Skip to main content

COMPEL Glossary / multi-armed-bandit

Multi-armed bandit

An online experimentation strategy that shifts traffic toward better-performing variants during the experiment — trading statistical power for exploitation of early wins.

What this means in practice

Appropriate for time-sensitive or high-cost decisions; inappropriate where regulatory evidence of statistical rigor is required.

Synonyms

bandit algorithm , contextual bandit , Thompson sampling

See also

  • Online evaluation — Assessment of an AI system under live traffic using randomized or sequential experimental designs — A/B test, multi-armed bandit, canary, or interleaving.