Quickest change detection under post-change non-stationarity and uncertainty

The problem of quickest change detection in a sequence of independent observations is considered. Given sequential observations, the problem aims to detect the change after it occurs as quickly as possible, subject to false alarm constraints. The goal of this dissertation is to extend classical theories of quickest change detection for cases where the post-change observations are non-stationary, and/or where the post-change distribution is not fully known nor belongs to known parametric families. The first problem considered is the mean-change detection problem, where a change in the mean of an observation sequence above some threshold is of interest. The post-change observations are allowed to be non-stationary, and no knowledge of the post-change distribution is assumed other than that its mean is above the threshold. The problem is formulated as a robust change detection problem, and the Mean-Change Test (MCT) is derived, which is shown to be asymptotically close to the minimax robust solution. In the second problem, the post-change observations are assumed non-stationary with possible parametric uncertainty in their distribution, where this non-stationarity is characterized by the cumulative Kullback-Leibler divergence between the post- and the pre-change distributions. A universal asymptotic lower bound on the delay is derived. For the case where the post-change distributions have parametric uncertainty, a window-limited (WL) generalized likelihood-ratio (GLR) CuSum test is developed which is shown to be asymptotically optimal. The use of the WL-GLR-CuSum test in monitoring pandemics is also demonstrated. The next two problems focus on observation models where there is a lack of concrete knowledge of the post-change distribution. In the third problem, it is assumed that the only information about the post-change distribution is through a (small) set of training data. The problem is formulated as a data-driven robust change detection problem, where the post-change uncertainty set is constructed using the Wasserstein distance from the empirical distribution. The distributionally robust (DR) CuSum test is constructed and is shown to be asymptotically minimax robust. The size of the uncertainty set is theoretically characterized using Wasserstein concentration bounds. In the last problem, it is assumed that the post-change distribution is completely unknown. Two tests, the window-limited non-parametric generalized likelihood ratio (NGLR) CuSum test and the non-parametric window-limited adaptive (NWLA) CuSum test, are developed with generic density estimators. Both tests do not require any pre-collected training samples. The tests are shown to achieve first-order asymptotic optimality under certain convergence conditions on the density estimator.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Liang, Yuchen

Contributors dc:contributor

• Veeravalli, Venugopal V.
• Fellouris, Georgios
• Bose, Subhonmesh
• Moustakides, George V.

• Quickest Change Detection (qcd)
• Non-stationary Observations
• Non-parametric Methods
• Density Estimation
• Minimax Robust Detection
• Wasserstein Distance

Statement dc:rights

• Copyright 2023 Yuchen Liang

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/121436

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F121436