I read http://synference.blogspot.com/2013/12/multi-armed-contextual-bandits.html and thought I could write a more concise version. I also had some questions.
A/b testing is great! Wikipedia uses it every year to pick the best fundraising banner and a/b testing makes them millions of dollars more.
The contextual bandits algorithm can do even better.
2 improvements over traditional a/b testing
- The algorithm doesn't have to have a single winner. It will try to show the most effective version for each user.
- The algorithm automatically optimizes. If it has a good guess at which version a user will prefer, it uses that. If not, it picks randomly to collect more data.
How to Use It
There's a paid service at http://www.synference.com/ with a free beta.
My Questions After Reading
- How do we generate the versions to use for contextual bandits?
- Does the algorithm support adding new versions later?
- Does the algorithm support removing poorly performing versions?