly time variant.
• SOLUTION: Consensus is the result of a mechanism design
(game theoretic approach).
• A supervisor imposes individual objectives.
• The agents reach asymptotically consensus as a side effect of the
optimization of their own individual objectives on a local basis.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 62
Outline
• Introduction
– Applications
– Why consensus problems
– Graph Theory
– The Consensus Problem
– References
• Distributed consensus protocols for coordinating buyers
• Mechanism Design for Optimal Consensus Problems
� Lazy consensus for networks with unknown but bounded
disturbances
• References
SCUOLA DOTTORATO- PALERMO 2011
DIETT 63
Illustrative example 1/2
1 2 3
x1 x2 x3
x1
x2
x3
= −
1 −1 0
−1 2 −1
0 −1 1
︸ ︷︷ ︸
L
x1
x2
x3
,
x1 = x2 − x1
x2 = (x1 − x2) + (x3 − x2)
x3 = x2 − x3
• Find equilibria by imposing x = 0 and obtain
x1 = x2 = x3 =x1(0) + x2(0) + x3(0)
3
SCUOLA DOTTORATO- PALERMO 2011
DIETT 64
Illustrative example 2/2
0 10 20 30 40 50 600
5
10
15
20
25
t
x i
SCUOLA DOTTORATO- PALERMO 2011
DIETT 65
UBB disturbances
1 2 3
x1+d21 x2 x3+d23
• neighbors’ measures y(i)
• closed-loop dynamics under protocol ui(xi, y(i))
x1 = (x2 + d12)− x1
x2 = [(x1 + d21)︸ ︷︷ ︸y21
−x2] + [(x3 + d23)︸ ︷︷ ︸y23
−x2]
x3 = (x2 + d32)− x3
• Find equilibria by imposing x = 0 and obtain
x1 = x2 + d12, 0 = d12 + d21 + d23 + d32, x3 = x2 + d32
• No equilibria unless d12 + d21 + d23 + d32 = 0.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 66
UBB disturbances
0 10 20 30 40 50 600
5
10
15
20
25
t
x i
• take dij = 1 for all i, j then x = −Lx+ (1, 2, 1)T has a drift
SCUOLA DOTTORATO- PALERMO 2011
DIETT 67
UBB disturbances
0 10 20 30 40 50 600
5
10
15
20
25
t
x i
• take d12 = d23 = 1, d32 = d21 = −1, then x = −Lx+ (1, 0,−1)T
SCUOLA DOTTORATO- PALERMO 2011
DIETT 68
ϵ-consensus
• bounded tube of radius ϵ,
T = {x ∈ Rn : |xi − xj | ≤ 2ϵ, ∀ i, j ∈ Γ} .
• given equilibrium x∗ ∈ T
• asymptotically ϵ-consensus
x(t) → x∗ for t → ∞
• for ϵ = 0 we have the standard consensus definition.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 69
Problem formulation
• find a rule to estimate yij ∈ [yij − ξ, yij + ξ] such that u(xi, y(i))
is ϵ-consensus protocol.
• study dependence of T on
– the topology E,
– disturbances D
• hypercube D = {d : −ξ ≤ dij ≤ ξ, ∀(i, j) ∈ Γ2}
SCUOLA DOTTORATO- PALERMO 2011
DIETT 70
Lazy rule
j=1 j=2 j=3
10
4
6
8
2
9
6
3
4
2
5
8
3
5
7
cg
ỹ2j
y2j
SCUOLA DOTTORATO- PALERMO 2011
DIETT 71
Lazy rule
j=1 j=2 j=3
10
4
6
8
2
9
6
3
4
2
5
8
3
5
7
cg
pick
4 4 4
ỹ2j
SCUOLA DOTTORATO- PALERMO 2011
DIETT 72
Lazy rule
• pick the solution of
y(i) = arg minyij∈[yij−ξ,yij+ξ], j∈Ni
|∑j∈Ni
(yij − xi)|.
• it maintains the sign: for each i ∈ Γ, for each t ≥ 0 either
sign(ui(xi, y(i))) = sign(
∑j∈Ni
(xj − xi)),
or
sign(ui(xi, y(i))) = 0.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 73
Equilibria
• all equilibria are such that
ui(xi, y(i)) = sign(ui(xi, y
(i))) = 0, for all i
• assume d(t) unknown but constant
• any solution x is an equilibrium if and only if it belongs to
P (d,E) =
{x : −
∑j∈Ni
dij
|Ni|− ξ ≤
∑j∈Ni
xj
|Ni|− xi ≤ −
∑j∈Ni
dij
|Ni|+ ξ, ∀i ∈ Γ
};
SCUOLA DOTTORATO- PALERMO 2011
DIETT 74
Polyhedron P (0, E)
x1 x2
x3
{π1}
SCUOLA DOTTORATO- PALERMO 2011
DIETT 75
Equilibria on ∂P (0, E)
j=1 j=2 j=3
10
4
6
8
2
cg
ỹij
-ξ
SCUOLA DOTTORATO- PALERMO 2011
DIETT 76
Equilibria on ∂P (0, E)
j=1 j=2 j=3
10
4
6
8
2
cg
ỹij ξ
SCUOLA DOTTORATO- PALERMO 2011
DIETT 77
Generic polyhedron P (D,E)
• For a generic {d(t) ∈ D, t ≥ 0}, define
P (D,E) :=∪d∈D
P (d,E) =
{−2ξ ≤
∑j∈Ni
xj
|Ni|− xi ≤ 2ξ, ∀i ∈ Γ
}.
• ∀x ∈ P (D,E), there exists a {d(t) ∈ D, t ≥ 0} s.t. x is equ.;
• Given a {d(t) ∈ D, t ≥ 0} all its equilibria are in P (D,E).
SCUOLA DOTTORATO- PALERMO 2011
DIETT 78
Stability
• Thm.: all system trajectories converges to equilibria in P (D,E)
• Proof: consider Lyapunov function V (x) = 12
∑(i,j)∈E(xj − xi)
2.
– V (x) = 0 if and only if x ∈ {π1};– V (x) > 0 for all x ∈ {π1}.– V (x) < 0 for all x ∈ P (D,E). Actually
V (x) =∑
(i,j)∈E
(xj − xi)(uj − ui) = −∑i∈Γ
ui
∑j∈Ni
(xj − xi) =
= −∑i∈Γ
sign(ui) sign
∑j∈Ni
(xj − xi)
|ui|
∣∣∣∣∣∣∑j∈Ni
(xj − xi)
∣∣∣∣∣∣SCUOLA DOTTORATO- PALERMO 2011
DIETT 79
Stability
• assume d bang-bang
• Thm.: all system trajectories converges to {π1}
• extensions to switching topology
SCUOLA DOTTORATO- PALERMO 2011
DIETT 80
Conclusions
• Despite the literature on consensus is now becoming extensive,
only few approaches have considered a disturbance affecting the
measurements.
• we have assumed an UBB noise in the neighbors’ state feedback
as it requires the least amount of a-priori knowledge on the
disturbance.
• Because of the presence of the disturbances convergence to
equilibria with all equal components is not possible.
• Main contribution: the introduction and solution of the
ϵ-consensus problem, where the states converge in a tube of ray ϵ
asymptotically or in finite time.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 81
Outline
� Introduction
– Applications
– Why consensus problems
– Graph Theory
– The Consensus Problem
– References
• Distributed consensus protocols for coordinating buyers
• Mechanism Design for Optimal Consensus Problems
• Lazy consensus for networks with unknown but bounded
disturbances
� References
SCUOLA DOTTORATO- PALERMO 2011
DIETT 82
References
• D. Bauso, L. Giarre, R. Pesenti, Consensus for networks with
unknown but bounded disturbances. SIAM JOURNAL ON
CONTROL AND OPTIMIZATION, 48(3); 1756-1770,2009,
• D. Bauso, L. Giarre, R. Pesenti. Distributed Consensus in
Noncooperative Inventory Games. EUROPEAN JOURNAL OF
OPERATIONAL RESEARCH; 192(3), 866-878, 2009;
• D. Bauso, L. Giarre, R. Pesenti. Consensus in Noncooperative
Dynamic Games: a Multi-Retailer Inventory Application. IEEE
TRANSACTIONS ON AUTOMATIC CONTROL, 53(4),
998-1003, 2008.
• D. Bauso, L. Giarre, R. Pesenti. Nonlinear protocols for optimal
distributed consensus in networks of dynamic agents. Systems
and Control Letters, 915-928, 55(11),2006.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 83
Conference papers
• D. Bauso, L. Giarre, R. Pesenti. Quantized Dissensus in
switching networks with nodes death and duplication, NecSys,
Venice, 2009.
• D. Bauso, L. Giarre, R. Pesenti. “Dissensus, Death and
Division”, IEEE ACC, 2307-2312, 2009.
• D. Bauso, L. Giarre, R. Pesenti. “Challenging aspects in
Consensus protocols for Networks”, Proceedings of the 3rd IEEE
ISCCSP, 2008.
• D. Bauso, L. Giarre, R. Pesenti. “Noncooperative Dynamic
games for Inventory Applications: a consensus approach ”, IEEE
CDC, p. 4819-4824,2008.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 84
Conference papers
• D. Bauso, L. Giarre, R. Pesenti. “Lazy consensus for network
with unknown but bounded noise”, IEEE CDC, new orleans,
december 12-14,2007, p. 2284-2290, 2007.
• D. Bauso, L. Giarre, R. Pesenti. “Distributed consensus for
switched networks with unknown but bounded noise”, 3rd
International workshop on : Networked Control Systems :
Tolerant to faults, 2007.
• D. Bauso, L. Giarre, R. Pesenti, Mechanism Design for Optimal
Consensus Problems, CDC- 2006.
• D. Bauso, L. Giarre, R. Pesenti, Distributed consensus in
Networks of Dynamic Agents, CDC-ECC, 2005.
SCUOLA DOTTORATO- PALERMO 2011
DIETT 85
• D. Bauso, L. Giarre, R. Pesenti, Existence and Optimality of
Nash Equilibria in Inventory Games, 16th IFAC World Congress,
2005.
• D. Bauso, L. Giarre, R. Pesenti, Multiple UAV Cooperative Path
Planning via Neuro-Dynamic Programming, CDC 2004.
• D. Bauso and L. Giarre and R. Pesenti, Neuro-Dynamic
Programming for Cooperative Inventory Control, ACC 2004.
• D. Bauso, L. Giarre, R. Pesenti, Distributed Consensus Protocols
for Coordinating Buyers, CDC 2003.
• D. Bauso, L. Giarre, R. Pesenti, Attitude Alignment for a Team
of UAVs under decentralized Information Structure, CCA 2003.
SCUOLA DOTTORATO- PALERMO 2011