Skip to main content
Multi-Armed Bandits are solutions that automatically find the best variant among a group of candidates, balancing between “exploring” options and “exploiting” the best option by dynamically allocating traffic. On Statsig, Bandits are used to pick the best user experience to drive a target metric or action. Statsig’s Autotune (the Multi-Armed Bandit solution) allocates traffic towards high-performing variants and can eventually identify a winning variant. Statsig’s Autotune AI (the Contextual Bandit solution) is a personalization tool that serves users the best variant determined by a machine learning model trained on previous observations.

​
How Autotune works

Autotune is Statsig’s Bayesian Multi-Armed Bandit, and Autotune AI is Statsig’s Contextual Bandit. Both Autotune products will test and measure different variations and their effect on a target metric.
  • The multi-armed bandit continuously adjusts traffic towards the best performing variations until it can confidently pick the best variation. The winning variation will then receive 100% of traffic.
  • The contextual bandit personalizes what variant a user sees based on “context” - or provided user/interaction attributes - to serve each user the variation predicted to be best (i.e. personalization).
Contextual Bandits are a subset of Multi-Armed Bandits; both seek to balance the “explore”/“exploit” problem - balancing between “exploiting” the current best known solution versus “exploring” to get more information about other solutions. Our blog posts on Multi-Armed Bandits and Contextual Bandits go into depth on use cases and considerations. The chart below describes some of the main considerations on when to use either bandits, a ranking engine, or an experiment.
A/B/n TestMulti-Armed Bandit (Autotune)Contextual Bandit (Autotune AI)Ranking Engine
Typical # Variants2-34-84-8Arbitrary #
Personalization FactorNoneNoneModerateHigh
Input Data RequiredNoneVery Little (100+ samples)Little - generally 1000+ samplesTens of thousands to millions of samples
Model EfficacyNoneBasicModerateHigh
Identifies Best VariantYesYesNoNo
Consistent User AssignmentYesNoNoNo