Skip to content. Skip to navigation
CIM Menus
 

McGill ECE Seminar

Distributed Estimation and Recursive Composite Hypothesis Testing: A Consensus + Innovations Approach


Anit Kumar Sahu
Electrical and Computer Engineering Carnegie Mellon University

July 19, 2017 at  11:00 AM
McConnell Engineering Room 603

Abstract:

The heterogeneity and enormity of static and streaming data nowadays calls for systematic distributed processing of the data while keeping up with the accuracy wth provable guarantees. For example, in any online digital advertising platform estimating the click through rate of an ad impression is paramount to the advertiser and also to the advertising platform. However, the impressions are shown across different advertisement sections in different apps which basically makes the data inherently distributed and the rate of an impression slotted at one in 30 ms makes the data volume enormous and of streaming nature which makes batch processing prohibitive.

In this talk, I will first present a distributed weighted non-linear least squares algorithm, namely CIWNLS, where the task is to estimate a possible high-dimensional parameter from data which is a time-series of noise corrupted non-linear function of the parameter distributed across agents in multi-agent network. Furthermore, I will establish the order-optimal convergence rate of the estimator and benchmark the performance with respect to the centralized estimator in terms of the asymptotic covariance. In the second part of the talk, I will focus on recursive distributed composite hypothesis testing, where the task is to estimate the possibly high-dimensional state of the environment through streaming data distributed across different agents in a multi-agent network, where the state of the estimator is indistinguishable locally at each agent. In particular, I will present two algorithms namely CIGLRT-L and CIGLRT-NL, of the consensus+innovations type, where instead of the classical batch processing based Generalized Likelihood Ratio Tests (GLRT), the data is processed in a sequential yet online manner. In spite of the online nature yielding sub-optimal test statistics, I will show that both the algorithms admit exponentially decaying probabilities of error.

Bio:

Anit Kumar Sahu is a PhD candidate in the Department of Electrical and Computer Engineering at Carnegie Mellon University, Pittsburgh. He received a B.Tech. in Electronics and Electrical Communication Engineering and M.Tech. in Telecommunication Systems Engineering from the Indian Institute of Technology, Kharagpur, India, in May 2013. Since Fall 2013, he has been working towards his Ph.D. in Electrical and Computer Engineering at Carnegie Mellon University, Pittsburgh, PA. His research interests include distributed inference in large-scale stochastic systems,statistical machine learning, and information theory.