Regenerative likelihood ratio control schemes

Abstract

We discuss the problem of monitoring where models driving the data are undergoing abrupt changes in time, such as shifts or drifts. We introduce a unified methodology based on the use of likelihood ratio tests that enables one to obtain control schemes that provide both good (and, under some conditions, optimal) statistical performance and are relatively easy to implement. These schemes depend on just one design parameter and require a limited computational effort that is dynamically adjusted based on process conditions. An example pertaining to multivariate control of the normal mean is discussed in detail.