def testNearZeroObjectivesAndDirectionCoordinates(self):
multi_objectives = tf.constant([[3, 0, 0]], dtype=tf.float32)
# direction with small non-zero coordinates.
nonzero = 2 * multi_objective_scalarizer.HyperVolumeScalarizer.ALMOST_ZERO
direction = [1, nonzero, nonzero]
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
self.assertAllClose(self.evaluate(scalarizer(multi_objectives)),
self.evaluate(tf.constant([0], dtype=tf.float32)))
# direction with coordinates that are so small in magnitude that would be
# considered zeros.
direction = [1, 1e-30, 1e-30]
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
self.assertAllClose(self.evaluate(scalarizer(multi_objectives)),
self.evaluate(tf.constant([3], dtype=tf.float32)))
# direction with absolute zero coordinates.
direction = [1, 0, 0]
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
self.assertAllClose(self.evaluate(scalarizer(multi_objectives)),
self.evaluate(tf.constant([3], dtype=tf.float32)))
def testCustomTransform(self):
transform = lambda m, s, o: m * s + o
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
[1, 0, 0], self._hv_params, transform)
self.assertAllClose(
self.evaluate(scalarizer(self._batch_multi_objectives)),
self.evaluate(tf.constant([3, 6], dtype=tf.float32)))
transform2 = lambda m, s, o: tf.multiply(m, m) * s + o
scalarizer2 = multi_objective_scalarizer.HyperVolumeScalarizer(
[0.1, 0.2, 0.2], self._hv_params, transform2)
self.assertAllClose(
self.evaluate(scalarizer2(self._batch_multi_objectives)),
self.evaluate(tf.constant([1.5, 24.], dtype=tf.float32)))
def testAxisAlignedDirection(self):
direction = [0, 1, 0]
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
self.assertAllEqual(
self.evaluate(scalarizer(self._batch_multi_objectives)),
self.evaluate(tf.constant([2, 5], dtype=tf.float32)))
def testDirectionNormalization(self):
for direction in [[1, 2, 2], [0.1, 0.2, 0.2]]:
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
self.assertAllClose(
self.evaluate(scalarizer(self._batch_multi_objectives)),
self.evaluate(tf.constant([1.5, 6], dtype=tf.float32)))
def testNegativeObjectives(self):
direction = [1, 0, 0]
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
self.assertAllEqual(
self.evaluate(scalarizer(-1.0 * self._batch_multi_objectives)),
self.evaluate(tf.constant([0, 0], dtype=tf.float32)))
def setUp(self):
super(ScalarizeObjectivesTest, self).setUp()
hv_params = [
multi_objective_scalarizer.HyperVolumeScalarizer.PARAMS(slope=1,
offset=0)
] * 3
self._scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
[0, 1, 0.01], hv_params)
def testInvalidObjectives(self):
with self.assertRaisesRegex(
ValueError, 'The number of input objectives should be'):
direction = [1, 0, 0]
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
multi_objectives = tf.constant([[1, 2, 3, 4]], dtype=tf.float32)
scalarizer(multi_objectives)
def testCallDirectionWrongBatchSizeRaisesValueError(self):
direction = [1, 0, 0]
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
with self.assertRaisesRegex(ValueError, 'does not match the shape of'):
scalarizer.set_parameters(direction=tf.constant(
[[0, 1, 0], [1, 0, 0], [0, 0, 1]], dtype=tf.float32),
transform_params={})
self.evaluate(scalarizer(self._batch_multi_objectives))
def testNegativeObjectives(self):
direction = [1, 0, 0]
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
multi_objectives = tf.constant([[-3, 2, 1], [-6, 5, 4]],
dtype=tf.float32)
self.assertAllEqual(
self.evaluate(scalarizer(multi_objectives)),
self.evaluate(tf.constant([0, 0], dtype=tf.float32)))
def testSetParametersDirectionOnly(self):
direction = [1, 0, 0]
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
scalarizer.set_parameters(direction=tf.constant([[0, 1, 0], [0, 0, 1]],
dtype=tf.float32),
transform_params={})
self.assertAllClose(
self.evaluate(scalarizer(self._batch_multi_objectives)),
self.evaluate(tf.constant([2, 4], dtype=tf.float32)))
def testObjectiveTranformation(self):
hv_params = copy.deepcopy(self._hv_params)
hv_params[0] = multi_objective_scalarizer.HyperVolumeScalarizer.PARAMS(
slope=-1, offset=2)
direction = [1, 0, 0]
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, hv_params)
self.assertAllEqual(
self.evaluate(scalarizer(-1.0 * self._batch_multi_objectives)),
self.evaluate(tf.constant([5, 8], dtype=tf.float32)))
def testSetParametersDirectionWithWrongNumberOfObjectivesRaisesValueError(
self):
direction = [1, 0, 0]
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
with self.assertRaisesRegex(
ValueError, 'The number of objectives in scalarization parameter'):
scalarizer.set_parameters(
direction=tf.constant([[0, 1]] * 2, dtype=tf.float32),
transform_params={})
def testSetParametersInvalidTransformationParamsKeysRaisesValueError(self):
direction = [1, 0, 0]
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
with self.assertRaisesRegex(ValueError,
'All transform_params keys should be'):
scalarizer.set_parameters(
direction=tf.constant([[0, 1, 0], [0, 0, 1]], dtype=tf.float32),
transform_params={
'weights': tf.constant([1, 1, 1], dtype=tf.float32)
})
def testSetParametersDirectionAndOffset(self):
direction = [1, 0, 0]
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
scalarizer.set_parameters(
direction=tf.constant([[0, 1, 0], [0, 0, 1]], dtype=tf.float32),
transform_params={
multi_objective_scalarizer.HyperVolumeScalarizer.OFFSET_KEY:
tf.constant([[0.5, 0.5, 0.5], [0.1, 0.1, 0.1]],
dtype=tf.float32)
})
self.assertAllClose(
self.evaluate(scalarizer(self._batch_multi_objectives)),
self.evaluate(tf.constant([2.5, 4.1], dtype=tf.float32)))
def setUp(self):
super(GreedyRewardPredictionPolicyTest, self).setUp()
self._obs_spec = tensor_spec.TensorSpec([2], tf.float32)
self._time_step_spec = ts.time_step_spec(self._obs_spec)
self._action_spec = tensor_spec.BoundedTensorSpec((), tf.int32, 0, 2)
hv_params = [
multi_objective_scalarizer.HyperVolumeScalarizer.PARAMS(slope=1,
offset=0)
] * 3
self._scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
[0, 1, 0.01], hv_params)
self._bias = np.asarray([-1, -1, -1])
self._kernel_weights = [
np.asarray([[1, 2, 3], [4, 5, 6]]),
np.asarray([[3, 1, 2], [5, 4, 6]]),
np.asarray([[2, 3, 1], [5, 6, 4]])
]
def setUp(self):
super(MultiObjectiveAgentTest, self).setUp()
tf.compat.v1.enable_resource_variables()
self._time_step_spec = ts.time_step_spec(
observation_spec=tensor_spec.TensorSpec([2], tf.float32),
reward_spec=tensor_spec.TensorSpec([3], tf.float32))
self._action_spec = tensor_spec.BoundedTensorSpec(
dtype=tf.int32, shape=(), minimum=0, maximum=2)
self._observation_spec = self._time_step_spec.observation
hv_params = [
multi_objective_scalarizer.HyperVolumeScalarizer.PARAMS(
slope=1, offset=0)
] * 3
self._scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
[0, 1, 0.01], hv_params)
self._bias = np.asarray([-1, -1, -1])
self._kernel_weights = [
np.asarray([[1, 2, 3], [4, 5, 6]]),
np.asarray([[3, 1, 2], [5, 4, 6]]),
np.asarray([[2, 3, 1], [5, 6, 4]])
]
def testCustomTransform(self):
transform = lambda m, s, o: m * s + o
scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
[1, 0, 0], self._hv_params, transform)
self.assertAllClose(scalarizer(self._batch_multi_objectives), [3, 6])
transform2 = lambda m, s, o: tf.multiply(m, m) * s + o
scalarizer2 = multi_objective_scalarizer.HyperVolumeScalarizer(
[0.1, 0.2, 0.2], self._hv_params, transform2)
self.assertAllClose(scalarizer2(self._batch_multi_objectives),
[1.5, 24.])
def default_transform(metrics, slopes, offsets):
transformed_metrics = metrics
return transformed_metrics * slopes + offsets
scalarizer3 = multi_objective_scalarizer.HyperVolumeScalarizer(
[0.1, 0.2, 0.2], self._hv_params, default_transform)
default_scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
[0.1, 0.2, 0.2], self._hv_params)
self.assertAllClose(
self.evaluate(scalarizer3(self._batch_multi_objectives)),
self.evaluate(default_scalarizer(self._batch_multi_objectives)))
# Apply sigmoid to the 2nd metric.
sigmoid_metric_mask = [False, True, False]
batch_size = self._batch_multi_objectives.shape[0]
sigmoid_batch_mask = tf.reshape(
tf.tile(sigmoid_metric_mask, [batch_size]),
[batch_size, len(sigmoid_metric_mask)])
def sigmoid_metric_transform(metrics: tf.Tensor, slopes, offsets):
transformed_values = tf.where(sigmoid_batch_mask,
tf.sigmoid(metrics), metrics)
return transformed_values * slopes + offsets
sigmoid_scalarizer = multi_objective_scalarizer.HyperVolumeScalarizer(
[0.1, 0.2, 0.2], self._hv_params, sigmoid_metric_transform)
self.assertAllClose(sigmoid_scalarizer(self._batch_multi_objectives),
[1.321196, 1.489961])
sigmoid_scalarizer2 = multi_objective_scalarizer.HyperVolumeScalarizer(
[0, 1.0, 0], self._hv_params, sigmoid_metric_transform)
self.assertAllClose(sigmoid_scalarizer2(self._batch_multi_objectives),
[0.880797, 0.993307])
def testZeroLengthDirection(self):
with self.assertRaisesRegex(ValueError, 'nearly-zero vector'):
direction = [0, 0, 0]
multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
def testDirectionWithWrongDimension(self):
with self.assertRaisesRegex(ValueError,
'direction has 2 elements but'):
direction = [1, 0]
multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)
def testInvalidParams(self):
with self.assertRaisesRegex(ValueError, 'nearly-zero vector'):
multi_objective_scalarizer.HyperVolumeScalarizer([], [])
with self.assertRaisesRegex(ValueError, 'at least two objectives'):
multi_objective_scalarizer.HyperVolumeScalarizer(
[1], [self._hv_params[0]])
def testDirectionWithNegativeCoordinates(self):
with self.assertRaisesRegex(ValueError, 'has negative coordinates'):
direction = [-1, 0, 0]
multi_objective_scalarizer.HyperVolumeScalarizer(
direction, self._hv_params)