Dynamic Thresholding for Distributed Multiple Hypotheses Testing

We consider a distributed multiple hypotheses testing problem in sensor networks using false discovery rate as the fidelity criterion. We propose an energy efficient, 1-bit per sensor algorithm that implements a dynamic thresholding strategy. The method takes the information that has been collected at each iteration of the distributed algorithm and uses it to estimate the number of significant observations. This estimate is then used to update the false discovery rate constraint. The learning approach leads to more aggressive thresholding strategies and leads to much larger detection results in comparison to previously developed distributed procedures. We then discuss extensions of the approach to parameter estimation problems with hidden variables.