Multi-armed Bandit

A multi-armed bandit is an experimentation method that, instead of a fixed traffic split, shifts more traffic toward the better-performing versions while running, to reduce the cost of learning.

The name comes from a gambler facing several slot machines who must simultaneously learn which is more generous and play it more; a multi-armed bandit automates this 'explore versus exploit' tension. Unlike A/B, which keeps traffic equal until the end and then decides, a bandit gradually starves the weaker version of traffic, so less money is wasted on the loser during the run. Its price is that, because allocation keeps changing, computing statistical significance is harder and interpreting the result more complex. A bandit shines when the goal is to maximize outcome while running (a short campaign, say), not when the goal is a clean scientific answer about the size of the effect.