0% found this document useful (0 votes)
16 views19 pages

Consenses in Multiagent

Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
16 views19 pages

Consenses in Multiagent

Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 19

Consensus Problem in Multi-Agent Systems

' $

Consensus Problem in Multi-Agent


Systems

Asu Ozdaglar

Joint work with Angelia Nedić UIUC, USA

Joint EUROPT-OMS Conference


July 2007

Department of Electrical Engineering & Computer Science


Massachusetts Institute of Technology, USA

& %
Consensus Problem in Multi-Agent Systems
' $

Motivation
• Increasing interest in distributed control and coordination of
networks consisting of multiple autonomous (potentially mobile)
agents

• Motivated by many emerging networking applications, such as ad


hoc wireless communication networks and sensor networks,
characterized by:
– Lack of centralized control and access to information
– Time-varying connectivity

• Control algorithms deployed in such networks should be:


– Completely distributed relying on local information
– Robust against changes in the network topology
– Easily implementable

& %
Consensus Problem in Multi-Agent Systems
' $
Consensus Problem
• Canonical problem that appears in the coordination of multi-agent
systems is the consensus problem
• Goal: Given initial values (scalar or vector) of agents, establish
conditions under which through local interactions and computations,
agents asymptotically agree upon a common value, i.e.,
reach a consensus
• Examples:
– Control of moving vehicles (UAVs): alignment of the
heading angles
– Information processing in sensor networks: computing
averages of initial local observations (i.e., consensus on a
particular value)
– Design of distributed optimization algorithms: need a
mechanism to align estimates of decision variables maintained by
different agents/processors
& %
Consensus Problem in Multi-Agent Systems
' $

Our Work
• We study consensus problem for multi-agent systems

• Main Results:
– General distributed asynchronous computational model for
reaching consensus
– Convergence analysis and convergence rate estimates for
time-varying topologies under general connectivity assumptions

• This talk has two parts:


– Analysis with no communication delay (i.e., there is no delay in
locally delivering information from one agent to another)
– Analysis with communication delay

• Part of the work not discussed here:


– Simultaneous optimization and consensus:
∗ See talk by Nedić in EURO XXII on Monday
& %
Consensus Problem in Multi-Agent Systems
' $
Related Literature
• Parallel and Distributed Algorithms:
– General computational model for dist asynchronous optimization
∗ Tsitsiklis 84, Bertsekas and Tsitsiklis 95
• Consensus and Cooperative Control:
– Analysis of group behavior (flocking) in dynamical-biological
systems
∗ Vicsek 95, Reynolds 87, Toner and Tu 98
– Mathematical models of consensus and averaging
∗ Jadbabaie et al. 03, Olfati-Saber Murray 04, Boyd et al. 05

• Previous literature:
– Focus on convergence to consensus
– No explicit convergence rate estimates (except for specific cases;
Olshevsky and Tsitsiklis 06, Cao et al. 06)
– Limited focus on communication delay case
& %
Consensus Problem in Multi-Agent Systems
' $

Model
• A network with m agents with node set V = {1, . . . , m}

• Agents update and (potentially) send their information at discrete


times t0 , t1 , t2 , . . .

• We use xi (k) ∈ Rn to denote information state of agent i at time tk

Agent Update Rule:


– Agent i updates his information state by
m
X
i
x (k + 1) = aij (k)xj (k),
j=1

where ai (k) = (ai1 (k), . . . , aim (k))0 is a vector of weights


– The vector ai (k) represents agent i’s neighbor relations at slot k
– Dynamics governed by a switched linear system

& %
Consensus Problem in Multi-Agent Systems
' $

Weights
Assumption (Weights Rule) For all k, we have

(a) There exists a scalar η ∈ (0, 1) s.t. for all i ∈ {1, . . . , m},
(i) aii (k) ≥ η
(ii) aij (k) ≥ η for all j communicating directly with i in (tk , tk+1 ).
(iii) aij (k) = 0 for all j otherwise.
Pm
(b) The vectors ai (k) are stochastic, i.e., j=1 aij (k) = 1 for all i.

Example: Equal neighbor weights aij (k) = 1


ni (k)+1
, where ni (k) is the
number of agents communicating with i (his neighbors) at slot k

& %
Consensus Problem in Multi-Agent Systems
' $

Information Exchange
At slot k, information exchange may be represented by a directed graph
(V, Ek ) where
Ek = {(j, i) | aij (k) > 0}

Assumption (Connectivity) The graph (V, E∞ ) is connected, where

E∞ = {(j, i) | (j, i) ∈ Ek for infinitely many indices k}.

• Information state of agent i influences information state of any other


agent infinitely often

Assumption (Bounded Intercomm Interval) There is some B ≥ 1 s.t.

(j, i) ∈ Ek ∪ Ek+1 ∪ · · · ∪ Ek+B−1 for all (j, i) ∈ E∞ and k ≥ 0.

• Agent j send his information to neighboring agent i at least once


every B consecutive time slots.

& %
Consensus Problem in Multi-Agent Systems
' $

Evolution of Information States


Notation: For a matrix A, we write

[A]ji : (i, j)th entry, [A]i : ith row, [A]j : j th column

• Let A(s) denote the matrix whose ith column is the vector ai (k)
– By Weights Rule(b), A0 (s) is a stochastic matrix

• By the linearity of the dynamics, the iterates satisfy


m
X
xi (k + 1) = [A(s)A(s + 1) · · · A(k − 1)ai (k)]j xj (s)
j=1

• We introduce the transition matrices

Φ(k, s) = A(s)A(s + 1) · · · A(k − 1)A(k) for all k ≥ s


Pm
• Then: xi (k + 1) = j=1 [Φ(k, s)] i j
j x (s)

& %
Consensus Problem in Multi-Agent Systems
' $

Properties of Transition Matrices


Lemma: Let Weights Rule (a), Connectivity, and Bounded
Intercommunication Interval assumptions hold. We then have

[Φ(s + (m − 1)B − 1, s)]ij ≥ η (m−1)B for all s, i, and j,

where η is the lower bound on weights and B is the intercommunication


interval bound.

• We introduce the matrices Dk (s) as follows: for a fixed s ≥ 0,

Dk (s) = Φ0 (s + kB0 − 1, s + (k − 1)B0 ) for k = 1, 2, . . . ,

where B0 = (m − 1)B.

• By the previous lemma, all entries of Dk (s) are positive.

& %
Consensus Problem in Multi-Agent Systems
' $

Convergence of Transition Matrices


Lemma: Let Weights Rule, Connectivity, and Bounded Intercommuni-
cation Interval assumptions hold. For each s ≥ 0, we have:

(a) The limit D̄(s) = limk→∞ Dk (s) · · · D1 (s) exists.

(b) The limit D̄(s) is a stochastic matrix with identical rows.

(c) The convergence of Dk (s) · · · D1 (s) to D̄(s) is geometric: ∀ x ∈ Rm ,


° ° ³ ´³ ´k
°(Dk (s) · · · D1 (s)) x − D̄(s)x° ≤ 2 1 + η 0 −B B
1 − η 0 kxk∞

In particular, for every j, the entries [Dk (s) · · · D1 (s)]ji , i = 1, . . . , m,


converge to the same limit φj (s) as k → ∞ with a geometric rate:
¯ ¯ ³ ´³ ´k
¯ j ¯ −B0 B0
¯[Dk (s) · · · D1 (s)]i − φj (s)¯ ≤ 2 1 + η 1−η
where η is the lower bound on weights, B is the intercommunication
interval bound, and B0 = (m − 1)B.

& %
Consensus Problem in Multi-Agent Systems
' $

Proof Idea
• We show that the sequence {(Dk · · · D1 )x} converges for every
x ∈ Rm

• Consider the sequence {xk } with xk = Dk · · · D1 x and write xk as

xk = zk + ck e, where ck = min [xk ]i


1≤i≤m

• Using the property that each entry of the matrix Dk is positive, we


show ³ ´k
B0
kzk k∞ ≤ 1 − η kz0 k∞ for all k.
Hence zk → 0 with a geometric rate.

• We then show that the sequence {ck } converges to some c̄ ∈ R and


use the contraction constant to establish the rate estimate

• The final relation follows by picking x = ej , the j th unit vector

& %
Consensus Problem in Multi-Agent Systems
' $

Convergence of Transition Matrices


Proposition: Let Weights Rule, Connectivity, and Bounded Inter-
communication Interval assumptions hold.

(a) The limit Φ̄(s) = limk→∞ Φ(k, s) exists for each s.

(b) The limit matrix Φ̄(s) has identical columns and the columns are
stochastic, i.e.,
Φ̄(s) = φ(s)e0 ,
where φ(s) ∈ Rm is a stochastic vector for each s.

(c) For every i, [Φ(k, s)]ji , j = 1, ..., m, converge to the same limit φi (s)
as k → ∞ with a geometric rate, i.e., for all i, j and all k ≥ s,
¯ ¯
¯ j ¯ 1 + η −B0 ³ B0
´ k−s
B0
¯[Φ(k, s)]i − φi (s)¯ ≤ 2 1 − η
1 − η B0
where η is the lower bound on weights, B is the intercommunication
interval bound, and B0 = (m − 1)B.

& %
Consensus Problem in Multi-Agent Systems
' $

Model with Delays


• Assume now that there is delay in delivering information of j to i
– Models communication delay over wireless links

• In the presence of delay, agent i updates his information state by


m
X
xi (k + 1) = aij (k)xj (k − tij (k)),
j=1

where tij (k) is the delay in passing information from j to i

Assumption (Bounded Delays)

(a) tii (k) = 0 for all agents i and all k ≥ 0.

(b) tij (k) = 0 for all agents i and j such that aij (k) = 0.

(c) There is an integer B1 such that 0 ≤ tij (k) ≤ B1 − 1 for all agents
i, j, and all k.
& %
Consensus Problem in Multi-Agent Systems
' $
Enlarged Linear System
• Under Bounded Delays assumption, the system with delays can be
reduced to a system without delays by state augmentation
• For each agent i, we associate a new agent for every possible delay
value
– Sufficient to add m(B1 − 1) new agents handling delays
• We refer to the original agents as computing agents (indexed by
1, . . . , m) and the new agents as non-computing agents (indexed
by m + 1, . . . , (B1 − 1)m)
• An example with 3 agents and delay bound B1 = 3

& %
Consensus Problem in Multi-Agent Systems
' $

Information State Evolution


• Let x̃i (k) denote information state of agent i of the enlarged system

• The evolution of the states x̃i (k) for all agents i in the enlarged
system:
mB1
i
X
x̃ (k + 1) = ãih (k)x̃h (k),
h=1

where weights ãih (k) for computing agents i ∈ {1, . . . , m} are



 ai (k) if h = j + tm, t = ti (k)
j j
ãih (k) = for all k ≥ 0,
 0 otherwise

while weights ãih (k) for noncomputing agents i ∈ {m + 1, . . . ,


mB1 } are

 1 for h = i − m
i
ãh (k) = for all k ≥ 0.
 0 otherwise

& %
Consensus Problem in Multi-Agent Systems
' $

Convergence Analysis
• Let Ã(k) denote the matrix whose columns are given by ãi (k).
– Under Weights Rule, Ã0 (k) is a stochastic matrix for all k.

• Similar to the previous analysis, we define the transition matrices

Φ̃(k, s) = Ã(s)Ã(s + 1) · · · Ã(k − 1)Ã(k) for all k ≥ s.

Lemma:
(a) For any computing nodes i, j ∈ {1, . . . , m},

[Φ̃(k, s)]ij ≥ η k−s+1 for all k ≥ s + (m − 1)(B + B1 ).

(b) For any computing node j ∈ {1, . . . , m}, we have

[Φ̃(s + (m − 1)B + mB1 − 1, s)]ij ≥ η (m−1)B+mB1 for all i.

& %
Consensus Problem in Multi-Agent Systems
' $

Convergence Rate
Proposition:

(a) The limit Φ̃(s) = limk→∞ Φ̃(k, s) exists for each s.

(b) The limit matrix Φ̃(s) has identical columns and the columns are
stochastic, i.e.,
Φ̃(s) = φ̃(s)e0 ,
where φ̃(s) ∈ RmB1 is a stochastic vector for each s.

(c) For every i ∈ {1, . . . , mB1 }, the entries [Φ̃(k, s)]ji , j = 1, ..., mB1 ,
converge to the same limit φ̃i (s) as k → ∞ with a geometric rate,
i.e.,
¯ ¯
¯ j ¯ 1 + η −B2 ³ B2
´ k−s
B2
¯[Φ̃(k, s)]i − φ̃i (s)¯ ≤ 2 B
1 − η ,
1−η 2

where B is the intercommunication interval bound, B1 is the delay


bound, and B2 = (m − 1)B + mB1 .

& %
Consensus Problem in Multi-Agent Systems
' $

Conclusions
• We presented a general distributed computational model for the
consensus problem

• We provided convergence analysis and convergence rate estimates


with and without communication delay

• Our estimates highlight the dependence of convergence rate on key


system parameters

• Ongoing work:
– Distributed asynchronous subgradient methods for constrained
multi-agent optimization
– Effects of quantization of information states on consensus

& %

You might also like