1: \begin{abstract}

2:   We consider $n$ agents located on the vertices of a connected graph.

3:   Each agent $v$ receives a signal $X_v(0)\sim N(\mu,1)$

4:   where $\mu$ is an unknown quantity. A natural iterative

5:   way of estimating $\mu$ is to perform the following

6:   procedure. At iteration $t+1$ let $X_v(t+1)$

7:   be the average of $X_v(t)$ and of $X_w(t)$ among all the

8:   neighbors $w$ of $v$. It is well known that this procedure

9:   converges to $X(\infty) = \frac{1}{2} |E|^{-1} \sum d_v X_v$ where

10:   $d_v$ is the degree of $v$.

11: 

12:   In this paper we consider a variant of simple iterative averaging,

13:   which models ``greedy'' behavior of the agents.

14:   At iteration $t$, each agent $v$ declares the value of its estimator

15:   $X_v(t)$ to all of its neighbors. Then, it updates $X_v(t+1)$

16:   by taking the maximum likelihood (or minimum variance) estimator of

17:   $\mu$, given $X_v(t)$ and

18:   $X_w(t)$ for all neighbors $w$ of $v$, and the structure of the graph.

19: 

20:   We give an explicit efficient procedure for calculating $X_v(t)$,

21:   study the convergence of the process as $t \to \infty$ and show that if

22:   the limit exists then $X_v(\infty) = X_w(\infty)$ for all $v$ and $w$.

23:  For graphs that are

24:   symmetric under actions of transitive groups, we show that

25:   the process is efficient.  Finally, we show that the greedy

26:   process is in some cases

27:   more efficient than simple averaging, while in other cases

28:   the converse is true, so that, in this model, ``greed'' of the

29:   individual agents may or may not have an adverse

30:   affect on the outcome.

31: 

32:   The model discussed here may be viewed as the Maximum-Likelihood

33:   version of models studied in Bayesian Economics.  The ML variant is

34:   more accessible and allows in particular to show the significance of

35:   symmetry in the efficiency of estimators using networks of agents.

36: 

37: 

38: \end{abstract}

39: