Optimal Design of the Shiryaev-Roberts Chart: Give Your Shiryaev-Roberts a Headstart

TL;DR

This paper investigates how optimal headstarts can significantly improve the speed and effectiveness of Shiryaev-Roberts charts in detecting mean shifts in normal processes, especially for faint changes.

Contribution

It introduces an optimal headstart design for SR charts, demonstrating near-uniform fastest detection across various change magnitudes and ARL levels.

Findings

Optimal headstart improves detection speed for faint shifts.

Nearly uniform fastest detection achieved with the designed headstart.

Performance gains are significant across diverse scenarios.

Abstract

We offer a numerical study of the effect of headstarting on the performance of a Shiryaev-Roberts (SR) chart set up to control the mean of a normal process. The study is a natural extension of that previously carried out by Lucas and Crosier for the CUSUM scheme in their seminal 1982 paper published in Technometrics. The Fast Initial Response (FIR) feature exhibited by a headstarted CUSUM turns out to be also characteristic of an SR chart (re-)started off a nonzero initial score. However, our main result is the observation that a FIR SR with a carefully designed {\em optimal} headstart is not just faster to react to an initial out-of-control situation, it is nearly {\em the} fastest {\em uniformly}, i.e., assuming the process under surveillance is equally likely to go out of control effective any sample number. The performance improvement is the greater, the fainter the change. We explain the optimization strategy, and tabulate the optimal initial score, control limit, and the corresponding "worst possible" out-of-control Average Run Length (ARL), considering mean-shifts of diverse magnitudes and a wide range of levels of the in-control ARL.