|Stochastic Systems Group|
Alex Ihler - SSG, MIT
Many problems involving nonlinear, non-Gaussian relationships can be solved approximately using particle- (sample-) based representations. Examples of this include particle filtering for tracking or other inference problems defined on a Markov chain, or nonparametric belief propagation (NBP), a sample-based approximate inference algorithm on general graphical models. One particularly interesting domain is that of ad-hoc sensor networks. Often, information fusion in these networks must be performed in a distributed fashion on a limited power budget, linking the process of inference with communications constraints -- which sensors may communicate, and how much they may transmit.
In the first half of the talk, we describe the canonical problem of sensor localization, or automatically determining each sensor's location given local, relative distance information, as a graphical model which can be approximately solved by NBP. The resulting distributed algorithm resolves both an estimate of the sensor locations, as well as an estimate of the residual uncertainty. To analyze the communications cost of NBP-based localization, in the second half of the talk we discuss the encoding of the particle-based messages required by this algorithm, and describe the fundamental cost of lossless coding along with ways to efficiently trade off communications cost with error in the overall inference goal (lossy encoding).
Problems with this site should be emailed to email@example.com