piabrota← proof index

A/B test as a special case of a bandit

A/B test as a special case of a bandit A workflow diagram generated by Archify. 01 / A/B/n test 02 / Contextual bandit Context and belief Decide Observe Learn Split traffic · 1/K per variant · A/B/n test › Decide Split traffic 1/K per variant Run test · fixed period · A/B/n test › Observe Run test fixed period Analyze · pick winner · A/B/n test › Learn Analyze pick winner Ship winner · 100% traffic · A/B/n test › Learn Ship winner 100% traffic Observe context · per request · Contextual bandit › Context and belief Observe context per request Estimate arms · mean + width · Contextual bandit › Context and belief Estimate arms mean + width Pick arm · explore / exploit · Contextual bandit › Decide Pick arm explore / exploit Get reward · chosen arm only · Contextual bandit › Observe Get reward chosen arm only Update model · posterior · Contextual bandit › Learn Update model posterior ε = 1 next round Legend Feedback Decision Model update Commit Input