def test_pull_returns_a_valid_arm(self, k, e, rounds):
egreedy = EpsilonGreedy(num_arms=k, epsilon=e)
for _ in range(rounds):
arm = egreedy.pull()
self.assertGreaterEqual(arm, 0)
self.assertLess(arm, k)
def test_pull_does_not_change_state(self, k, e, rounds):
egreedy = EpsilonGreedy(num_arms=k, epsilon=e)
for _ in range(rounds):
egreedy.pull()
self.assertEqual(0, sum(egreedy.pull_counts))
self.assertEqual(0, sum(egreedy.rewards))
def test_the_best_arm_based_on_feedback(self, k, e, rounds):
p = np.random.random(k)
p = p / np.sum(p)
egreedy = EpsilonGreedy.create(e, rewards=p)
pulls = np.zeros(k, dtype=np.uint64)
for _ in range(rounds):
arm = egreedy.pull()
pulls[arm] += 1
self.assertEqual(np.argmax(p), np.argmax(pulls))
def test_pull_does_not_change_state_after_any_updates(
self, k, e, rounds, updates):
egreedy = EpsilonGreedy(num_arms=k, epsilon=e)
for _ in range(updates):
egreedy.update(chosen_arm=np.random.randint(0, k),
reward=np.random.random())
pulls = sum(egreedy.pull_counts)
rewards = sum(egreedy.rewards)
for _ in range(rounds):
egreedy.pull()
self.assertEqual(pulls, sum(egreedy.pull_counts))
self.assertEqual(rewards, sum(egreedy.rewards))
def test_init_fails_with_epsilon_above_one(self, k, e):
assume(e > 1.0)
with self.assertRaises(AssertionError,
msg="epsilon should be between [0, 1]"):
_ = EpsilonGreedy(num_arms=k, epsilon=e)
def test_init_fails_with_invalid_nums_arms(self, k, e):
with self.assertRaises(AssertionError,
msg="there should be more than one arm"):
_ = EpsilonGreedy(num_arms=k, epsilon=e)
def test_init_set_config_correctly(self, k, e):
egreedy = EpsilonGreedy(num_arms=k, epsilon=e)
self.assertEqual(k, egreedy.num_arms)
self.assertEqual(e, egreedy.epsilon)