Reflections on Bandit Problems and Selection Methods in Uncertain Environments

The behavior of selection methods used in evolutionary algorithms that operate in uncertain environments is investigated in the framework of parametric two{armed bandit problems. Asymptotically optimal selection strategies are based on the sequential probability ratio test which i s p r o ved to perform up to four times better than analogous strategies based on the optimal xed size sample test. A variant of local binary tournament selection in a spatially structured population is shown to behave like a sequential test provided that the population size is optimally adjusted.