Sequentially designed experiments for screening out bad models with F tests

Techniques for discrimination should depend on the purpose of discrimination. That considered here is to screen out invalid models from a set of contemplated models. The function of the experimental design is to provide opportunities for evidence of invalidity of false models to be as strong as possible. Once some observations have been made, information is available to help design subsequent trials. One way to use this information is, for each possible location of the next trial, to forecast the next response and hence the next set of significance levels for the contemplated models. The location of the next trial may then be chosen to optimize some property of these forecast significance levels. In ? 2 these levels are defined. In ? 3 a design procedure is proposed and in ? 4 this procedure is illustrated with a simple polynomial example involving four hypothesized models.