---
title: What is Multi-armed bandit?
url: https://markin.ai/glossary/multi-armed-bandit
category: Experimentation and causality
---

# Multi-armed bandit

> A multi-armed bandit allocates traffic adaptively toward variants that are performing well while continuing to sample the others. It reduces the cost of testing inferior options, at the price of a less clean estimate of each variant's true effect.

## Why it matters for ARPU

Bandits suit high-volume, low-stakes choices such as creative selection. Pricing, incentives and retention offers usually deserve a clean experiment instead.

## Related terms

- [A/B test](https://markin.ai/glossary/ab-test), An A/B test randomly assigns customers to two or more variants and compares a pre-declared metric between them.
- [Reinforcement learning for marketing](https://markin.ai/glossary/reinforcement-learning-for-marketing), Reinforcement learning treats customer decisions as a sequence, optimising cumulative long-term reward rather than the next click.
- [Sequential testing](https://markin.ai/glossary/sequential-testing), Sequential testing allows results to be monitored continuously and stopped early without inflating false positives, using always-valid confidence intervals instead of a single fixed-horizon check.
- [Novelty effect](https://markin.ai/glossary/novelty-effect), The novelty effect is a temporary response lift caused by a change being new rather than better.

## Go deeper

- [The next best action model](https://markin.ai/blog/next-best-action-model), Explore and exploit in decisioning.
- [Optimizely comparison](https://markin.ai/compare/markin-and-optimizely), Testing tools vs a decision layer.

Source: https://markin.ai/glossary/multi-armed-bandit