|Sequential anomaly detection with observation control under a generalized error metric
|Aristomenis Tsopelakos, Georgios Fellouris, University of Illinois at Urbana Champaign, United States
|E.2: Detection and Applications
|Detection and Estimation
|Click here to download the manuscript
|Click here to watch in the Virtual Symposium
|The problem of anomaly detection is considered when multiple processes are observed sequentially, but it is possible to sample only a subset of them at a time. The goal is to stop sampling as soon as possible and identify the anomalous processes. In large scale systems, we may tolerate less than k errors in order to decrease the expected time we make our decision. Therefore, we use a generalized error metric and we control the probability of making at least k errors of any kind. No prior information is assumed for the number of anomalies. We obtain the optimal asymptotic performance as the probability of making at least k misclassification errors, vanishes to zero, and we characterize the sampling rules that achieve asymptotic optimality. Moreover, we present three sampling rules, which differ in terms of the computational complexity and the actual performance they imply. A simulation illustrates the performance of the sampling rules in finite regime for all possible values of k.