Skip to Main Content
Some numeric results are presented from a distribution-free analysis of how many design patterns to sample from each of two classes, for stationary, discrete-measurement pattern classifiers. A table of design sample size partitions is presented, that maximizes the mean accuracy, given the total number of design patterns and the prior class probabilities. The maximization is also over the measurement complexity (total number of resolvable measurement values). It is shown that the optimal sample size partitions are biased, in that fewer patterns should be taken of the more probable class than is implied by the prior class probabilities. An insensitivity to nonoptimal partitions is exhibited, which appears to be similar to that seen in other Bayesian decision problems.