Confidence intervals for policy evaluation in adaptive experiments

Significance Randomized controlled trials are central to the scientific process, but they can be costly. For example, a clinical trial may assign patients treatments that detrimental them. Adaptive experimental designs, such as multiarmed bandit algorithms, reduce costs by increasing probability of assigning promising over course experiment. However, because observations collected these methods dependent and their distribution is nonstationary, statistical inference challenging. We propose treatment-effect estimator has an asymptotically unbiased normal test statistic under straightforward, relatively weak conditions on adaptive design. This generalizes for variety parameters interest.