def testScalarizeObjectivesWrongInputRankRaisesError(self):
objectives_tensor = tf.constant([1], dtype=tf.float32)
with self.assertRaisesRegexp(
ValueError,
'The objectives_tensor should be rank-3, but is rank-1'):
greedy_multi_objective_policy.scalarize_objectives(
objectives_tensor, self._scalarizer)
def testPredictedRewards(self):
policy = greedy_multi_objective_policy.GreedyMultiObjectiveNeuralPolicy(
self._time_step_spec,
self._action_spec,
self._scalarizer,
self._create_objective_networks(),
emit_policy_info=(
policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN,
policy_utilities.InfoFields.
MULTIOBJECTIVE_SCALARIZED_PREDICTED_REWARDS_MEAN))
observations = tf.constant([[1, 2], [2, 1]], dtype=tf.float32)
time_step = ts.restart(observations, batch_size=2)
action_step = policy.action(time_step)
self.assertEqual(action_step.action.shape.as_list(), [2])
self.assertEqual(action_step.action.dtype, tf.int32)
# Initialize all variables
self.evaluate(tf.compat.v1.global_variables_initializer())
self.assertAllEqual(self.evaluate(action_step.action), [2, 0])
# The expected values are obtained by passing the observation through the
# Keras dense layer of the DummyNet (defined above).
predicted_rewards_expected_array = np.array([[[8, 11, 14], [12, 8, 13],
[11, 14, 8]],
[[5, 8, 11], [10, 5, 9],
[8, 11, 5]]])
p_info = self.evaluate(action_step.info)
predicted_rewards = getattr(
p_info, policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN)
self.assertAllClose(predicted_rewards,
predicted_rewards_expected_array)
self.assertAllClose(
getattr(
p_info, policy_utilities.InfoFields.
MULTIOBJECTIVE_SCALARIZED_PREDICTED_REWARDS_MEAN),
greedy_multi_objective_policy.scalarize_objectives(
predicted_rewards, policy.scalarizer))
def testScalarizeObjectivesWrongNumberOfObjectiveRaisesError(self):
objectives_tensor = tf.constant([[[1, 2, 3]], [[4, 5, 6]]],
dtype=tf.float32)
with self.assertRaisesRegexp(
ValueError,
'The number of input objectives should be 3, but is 1'):
self.evaluate(
greedy_multi_objective_policy.scalarize_objectives(
objectives_tensor, self._scalarizer))
def testScalarizeObjectives(self):
objectives_tensor = tf.constant(
[[[1, 2], [3, 4], [5, 6]], [[7, 8], [9, 10], [11, 12]]],
dtype=tf.float32)
scalarized_reward = greedy_multi_objective_policy.scalarize_objectives(
objectives_tensor, self._scalarizer)
self.assertAllClose(self.evaluate(scalarized_reward),
[[3, 4], [9, 10]],
rtol=1e-4,
atol=1e-3)