1: \begin{abstract}

2: An energy efficient use of large scale sensor networks necessitates activating a subset of possible sensors for estimation at a fusion center. The problem is inherently combinatorial; to this end, a set of  iterative, randomized  algorithms are developed for sensor subset selection by exploiting the underlying statistics. Gibbs sampling-based methods are designed to optimize the estimation error and the mean number of activated sensors.  {\em The optimality of the proposed strategy is proven}, along with guarantees on their convergence speeds. Also, another new algorithm  exploiting stochastic approximation in conjunction with Gibbs sampling is derived for a constrained version of the sensor selection problem.  The methodology is extended to the scenario where the fusion center has access to only a parametric form of the joint statistics, but not the true underlying distribution.  Therein, expectation-maximization is effectively employed to  learn the distribution.  Strategies for iid time-varying data are also outlined.  Numerical results show that the proposed methods converge very fast to the respective {\em optimal} solutions, and therefore can be employed for optimal sensor subset selection  in practical sensor networks.

3: \end{abstract}

4: