The last success problem is an optimal stopping problem that aims to maximize the probability of stopping on the last success in a sequence of $n$ Bernoulli trials. In a typical setting where complete information about the distributions is available, Bruss provided an optimal stopping policy ensuring a winning probability of $1/e$. However, assuming complete knowledge of the distributions is unrealistic in many practical applications. In this paper, we investigate a variant of the last success problem where we have single-sample access from each distribution instead of having comprehensive knowledge of the distributions. Nuti and Vondr\'{a}k demonstrated that a winning probability exceeding $1/4$ is unachievable for this setting, but it remains unknown whether a stopping policy that meets this bound exists. We reveal that Bruss's policy, when applied with the estimated success probabilities, cannot ensure a winning probability greater than $(1-e^{-4})/4\approx 0.2454~(< 1/4)$, irrespective of the estimations from the given samples. Nevertheless, we demonstrate that by setting the threshold the second-to-last success in samples and stopping on the first success observed \emph{after} this threshold, a winning probability of $1/4$ can be guaranteed.
翻译:暂无翻译