Best-Arm Identification

Best-arm identification is the problem of choosing the best among several options when each evaluation is noisy, as in bandit experiments, model evaluation, or hyperparameter search. Algorithms such as successive elimination and halving, as well as Track-and-Stop, allocate samples adaptively and use confidence bounds or likelihood tests to decide what to keep and when to stop. Work considers fixed budgets or prescribed confidence levels, and extends these methods to pairwise comparisons, multiple objectives, costs, fairness constraints, and nonstationary rewards; evaluations often use synthetic Gaussian or Bernoulli bandits and real model or policy benchmarks. A central challenge is reducing evaluations by exploiting gaps, correlations, or prior information without losing reliable guarantees when assumptions fail or the best options are nearly tied.