Yajun Mei (Georgia Tech)

​

Title: Bandit multi-stream sequential change-point detection

 

Abstract -Bandit sequential change-point detection problem occurs in many real-world problems such smart manufacturing or biosurveillance when one monitors multi-dimensional data streams under the sampling control due to limited capacity in data acquisition, transmission or processing. In such a scenario, one needs decide how to smartly observe which local components or features of multi-dimensional streaming data at each and every time, and then uses the observed incomplete data to quickly raise an alarm once a change has occurred subject to the false alarm constraint. In this talk, we present two of our latest research by developing efficient change-point detection algorithms under the sampling control through bandit sampling policies: one is Robbins' win-stay, lose-switch policy that leads to asymptotically optimal algorithm when monitoring low-dimensional data, and the other is Thompson sampling policy that yields efficient scalable schemes for online monitoring high-dimensional data. Numerical simulations and case studies will be presented to demonstrate the usefulness of our proposed algorithms, and future potential research directions will also be discussed.

​

​

RETURN TO SEMINAR HOMEPAGE