Second order efficiency in the sequential design of experiments (Q762867): Difference between revisions

In sequential design of experiments, an experimenter performs experiments sequentially to make an inference about the true state of nature. The parameter space is assumed finite and there are a finite number of repeatable experiments. A Bayesian formulation is used with a fairly general loss structure. All costs of sampling are scaled by a constant a, and the asymptotic limit is considered as \(a\to 0\) which corresponds with large samples. The main theorem 2.1 gives an approximation for the optimal Bayes risk which has an error o(a \(\sqrt{-\ln a})\) as \(a\to 0\) uniformly over prior distributions. The bulk of the paper deals with the asymptotic analysis of a Markov control problem which concerns driving a discrete time vector valued process from an initial position s to the first quadrant if a finite set \(\{P_ 1,...,P_ k\}\) of allowed distributions for steps are given and the distance from s to the first quadrant approaches infinity.