Multi-agent Optimization
Asu Ozdaglar
February 2009
Examples:
(a) The weights aij (k) are nonnegative for all agents i, j.
(b) There exists a scalar (0, 1) such that for all agents i,
(c) The agent weight vectors ai (k) = [ai1 (k), . . . , aim (k)]T are stochastic, i.e.,
Pm i
j=1 aj (k) = 1 for all i and k.
where B0 = (m 1)B.
(b) The limit D(s) has identical rows and the rows are stochastic.
(c) For every j, the entries [Dk (s) D1 (s)]ji , i = 1, . . . , m, converge to the same
limit j (s) as k with a geometric rate:
k
j B0 B0
[Dk (s) D1 (s)]i j (s) 2 1 + 1
Using the property that each entry of the matrix Dk is positive, we show
k
B0
kzk k 1 kz0 k for all k.
We then show that the sequence {ck } converges to some c R and use the
contraction constant to establish the rate estimate
(b) The limit matrix (s) has identical columns and the columns are stochastic, i.e.,
(s) = (s)e0 ,
(c) For every i, [(k, s)]ji , j = 1, ..., m, converge to the same limit i (s) as k
with a geometric rate, i.e., for all i, j and all k s,
j 1 + B0 B0
ks
B0
[(k, s)]i i (s) 2 1
1 B0
where is the lower bound on weights, B is the intercommunication interval
bound, and B0 = (m 1)B.
The rate estimate in part (c) recently improved in [Nedic, Olshevsky, Ozdaglar,
Tsitsiklis 08]
MIT Laboratory for Information and Decision Systems 13
Convergence Analysis
Recall the evolution of the estimates (with i (s) = ):
m k1
m !
X X X
xi (k + 1) = [(k, s)]ij xj (s) [(k, r + 1)]ij dj (r) di (k).
j=1 r=s j=1
Proof method: Consider a stopped process, where after time k, agents stop
computing subgradients but keep exchanging their estimates: di (k) = 0 for all
k k and all i.
Using limk [(k, s)]ij = j (s) for all i, we see that the limit vector
limk xi (k) exists, independent of i, but depends on k:
1 X i
y(0) = x (0), C = 1 + 8mC1
m i
m 1 + B0
C1 = 1 + 1 , B0 = (m 1)B
1 (1 B0 ) B0 1 B0
Using the convergence rate estimate for the transition matrices (k, s), we show
i 1
Pk1 i
that y(k) are close to the agents averages x (k) = k h=0 x (h)
Implication:
Pm
limk i=1 kwi (k) xk2 kxi (k + 1) xk2 = 0, for all x X
Using the transition matrices and the decomposition of the estimate evolution,
m k
m !
i
X i j
X X
x (k + 1) = [(k, s)]j x (s) + [(k, r)]ij ej (r 1) + ei (k).
j=1 r=s+1 j=1
Proposition: Let Interior Point Assumption hold. Let the weight vectors ai (k) be
given by ai (k) = (1/m, . . . , 1/m)0 for all i and k. Then,
Xm k X m
1
kxi (k) xk2 1 2
kxi
(0) xk 2
for all k 0,
i=1
4R i=1
Pm
where x X is the limit of the sequence {xi (k)}, and R = 1
i=1 kxi
(0) xk,
i.e., the convergence rate is linear.
Convergence rate for time-varying weights and different local constraints open
MIT Laboratory for Information and Decision Systems 24
Summary
We presented a distributed subgradient method for multi-agent optimization
We proposed a constrained consensus policy for the case when agents have local
constraints on their estimates
Results:
We assume that the matrix A(k) is a random matrix drawn independently over
time from a probability space of stochastic matrices.
Results: