Be the Bandit

Three versions of an AI tutor. Sixty learners. Who gets which?

You wrote three versions of an AI tutor, A, B and C. One of them helps more learners finish, but you don't know which.

Learners arrive one at a time. You pick a version for each one and see if they finished. Send too many to a weak version and they lose out. Never try a version and you'll never learn it was the best.

How to play
  1. Pick A, B or C for each learner.
  2. Watch who finishes.
  3. After 60, see how a classic algorithm did with the same learners.
Learner 1 of 60 is waiting. Which version do they get?

Talk about it

  1. When did you stop trying the version that looked worst? Were you wrong?
  2. Play a few games. Sometimes Thompson sampling backs the wrong version for a while. How would you make sure every version keeps getting a fair chance?
  3. Here the reward was "finished". What should it be in a real course? What could go wrong?
  4. A new version D shows up halfway. How should it join?

© Stanford 2026 | Website designed and made by Chris Piech. | Past versions of this course: Spring 2022, Fall 2019