def testCtorValueErrorMissingEpsilonEndValue(self):
with self.assertRaises(ValueError):
py_epsilon_greedy_policy.EpsilonGreedyPolicy(
self.greedy_policy,
0.99,
random_policy=self.random_policy,
epsilon_decay_end_count=100)
def testActionSelectionWithEpsilonDecay(self):
policy = py_epsilon_greedy_policy.EpsilonGreedyPolicy(
self.greedy_policy, 0.9, random_policy=self.random_policy,
epsilon_decay_end_count=10,
epsilon_decay_end_value=0.4)
time_step = mock.MagicMock()
# Replace the random generator with fixed behaviour
random = mock.MagicMock()
policy._rng = random
# 0.8 < 0.9 and 0.8 < 0.85, so random policy should be used.
policy._rng.rand.return_value = 0.8
for _ in range(2):
policy.action(time_step)
self.random_policy.action.assert_called_with(time_step)
self.assertEqual(2, self.random_policy.action.call_count)
self.assertEqual(0, self.greedy_policy.action.call_count)
# epislon will change from [0.8 to 0.4], and greedy policy should be used
for _ in range(8):
policy.action(time_step)
self.greedy_policy.action.assert_called_with(time_step, policy_state=())
self.assertEqual(2, self.random_policy.action.call_count)
self.assertEqual(8, self.greedy_policy.action.call_count)
# 0.399 < 0.4, random policy should be used.
policy._rng.rand.return_value = 0.399
self.random_policy.reset_mock()
for _ in range(5):
policy.action(time_step)
self.random_policy.action.assert_called_with(time_step)
self.assertEqual(5, self.random_policy.action.call_count)
# greedy policy should not be called any more
self.assertEqual(8, self.greedy_policy.action.call_count)
def testZeroState(self):
policy = py_epsilon_greedy_policy.EpsilonGreedyPolicy(
self.greedy_policy, 0.5, random_policy=self.random_policy)
policy.get_initial_state()
self.greedy_policy.get_initial_state.assert_called_once_with(
batch_size=None)
self.random_policy.get_initial_state.assert_called_once_with(
batch_size=None)
def testActionAlwaysRandom(self):
policy = py_epsilon_greedy_policy.EpsilonGreedyPolicy(
self.greedy_policy, 1, random_policy=self.random_policy)
time_step = mock.MagicMock()
for _ in range(5):
policy.action(time_step)
self.random_policy.action.assert_called_with(time_step)
self.assertEqual(5, self.random_policy.action.call_count)
self.assertEqual(0, self.greedy_policy.action.call_count)
def testActionSelection(self):
policy = py_epsilon_greedy_policy.EpsilonGreedyPolicy(
self.greedy_policy, 0.9, random_policy=self.random_policy)
time_step = mock.MagicMock()
# Replace the random generator with fixed behaviour
random = mock.MagicMock()
policy._rng = random
# 0.8 < 0.9, so random policy should be used.
policy._rng.rand.return_value = 0.8
policy.action(time_step)
self.random_policy.action.assert_called_with(time_step)
self.assertEqual(1, self.random_policy.action.call_count)
self.assertEqual(0, self.greedy_policy.action.call_count)
# 0.91 > 0.9, so greedy policy should be used.
policy._rng.rand.return_value = 0.91
policy.action(time_step)
self.greedy_policy.action.assert_called_with(time_step, policy_state=())
self.assertEqual(1, self.random_policy.action.call_count)
self.assertEqual(1, self.greedy_policy.action.call_count)
def testCtorValueErrorEpsilonMorThanOne(self):
with self.assertRaises(ValueError):
py_epsilon_greedy_policy.EpsilonGreedyPolicy(
self.greedy_policy, 1.00001, random_policy=self.random_policy)
def testCtorValueErrorNegativeEpsilon(self):
with self.assertRaises(ValueError):
py_epsilon_greedy_policy.EpsilonGreedyPolicy(
self.greedy_policy, -0.00001, random_policy=self.random_policy)
def testCtorAutoRandomPolicy(self):
self.greedy_policy.action_spec = mock.MagicMock()
policy = py_epsilon_greedy_policy.EpsilonGreedyPolicy(
self.greedy_policy, 0.5)
self.assertEqual(self.greedy_policy.action_spec(),
policy._random_policy.action_spec())
def testCtorAutoRandomPolicy(self):
policy = py_epsilon_greedy_policy.EpsilonGreedyPolicy(
self.greedy_policy, 0.5)
self.assertEqual(self.greedy_policy.action_spec,
policy._random_policy.action_spec)