Robust Quickest Change Detection with Sampling Control

TL;DR

This paper develops a robust, cost-efficient quickest change detection algorithm with sampling control, effective in minimizing detection delay while managing false alarms and observation costs, validated on real and simulated data.

Contribution

It introduces a computationally efficient, robust on-off observation control algorithm for quickest change detection under uncertain post-change distributions.

Findings

Algorithm reduces observation costs while maintaining detection performance.

Robustness to distributional uncertainty is achieved.

Validated on both simulated and real public health data.

Abstract

The problem of quickest detection of a change in the distribution of a sequence of random variables is studied. The objective is to detect the change with the minimum possible delay, subject to constraints on the rate of false alarms and the cost of observations used in the decision-making process. The post-change distribution of the data is known only within a distribution family. It is shown that if the post-change family has a distribution that is least favorable in a well-defined sense, then a computationally efficient algorithm can be designed that uses an on-off observation control strategy to save the cost of observations. In addition, the algorithm can detect the change robustly while avoiding unnecessary false alarms. It is shown that the algorithm is also asymptotically robust optimal as the rate of false alarms goes to zero for every fixed constraint on the cost of observations. The algorithm's effectiveness is validated on simulated data and real public health data.