Approaches in Sequential Design of Experiments

Abstract : Sequential design of experiments refers to problems of inference characterized by the fact that as data accumulate, the experimenter can choose whether or not to experiment further. If he decides to experiment further, he can decide which experiment to carry out next and if he decides to stop experimentation, he must decide what terminal decision to make. The literature contains two broad types of general approach and several major classes of applications. One general approach is that of stochastic approximation. Three variations are the Robbins-Monro methods, Box-Wilson response surface methods and the up-and-down methods. The other general approach consists of finding optimal or asymptotically optimal designs, generally in a Bayesian decision theoretic context. Special classes of applications include survey sampling, multilevel continuous sampling inspection, selecting the largest of k populations, which includes clinical trials and two-armed bandit-type problems, screening experiments, group testing, and search problems.