def test_pull_returns_a_valid_arm(self, k, tau, rounds):
softmax = Softmax(num_arms=k, temperature=tau)
for _ in range(rounds):
arm = softmax.pull()
self.assertGreaterEqual(arm, 0)
self.assertLess(arm, k)
def test_pull_does_not_change_state(self, k, tau, rounds):
softmax = Softmax(num_arms=k, temperature=tau)
for _ in range(rounds):
softmax.pull()
self.assertEqual(0, sum(softmax.pull_counts))
self.assertEqual(0, sum(softmax.rewards))
def test_the_best_arm_based_on_feedback(self, k, rounds):
p = np.random.random(k)
p = p / np.sum(p)
softmax = Softmax.create(temperature=1.0, rewards=p)
pulls = np.zeros(k, dtype=np.uint64)
for _ in range(rounds):
arm = softmax.pull()
pulls[arm] += 1
self.assertEqual(np.argmax(p), np.argmax(pulls))
def test_pull_does_not_change_state_after_any_updates(
self, k, tau, rounds, updates):
softmax = Softmax(num_arms=k, temperature=tau)
for _ in range(updates):
softmax.update(chosen_arm=np.random.randint(0, k),
reward=np.random.random())
pulls = sum(softmax.pull_counts)
rewards = sum(softmax.rewards)
for _ in range(rounds):
softmax.pull()
self.assertEqual(pulls, sum(softmax.pull_counts))
self.assertEqual(rewards, sum(softmax.rewards))
def test_init_fails_with_negative_temperature(self, k, tau):
assume(tau < 0.0)
with self.assertRaises(AssertionError,
msg="temperature should positive"):
_ = Softmax(num_arms=k, temperature=tau)
def test_init_fails_with_invalid_nums_arms(self, k, tau):
with self.assertRaises(AssertionError,
msg="there should be more than one arm"):
_ = Softmax(num_arms=k, temperature=tau)
def test_init_set_config_correctly(self, k, tau):
softmax = Softmax(num_arms=k, temperature=tau)
self.assertEqual(k, softmax.num_arms)
self.assertEqual(tau, softmax.temperature)