Multivariate ranking and selection without reduction to a univariate problem

• Authors: Edward J. Dudewicz;Vidya S. Taneja
• Venue: WSC '78 Proceedings of the 10th conference on Winter simulation - Volume 1
• Year: 1978

“Ranking and selection” procedures are statistical procedures appropriate for use in situations where the experimenter's goal is to “select the best” (selection) or to “rank competing alternatives” (ranking). These goals are often present in simulation studies, which are often performed in order to select that one of several procedures (for running a real-world system) which is “best.” (For a discussion of ranking in simulation, see (2).) Most previous work in this area has dealt with situations where either one has a univariate response, or where a simple univariate function of the multivariate response characterizes the “goodness” of a procedure. (For example, see (3) for an introduction and some procedures especially useful in univariate-response simulation settings, (8) for a comprehensive review of the area, and (4) for a discussion of selection in simulation and related statistical problems and procedures.) In a recent excellent expository book on the area (7), Gibbons, Olkin, and Sobel noted (Chapter 15, p. 390) that “The whole field [of multivariate-response ranking and selection] is as yet undeveloped and the reader is encouraged to regard this chapter as an introduction to a wide area that will see considerable development in the future as more meaningful models are formulated.” In this paper we outline a selection model recently developed for this multiple-response problem (6) and develop an example of its use and recommendations for its implementation.