Multi-Armed Bandit

A multi-armed bandit balances trying new choices with using the choice that is winning so far.

What it does

A multi-armed bandit balances trying new choices with using the choice that is winning so far.

Explore: Testing unknowns • Exploit: Using the best option

How it works

• Try several options to learn about them.
• Use the option that works best so far.
• Keep testing a little so it can find a better option later.

Uses and applications

• A website can test which button color gets more clicks.
• A game can test which reward keeps players interested.
• A learning app can try different practice styles and use the one that helps most.

In short

It is like trying several snacks, then choosing the favorite.

Related methods

• Q-Learning: Handles actions whose effects carry over to later states.
• Collaborative Filtering: Recommends from similar users rather than live trial and reward.

Back to Reinforcement learning, agents and planning · All topics

The interactive demonstration and the Python walkthrough for this topic need JavaScript.