Distributed and Sequential Detection of Many Interacting Change Points
Long Nguyen
Statistics, University of Michigan, Ann Arbor, MI, United States

A common application in sensor networks involve simultaneously detecting many change points on the basis of sequences of sensor data. The key departure from the classical literature of sequential analysis is that we have many change points that are statistically dependent. Moreover, there are constraints in the amount of information (statistics) a sensor may receive from its neighbors, potentially affecting the overall quality of the detection procedures. We propose a sequential detection procedure for detecting simultaneously interacting change points. This procedure can be implemented in a distributed fashion by having to pass statistics between only neighboring sensors in the network. We also provide an asymptotic theory for understanding the tradeoff between false alarm probability and the detection delay in the network setting.

Keywords: Change point detection; Sequential analysis; Graphical models; Message-passing algorithms

Biography: Long Nguyen is Assistant Professor in Department of Statistics, University of Michigan. His research interests lie in statistical and computational modeling, machine learning and optimization methods for structured data and distributed systems.