To begin with, let's start building an extremely simplistic bandit algorithm. The actual bandit class will have two methods: choose_treatment and log_payout. The first method will recommend the best treatment to choose, and log_payout will be used to report back on how effective the recommended treatment was.

The simplest way to approach this algorithm from a test-driven perspective is to start with a single treatment so that the algorithm only has one thing to recommend. The code for this test looks like the following:

from nose.tools import assert_equal
import simple_bandit

def given_a_single_treatment_test():
  bandit = simple_bandit.SimpleBandit(['A'])
  chosen_treatment = bandit.choose_treatment()
  assert_equal(chosen_treatment, 'A', 'Should choose the only available option.')

Again, we'll start from such a simple place that making the test pass will seem too easy:

class SimpleBandit:
  def __init__(self, treatments):
    self._treatments = treatments
  def choose_treatment...