A multi-armed bandit optmization framework for Rails
Ruby JavaScript
Pull request Compare This branch is 12 commits ahead, 40 commits behind bmuller:master.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.



Bandido is a fork of the bandit gem, the README for which you may find below this section. It still uses the same code namespace as bandit (Bandit), should be compatible with projects already using bandit. If you run into any problems because of this, consider adding:

require 'bandit'

in your bandit initializer, e.g. config/initializers/bandit.rb

Bandido provides more control over cookies, participant counts and the like. Instead of calling bandit_choose and bandit_convert! in your code, you may use the following method combinations:

bandit_simple_choose/bandit_simple_convert!   (no cookies involved, same as raw impressions)
bandit_session_choose/bandit_session_convert! (session cookie, same as bandit_choose with a fix)
bandit_sticky_choose/bandit_sticky_convert!   (persistent cookies)

bandit_sticky_convert! creates a _converted cookie which stops additional conversions from being counted. For example, if you are collecting email addresses, a user may enter their email address more than once, but you may want to count all these attempts as one conversion.

You may also use the bandit_sticky_choose/bandit_session_convert! combination if you wish to use persistent cookies but allow each user to convert multiple times. This way visitors are always presented with the same alternative until they convert, at which point they are presented with another and so on.

Finally, if you use bandit_simple_convert!, please remember the second argument (alternative) is not optional, as we have no cookie to read from.


Bandit is a multi-armed bandit optimization framework for Rails. It provides an alternative to A/B testing in Rails. For background and a comparison with A/B testing, see the whybandit.rdoc document or the blog post here.


First, add the following to your Gemfile in your Rails 3 app:

gem 'bandit'

Then, run the following:

bundle install
rails generate bandit:install

You can then edit the bandit.yml file in your config directory to set your storage and player parameters. Redis, memcache, and memory storage options are available. Memory storage should only be used for testing.

See the file players.rdoc for information about available players.


To set up an experiment, add it either somewhere in your code or in the bandit initializer. Creating an experiment is simple:

Bandit::Experiment.create(:click_test) { |exp|
  exp.alternatives = [ 20, 30, 40 ]
  exp.title = "Click Test"
  exp.description = "A test of clicks on purchase page with varying link sizes."


To get an alternative (per viewer, based on cookies):

<%= bandit_choose :click_test %>

For instance, in a link:

<%= link_to "new purchase", new_purchase_path, :style => "font-size: #{bandit_choose(:click_test)}px;" %>

You can force a particular alternative by adding a query parameter named “bandit_<experiment name>” and setting it's value to the alternative you want. For instance, given the above experiment in the configuration example:


will then force the alternative to be “40”.


To track a conversion in your controller:

bandit_convert! :click_test


rails generate bandit:dashboard

Then, add the following to your config/routes.rb file:

resources :bandit

To see a dashboard with relevant information, go to:



To run tests:

rake test_memory
rake test_memcache
rake test_redis

To produce fake data for the past week, first create an experiment definition. Then, run the following rake task:

rake bandit:populate_data[<experiment_name>]

For instance, to generate a week's worth of fake data for the click_test above:

rake bandit:populate_data[click_test]

Fault Tolerance

If the storage mechanism fails, then Bandit will automatically switch to in memory storage. It will then check every 5 minutes after that to see if the original storage mechanism is back up. If you have distributed front ends then each front end will continue to optimize (based on the in memory storage), but this optimization will be inefficient compared to shared storage among all front ends.