def __init__(self, encoding_network: types.Network, encoding_dim: int, reward_layer: tf.keras.layers.Dense, epsilon_greedy: float, actions_from_reward_layer: types.Bool, cov_matrix: Sequence[types.Float], data_vector: Sequence[types.Float], num_samples: Sequence[types.Int], time_step_spec: types.TimeStep, alpha: float = 1.0, emit_policy_info: Sequence[Text] = (), emit_log_probability: bool = False, accepts_per_arm_features: bool = False, distributed_use_reward_layer: bool = False, observation_and_action_constraint_splitter: Optional[ types.Splitter] = None, name: Optional[Text] = None): """Initializes `NeuralLinUCBPolicy`. Args: encoding_network: network that encodes the observations. encoding_dim: (int) dimension of the encoded observations. reward_layer: final layer that predicts the expected reward per arm. In case the policy accepts per-arm features, the output of this layer has to be a scalar. This is because in the per-arm case, all encoded observations have to go through the same computation to get the reward estimates. The `num_actions` dimension of the encoded observation is treated as a batch dimension in the reward layer. epsilon_greedy: (float) representing the probability of choosing a random action instead of the greedy action. actions_from_reward_layer: (boolean variable) whether to get actions from the reward layer or from LinUCB. cov_matrix: list of the covariance matrices. There exists one covariance matrix per arm, unless the policy accepts per-arm features, in which case this list must have a single element. data_vector: list of the data vectors. A data vector is a weighted sum of the observations, where the weight is the corresponding reward. Each arm has its own data vector, unless the policy accepts per-arm features, in which case this list must have a single element. num_samples: list of number of samples per arm. If the policy accepts per- arm features, this is a single-element list counting the number of steps. time_step_spec: A `TimeStep` spec of the expected time_steps. alpha: (float) non-negative weight multiplying the confidence intervals. emit_policy_info: (tuple of strings) what side information we want to get as part of the policy info. Allowed values can be found in `policy_utilities.PolicyInfo`. emit_log_probability: (bool) whether to emit log probabilities. accepts_per_arm_features: (bool) Whether the policy accepts per-arm features. distributed_use_reward_layer: (bool) Whether to pick the actions using the network or use LinUCB. This applies only in distributed training setting and has a similar role to the `actions_from_reward_layer` mentioned above. observation_and_action_constraint_splitter: A function used for masking valid/invalid actions with each state of the environment. The function takes in a full observation and returns a tuple consisting of 1) the part of the observation intended as input to the bandit policy and 2) the mask. The mask should be a 0-1 `Tensor` of shape `[batch_size, num_actions]`. This function should also work with a `TensorSpec` as input, and should output `TensorSpec` objects for the observation and mask. name: The name of this policy. """ policy_utilities.check_no_mask_with_arm_features( accepts_per_arm_features, observation_and_action_constraint_splitter) encoding_network.create_variables() self._encoding_network = encoding_network self._reward_layer = reward_layer self._encoding_dim = encoding_dim if accepts_per_arm_features and reward_layer.units != 1: raise ValueError( 'The output dimension of the reward layer must be 1, got' ' {}'.format(reward_layer.units)) if not isinstance(cov_matrix, (list, tuple)): raise ValueError( 'cov_matrix must be a list of matrices (Tensors).') self._cov_matrix = cov_matrix if not isinstance(data_vector, (list, tuple)): raise ValueError( 'data_vector must be a list of vectors (Tensors).') self._data_vector = data_vector if not isinstance(num_samples, (list, tuple)): raise ValueError( 'num_samples must be a list of vectors (Tensors).') self._num_samples = num_samples self._alpha = alpha self._actions_from_reward_layer = actions_from_reward_layer self._epsilon_greedy = epsilon_greedy self._dtype = self._data_vector[0].dtype self._distributed_use_reward_layer = distributed_use_reward_layer if len(cov_matrix) != len(data_vector): raise ValueError( 'The size of list cov_matrix must match the size of ' 'list data_vector. Got {} for cov_matrix and {} ' 'for data_vector'.format(len(self._cov_matrix), len((data_vector)))) if len(num_samples) != len(cov_matrix): raise ValueError('The size of num_samples must match the size of ' 'list cov_matrix. Got {} for num_samples and {} ' 'for cov_matrix'.format(len(self._num_samples), len((cov_matrix)))) self._accepts_per_arm_features = accepts_per_arm_features if observation_and_action_constraint_splitter is not None: context_spec, _ = observation_and_action_constraint_splitter( time_step_spec.observation) else: context_spec = time_step_spec.observation if accepts_per_arm_features: self._num_actions = tf.nest.flatten(context_spec[ bandit_spec_utils.PER_ARM_FEATURE_KEY])[0].shape.as_list()[0] self._num_models = 1 else: self._num_actions = len(cov_matrix) self._num_models = self._num_actions cov_matrix_dim = tf.compat.dimension_value(cov_matrix[0].shape[0]) if self._encoding_dim != cov_matrix_dim: raise ValueError('The dimension of matrix `cov_matrix` must match ' 'encoding dimension {}.' 'Got {} for `cov_matrix`.'.format( self._encoding_dim, cov_matrix_dim)) data_vector_dim = tf.compat.dimension_value(data_vector[0].shape[0]) if self._encoding_dim != data_vector_dim: raise ValueError( 'The dimension of vector `data_vector` must match ' 'encoding dimension {}. ' 'Got {} for `data_vector`.'.format(self._encoding_dim, data_vector_dim)) action_spec = tensor_spec.BoundedTensorSpec(shape=(), dtype=tf.int32, minimum=0, maximum=self._num_actions - 1, name='action') self._emit_policy_info = emit_policy_info predicted_rewards_mean = () if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN in emit_policy_info: predicted_rewards_mean = tensor_spec.TensorSpec( [self._num_actions], dtype=tf.float32) predicted_rewards_optimistic = () if (policy_utilities.InfoFields.PREDICTED_REWARDS_OPTIMISTIC in emit_policy_info): predicted_rewards_optimistic = tensor_spec.TensorSpec( [self._num_actions], dtype=tf.float32) if accepts_per_arm_features: chosen_arm_features_info_spec = ( policy_utilities.create_chosen_arm_features_info_spec( time_step_spec.observation)) info_spec = policy_utilities.PerArmPolicyInfo( predicted_rewards_mean=predicted_rewards_mean, predicted_rewards_optimistic=predicted_rewards_optimistic, chosen_arm_features=chosen_arm_features_info_spec) else: info_spec = policy_utilities.PolicyInfo( predicted_rewards_mean=predicted_rewards_mean, predicted_rewards_optimistic=predicted_rewards_optimistic) super(NeuralLinUCBPolicy, self).__init__(time_step_spec=time_step_spec, action_spec=action_spec, emit_log_probability=emit_log_probability, observation_and_action_constraint_splitter=( observation_and_action_constraint_splitter), info_spec=info_spec, name=name)
def __init__(self, time_step_spec: types.TimeStep, action_spec: types.NestedTensorSpec, reward_network: types.Network, temperature: types.FloatOrReturningFloat = 1.0, boltzmann_gumbel_exploration_constant: Optional[ types.Float] = None, observation_and_action_constraint_splitter: Optional[ types.Splitter] = None, accepts_per_arm_features: bool = False, constraints: Tuple[constr.NeuralConstraint, ...] = (), emit_policy_info: Tuple[Text, ...] = (), num_samples_list: Sequence[tf.Variable] = (), name: Optional[Text] = None): """Builds a BoltzmannRewardPredictionPolicy given a reward network. This policy takes a tf_agents.Network predicting rewards and chooses an action with weighted probabilities (i.e., using a softmax over the network estimates of value for each action). Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of BoundedTensorSpec representing the actions. reward_network: An instance of a `tf_agents.network.Network`, callable via `network(observation, step_type) -> (output, final_state)`. temperature: float or callable that returns a float. The temperature used in the Boltzmann exploration. boltzmann_gumbel_exploration_constant: optional positive float. When provided, the policy implements Neural Bandit with Boltzmann-Gumbel exploration from the paper: N. Cesa-Bianchi et al., "Boltzmann Exploration Done Right", NIPS 2017. observation_and_action_constraint_splitter: A function used for masking valid/invalid actions with each state of the environment. The function takes in a full observation and returns a tuple consisting of 1) the part of the observation intended as input to the network and 2) the mask. The mask should be a 0-1 `Tensor` of shape `[batch_size, num_actions]`. This function should also work with a `TensorSpec` as input, and should output `TensorSpec` objects for the observation and mask. accepts_per_arm_features: (bool) Whether the policy accepts per-arm features. constraints: iterable of constraints objects that are instances of `tf_agents.bandits.agents.NeuralConstraint`. emit_policy_info: (tuple of strings) what side information we want to get as part of the policy info. Allowed values can be found in `policy_utilities.PolicyInfo`. num_samples_list: list or tuple of tf.Variable's. Used only in Boltzmann-Gumbel exploration. Otherwise, empty. name: The name of this policy. All variables in this module will fall under that name. Defaults to the class name. Raises: NotImplementedError: If `action_spec` contains more than one `BoundedTensorSpec` or the `BoundedTensorSpec` is not valid. """ policy_utilities.check_no_mask_with_arm_features( accepts_per_arm_features, observation_and_action_constraint_splitter) flat_action_spec = tf.nest.flatten(action_spec) if len(flat_action_spec) > 1: raise NotImplementedError( 'action_spec can only contain a single BoundedTensorSpec.') self._temperature = temperature action_spec = flat_action_spec[0] if (not tensor_spec.is_bounded(action_spec) or not tensor_spec.is_discrete(action_spec) or action_spec.shape.rank > 1 or action_spec.shape.num_elements() != 1): raise NotImplementedError( 'action_spec must be a BoundedTensorSpec of type int32 and shape (). ' 'Found {}.'.format(action_spec)) self._expected_num_actions = action_spec.maximum - action_spec.minimum + 1 self._action_offset = action_spec.minimum reward_network.create_variables() self._reward_network = reward_network self._constraints = constraints self._boltzmann_gumbel_exploration_constant = ( boltzmann_gumbel_exploration_constant) self._num_samples_list = num_samples_list if self._boltzmann_gumbel_exploration_constant is not None: if self._boltzmann_gumbel_exploration_constant <= 0.0: raise ValueError( 'The Boltzmann-Gumbel exploration constant is expected to be ', 'positive. Found: ', self._boltzmann_gumbel_exploration_constant) if self._action_offset > 0: raise NotImplementedError( 'Action offset is not supported when ', 'Boltzmann-Gumbel exploration is enabled.') if accepts_per_arm_features: raise NotImplementedError( 'Boltzmann-Gumbel exploration is not supported ', 'for arm features case.') if len(self._num_samples_list) != self._expected_num_actions: raise ValueError( 'Size of num_samples_list: ', len(self._num_samples_list), ' does not match the expected number of actions:', self._expected_num_actions) self._emit_policy_info = emit_policy_info predicted_rewards_mean = () if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN in emit_policy_info: predicted_rewards_mean = tensor_spec.TensorSpec( [self._expected_num_actions]) bandit_policy_type = () if policy_utilities.InfoFields.BANDIT_POLICY_TYPE in emit_policy_info: bandit_policy_type = ( policy_utilities.create_bandit_policy_type_tensor_spec( shape=[1])) if accepts_per_arm_features: # The features for the chosen arm is saved to policy_info. chosen_arm_features_info = ( policy_utilities.create_chosen_arm_features_info_spec( time_step_spec.observation)) info_spec = policy_utilities.PerArmPolicyInfo( predicted_rewards_mean=predicted_rewards_mean, bandit_policy_type=bandit_policy_type, chosen_arm_features=chosen_arm_features_info) else: info_spec = policy_utilities.PolicyInfo( predicted_rewards_mean=predicted_rewards_mean, bandit_policy_type=bandit_policy_type) self._accepts_per_arm_features = accepts_per_arm_features super(BoltzmannRewardPredictionPolicy, self).__init__(time_step_spec, action_spec, policy_state_spec=reward_network.state_spec, clip=False, info_spec=info_spec, emit_log_probability='log_probability' in emit_policy_info, observation_and_action_constraint_splitter=( observation_and_action_constraint_splitter), name=name)
def testLinearAgentUpdatePerArmFeatures(self, batch_size, context_dim, exploration_policy, dtype, use_eigendecomp=False, set_example_weights=False): """Check that the agent updates for specified actions and rewards.""" # Construct a `Trajectory` for the given action, observation, reward. num_actions = 5 global_context_dim = context_dim arm_context_dim = 3 initial_step, final_step = ( _get_initial_and_final_steps_with_per_arm_features( batch_size, global_context_dim, num_actions, arm_context_dim, num_actions_feature=True)) action = np.random.randint(num_actions, size=batch_size, dtype=np.int32) action_step = policy_step.PolicyStep( action=tf.convert_to_tensor(action), info=policy_utilities.PerArmPolicyInfo( chosen_arm_features=np.arange( batch_size * arm_context_dim, dtype=np.float32).reshape( [batch_size, arm_context_dim]))) experience = _get_experience(initial_step, action_step, final_step) # Construct an agent and perform the update. observation_spec = bandit_spec_utils.create_per_arm_observation_spec( context_dim, arm_context_dim, num_actions, add_num_actions_feature=True) time_step_spec = time_step.time_step_spec(observation_spec) action_spec = tensor_spec.BoundedTensorSpec( dtype=tf.int32, shape=(), minimum=0, maximum=num_actions - 1) agent = linear_agent.LinearBanditAgent( exploration_policy=exploration_policy, time_step_spec=time_step_spec, action_spec=action_spec, use_eigendecomp=use_eigendecomp, accepts_per_arm_features=True, dtype=dtype) self.evaluate(agent.initialize()) weights = tf.linspace( start=1.5, stop=10.5, num=batch_size) if set_example_weights else None loss_info = agent.train(experience, weights) self.evaluate(loss_info) final_a = self.evaluate(agent.cov_matrix) final_b = self.evaluate(agent.data_vector) # Compute the expected updated estimates. global_observation = experience.observation[ bandit_spec_utils.GLOBAL_FEATURE_KEY] arm_observation = experience.policy_info.chosen_arm_features overall_observation = tf.squeeze( tf.concat([global_observation, arm_observation], axis=-1), axis=1) squeezed_rewards = tf.squeeze(experience.reward, axis=1) observation, rewards = _maybe_weight_observation_and_reward( overall_observation, squeezed_rewards, weights) expected_a_new = tf.matmul(observation, observation, transpose_a=True) expected_b_new = bandit_utils.sum_reward_weighted_observations( rewards, observation) self.assertAllClose(expected_a_new, final_a[0]) self.assertAllClose(expected_b_new, final_b[0])
def _distribution(self, time_step, policy_state): observation = time_step.observation if self.observation_and_action_constraint_splitter is not None: observation, _ = self.observation_and_action_constraint_splitter( observation) predictions, policy_state = self._reward_network( observation, time_step.step_type, policy_state) batch_size = tf.shape(predictions)[0] if isinstance(self._reward_network, heteroscedastic_q_network.HeteroscedasticQNetwork): predicted_reward_values = predictions.q_value_logits else: predicted_reward_values = predictions predicted_reward_values.shape.with_rank_at_least(2) predicted_reward_values.shape.with_rank_at_most(3) if predicted_reward_values.shape[ -1] is not None and predicted_reward_values.shape[ -1] != self._expected_num_actions: raise ValueError( 'The number of actions ({}) does not match the reward_network output' ' size ({}).'.format(self._expected_num_actions, predicted_reward_values.shape[1])) mask = constr.construct_mask_from_multiple_sources( time_step.observation, self._observation_and_action_constraint_splitter, self._constraints, self._expected_num_actions) if self._boltzmann_gumbel_exploration_constant is not None: logits = predicted_reward_values # Apply masking if needed. Overwrite the logits for invalid actions to # logits.dtype.min. if mask is not None: almost_neg_inf = tf.constant(logits.dtype.min, dtype=logits.dtype) logits = tf.compat.v2.where(tf.cast(mask, tf.bool), logits, almost_neg_inf) gumbel_dist = tfp.distributions.Gumbel(loc=0., scale=1.) gumbel_samples = gumbel_dist.sample(tf.shape(logits)) num_samples_list_float = tf.stack([ tf.cast(x.read_value(), tf.float32) for x in self._num_samples_list ], axis=-1) exploration_weights = tf.math.divide_no_nan( self._boltzmann_gumbel_exploration_constant, tf.sqrt(num_samples_list_float)) final_logits = logits + exploration_weights * gumbel_samples actions = tf.cast(tf.math.argmax(final_logits, axis=1), self._action_spec.dtype) # Log probability is not available in closed form. We treat this as a # deterministic policy at the moment. log_probability = tf.zeros([batch_size], tf.float32) else: # Apply the temperature scaling, needed for Boltzmann exploration. logits = predicted_reward_values / self._get_temperature_value() # Apply masking if needed. Overwrite the logits for invalid actions to # logits.dtype.min. if mask is not None: almost_neg_inf = tf.constant(logits.dtype.min, dtype=logits.dtype) logits = tf.compat.v2.where(tf.cast(mask, tf.bool), logits, almost_neg_inf) if self._action_offset != 0: distribution = shifted_categorical.ShiftedCategorical( logits=logits, dtype=self._action_spec.dtype, shift=self._action_offset) else: distribution = tfp.distributions.Categorical( logits=logits, dtype=self._action_spec.dtype) actions = distribution.sample() log_probability = distribution.log_prob(actions) bandit_policy_values = tf.fill( [batch_size, 1], policy_utilities.BanditPolicyType.BOLTZMANN) if self._accepts_per_arm_features: # Saving the features for the chosen action to the policy_info. def gather_observation(obs): return tf.gather(params=obs, indices=actions, batch_dims=1) chosen_arm_features = tf.nest.map_structure( gather_observation, observation[bandit_spec_utils.PER_ARM_FEATURE_KEY]) policy_info = policy_utilities.PerArmPolicyInfo( log_probability=log_probability if policy_utilities.InfoFields.LOG_PROBABILITY in self._emit_policy_info else (), predicted_rewards_mean=( predicted_reward_values if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN in self._emit_policy_info else ()), bandit_policy_type=( bandit_policy_values if policy_utilities.InfoFields.BANDIT_POLICY_TYPE in self._emit_policy_info else ()), chosen_arm_features=chosen_arm_features) else: policy_info = policy_utilities.PolicyInfo( log_probability=log_probability if policy_utilities.InfoFields.LOG_PROBABILITY in self._emit_policy_info else (), predicted_rewards_mean=( predicted_reward_values if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN in self._emit_policy_info else ()), bandit_policy_type=( bandit_policy_values if policy_utilities.InfoFields.BANDIT_POLICY_TYPE in self._emit_policy_info else ())) return policy_step.PolicyStep( tfp.distributions.Deterministic(loc=actions), policy_state, policy_info)
def __init__(self, time_step_spec: types.TimeStep, action_spec: types.NestedTensorSpec, reward_network: types.Network, observation_and_action_constraint_splitter: Optional[ types.Splitter] = None, accepts_per_arm_features: bool = False, constraints: Tuple[constr.NeuralConstraint, ...] = (), emit_policy_info: Tuple[Text, ...] = (), name: Optional[Text] = None): """Builds a GreedyRewardPredictionPolicy given a reward tf_agents.Network. This policy takes a tf_agents.Network predicting rewards and generates the action corresponding to the largest predicted reward. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of BoundedTensorSpec representing the actions. reward_network: An instance of a `tf_agents.network.Network`, callable via `network(observation, step_type) -> (output, final_state)`. observation_and_action_constraint_splitter: A function used for masking valid/invalid actions with each state of the environment. The function takes in a full observation and returns a tuple consisting of 1) the part of the observation intended as input to the network and 2) the mask. The mask should be a 0-1 `Tensor` of shape `[batch_size, num_actions]`. This function should also work with a `TensorSpec` as input, and should output `TensorSpec` objects for the observation and mask. accepts_per_arm_features: (bool) Whether the policy accepts per-arm features. constraints: iterable of constraints objects that are instances of `tf_agents.bandits.agents.NeuralConstraint`. emit_policy_info: (tuple of strings) what side information we want to get as part of the policy info. Allowed values can be found in `policy_utilities.PolicyInfo`. name: The name of this policy. All variables in this module will fall under that name. Defaults to the class name. Raises: NotImplementedError: If `action_spec` contains more than one `BoundedTensorSpec` or the `BoundedTensorSpec` is not valid. """ policy_utilities.check_no_mask_with_arm_features( accepts_per_arm_features, observation_and_action_constraint_splitter) flat_action_spec = tf.nest.flatten(action_spec) if len(flat_action_spec) > 1: raise NotImplementedError( 'action_spec can only contain a single BoundedTensorSpec.') action_spec = flat_action_spec[0] if (not tensor_spec.is_bounded(action_spec) or not tensor_spec.is_discrete(action_spec) or action_spec.shape.rank > 1 or action_spec.shape.num_elements() != 1): raise NotImplementedError( 'action_spec must be a BoundedTensorSpec of type int32 and shape (). ' 'Found {}.'.format(action_spec)) self._expected_num_actions = action_spec.maximum - action_spec.minimum + 1 self._action_offset = action_spec.minimum reward_network.create_variables() self._reward_network = reward_network self._constraints = constraints self._emit_policy_info = emit_policy_info predicted_rewards_mean = () if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN in emit_policy_info: predicted_rewards_mean = tensor_spec.TensorSpec( [self._expected_num_actions]) bandit_policy_type = () if policy_utilities.InfoFields.BANDIT_POLICY_TYPE in emit_policy_info: bandit_policy_type = ( policy_utilities.create_bandit_policy_type_tensor_spec(shape=[1])) if accepts_per_arm_features: # The features for the chosen arm is saved to policy_info. chosen_arm_features_info = ( policy_utilities.create_chosen_arm_features_info_spec( time_step_spec.observation)) info_spec = policy_utilities.PerArmPolicyInfo( predicted_rewards_mean=predicted_rewards_mean, bandit_policy_type=bandit_policy_type, chosen_arm_features=chosen_arm_features_info) else: info_spec = policy_utilities.PolicyInfo( predicted_rewards_mean=predicted_rewards_mean, bandit_policy_type=bandit_policy_type) self._accepts_per_arm_features = accepts_per_arm_features super(GreedyRewardPredictionPolicy, self).__init__( time_step_spec, action_spec, policy_state_spec=reward_network.state_spec, clip=False, info_spec=info_spec, emit_log_probability='log_probability' in emit_policy_info, observation_and_action_constraint_splitter=( observation_and_action_constraint_splitter), name=name)
def _distribution(self, time_step, policy_state): observation = time_step.observation if self.observation_and_action_constraint_splitter is not None: observation, _ = self.observation_and_action_constraint_splitter( observation) predictions, policy_state = self._reward_network( observation, time_step.step_type, policy_state) batch_size = tf.shape(predictions)[0] if isinstance(self._reward_network, heteroscedastic_q_network.HeteroscedasticQNetwork): predicted_reward_values = predictions.q_value_logits else: predicted_reward_values = predictions predicted_reward_values.shape.with_rank_at_least(2) predicted_reward_values.shape.with_rank_at_most(3) if predicted_reward_values.shape[ -1] is not None and predicted_reward_values.shape[ -1] != self._expected_num_actions: raise ValueError( 'The number of actions ({}) does not match the reward_network output' ' size ({}).'.format(self._expected_num_actions, predicted_reward_values.shape[1])) mask = constr.construct_mask_from_multiple_sources( time_step.observation, self._observation_and_action_constraint_splitter, self._constraints, self._expected_num_actions) # Argmax. if mask is not None: actions = policy_utilities.masked_argmax( predicted_reward_values, mask, output_type=self.action_spec.dtype) else: actions = tf.argmax( predicted_reward_values, axis=-1, output_type=self.action_spec.dtype) actions += self._action_offset bandit_policy_values = tf.fill([batch_size, 1], policy_utilities.BanditPolicyType.GREEDY) if self._accepts_per_arm_features: # Saving the features for the chosen action to the policy_info. def gather_observation(obs): return tf.gather(params=obs, indices=actions, batch_dims=1) chosen_arm_features = tf.nest.map_structure( gather_observation, observation[bandit_spec_utils.PER_ARM_FEATURE_KEY]) policy_info = policy_utilities.PerArmPolicyInfo( log_probability=tf.zeros([batch_size], tf.float32) if policy_utilities.InfoFields.LOG_PROBABILITY in self._emit_policy_info else (), predicted_rewards_mean=( predicted_reward_values if policy_utilities.InfoFields .PREDICTED_REWARDS_MEAN in self._emit_policy_info else ()), bandit_policy_type=(bandit_policy_values if policy_utilities.InfoFields.BANDIT_POLICY_TYPE in self._emit_policy_info else ()), chosen_arm_features=chosen_arm_features) else: policy_info = policy_utilities.PolicyInfo( log_probability=tf.zeros([batch_size], tf.float32) if policy_utilities.InfoFields.LOG_PROBABILITY in self._emit_policy_info else (), predicted_rewards_mean=( predicted_reward_values if policy_utilities.InfoFields .PREDICTED_REWARDS_MEAN in self._emit_policy_info else ()), bandit_policy_type=(bandit_policy_values if policy_utilities.InfoFields.BANDIT_POLICY_TYPE in self._emit_policy_info else ())) return policy_step.PolicyStep( tfp.distributions.Deterministic(loc=actions), policy_state, policy_info)
def __init__( self, time_step_spec: Optional[ts.TimeStep], action_spec: Optional[types.NestedBoundedTensorSpec], scalarizer: multi_objective_scalarizer.Scalarizer, objective_networks: Sequence[Network], observation_and_action_constraint_splitter: types.Splitter = None, accepts_per_arm_features: bool = False, emit_policy_info: Tuple[Text, ...] = (), name: Optional[Text] = None): """Builds a GreedyMultiObjectiveNeuralPolicy based on multiple networks. This policy takes an iterable of `tf_agents.Network`, each responsible for predicting a specific objective, along with a `Scalarizer` object to generate an action by maximizing the scalarized objective, i.e., the output of the `Scalarizer` applied to the multiple predicted objectives by the networks. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of `BoundedTensorSpec` representing the actions. scalarizer: A `tf_agents.bandits.multi_objective.multi_objective_scalarizer.Scalarizer` object that implements scalarization of multiple objectives into a single scalar reward. objective_networks: A Sequence of `tf_agents.network.Network` objects to be used by the policy. Each network will be called with call(observation, step_type) and is expected to provide a prediction for a specific objective for all actions. observation_and_action_constraint_splitter: A function used for masking valid/invalid actions with each state of the environment. The function takes in a full observation and returns a tuple consisting of 1) the part of the observation intended as input to the network and 2) the mask. The mask should be a 0-1 `Tensor` of shape `[batch_size, num_actions]`. This function should also work with a `TensorSpec` as input, and should output `TensorSpec` objects for the observation and mask. accepts_per_arm_features: (bool) Whether the policy accepts per-arm features. emit_policy_info: (tuple of strings) what side information we want to get as part of the policy info. Allowed values can be found in `policy_utilities.PolicyInfo`. name: The name of this policy. All variables in this module will fall under that name. Defaults to the class name. Raises: NotImplementedError: If `action_spec` contains more than one `BoundedTensorSpec` or the `BoundedTensorSpec` is not valid. NotImplementedError: If `action_spec` is not a `BoundedTensorSpec` of type int32 and shape (). ValueError: If `objective_networks` has fewer than two networks. ValueError: If `accepts_per_arm_features` is true but `time_step_spec` is None. """ policy_utilities.check_no_mask_with_arm_features( accepts_per_arm_features, observation_and_action_constraint_splitter) flat_action_spec = tf.nest.flatten(action_spec) if len(flat_action_spec) > 1: raise NotImplementedError( 'action_spec can only contain a single BoundedTensorSpec.') action_spec = flat_action_spec[0] if (not tensor_spec.is_bounded(action_spec) or not tensor_spec.is_discrete(action_spec) or action_spec.shape.rank > 1 or action_spec.shape.num_elements() != 1): raise NotImplementedError( 'action_spec must be a BoundedTensorSpec of type int32 and shape (). ' 'Found {}.'.format(action_spec)) self._expected_num_actions = action_spec.maximum - action_spec.minimum + 1 self._action_offset = action_spec.minimum policy_state_spec = [] for network in objective_networks: policy_state_spec.append(network.state_spec) network.create_variables() self._objective_networks = objective_networks self._scalarizer = scalarizer self._num_objectives = len(self._objective_networks) if self._num_objectives < 2: raise ValueError( 'Number of objectives should be at least two, but found to be {}' .format(self._num_objectives)) self._emit_policy_info = emit_policy_info predicted_rewards_mean = () if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN in emit_policy_info: predicted_rewards_mean = tensor_spec.TensorSpec( [self._num_objectives, self._expected_num_actions]) scalarized_predicted_rewards_mean = () if (policy_utilities.InfoFields. MULTIOBJECTIVE_SCALARIZED_PREDICTED_REWARDS_MEAN in emit_policy_info): scalarized_predicted_rewards_mean = tensor_spec.TensorSpec( [self._expected_num_actions]) bandit_policy_type = () if policy_utilities.InfoFields.BANDIT_POLICY_TYPE in emit_policy_info: bandit_policy_type = ( policy_utilities.create_bandit_policy_type_tensor_spec( shape=[1])) if accepts_per_arm_features: if time_step_spec is None: raise ValueError( 'time_step_spec should not be None for per-arm-features policies, ' 'but found to be.') # The features for the chosen arm is saved to policy_info. chosen_arm_features_info = ( policy_utilities.create_chosen_arm_features_info_spec( time_step_spec.observation)) info_spec = policy_utilities.PerArmPolicyInfo( predicted_rewards_mean=predicted_rewards_mean, multiobjective_scalarized_predicted_rewards_mean= scalarized_predicted_rewards_mean, bandit_policy_type=bandit_policy_type, chosen_arm_features=chosen_arm_features_info) else: info_spec = policy_utilities.PolicyInfo( predicted_rewards_mean=predicted_rewards_mean, multiobjective_scalarized_predicted_rewards_mean= scalarized_predicted_rewards_mean, bandit_policy_type=bandit_policy_type) self._accepts_per_arm_features = accepts_per_arm_features super(GreedyMultiObjectiveNeuralPolicy, self).__init__(time_step_spec, action_spec, policy_state_spec=policy_state_spec, clip=False, info_spec=info_spec, emit_log_probability='log_probability' in emit_policy_info, observation_and_action_constraint_splitter=( observation_and_action_constraint_splitter), name=name)
def _distribution( self, time_step: ts.TimeStep, policy_state: Sequence[types.TensorSpec] ) -> policy_step.PolicyStep: observation = time_step.observation if self.observation_and_action_constraint_splitter is not None: observation, _ = self.observation_and_action_constraint_splitter( observation) predicted_objective_values_tensor, policy_state = self._predict( observation, time_step.step_type, policy_state) scalarized_reward = scalarize_objectives( predicted_objective_values_tensor, self._scalarizer) # Preserve static batch size values when they are available. batch_size = (tf.compat.dimension_value(scalarized_reward.shape[0]) or tf.shape(scalarized_reward)[0]) mask = constraints.construct_mask_from_multiple_sources( time_step.observation, self._observation_and_action_constraint_splitter, (), self._expected_num_actions) # Argmax. if mask is not None: actions = policy_utilities.masked_argmax( scalarized_reward, mask, output_type=self.action_spec.dtype) else: actions = tf.argmax(scalarized_reward, axis=-1, output_type=self.action_spec.dtype) actions += self._action_offset bandit_policy_values = tf.fill( [batch_size, 1], policy_utilities.BanditPolicyType.GREEDY) if self._accepts_per_arm_features: # Saving the features for the chosen action to the policy_info. def gather_observation(obs): return tf.gather(params=obs, indices=actions, batch_dims=1) chosen_arm_features = tf.nest.map_structure( gather_observation, observation[bandit_spec_utils.PER_ARM_FEATURE_KEY]) policy_info = policy_utilities.PerArmPolicyInfo( log_probability=tf.zeros([batch_size], tf.float32) if policy_utilities.InfoFields.LOG_PROBABILITY in self._emit_policy_info else (), predicted_rewards_mean=( predicted_objective_values_tensor if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN in self._emit_policy_info else ()), multiobjective_scalarized_predicted_rewards_mean=( scalarized_reward if policy_utilities.InfoFields. MULTIOBJECTIVE_SCALARIZED_PREDICTED_REWARDS_MEAN in self._emit_policy_info else ()), bandit_policy_type=( bandit_policy_values if policy_utilities.InfoFields.BANDIT_POLICY_TYPE in self._emit_policy_info else ()), chosen_arm_features=chosen_arm_features) else: policy_info = policy_utilities.PolicyInfo( log_probability=tf.zeros([batch_size], tf.float32) if policy_utilities.InfoFields.LOG_PROBABILITY in self._emit_policy_info else (), predicted_rewards_mean=( predicted_objective_values_tensor if policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN in self._emit_policy_info else ()), multiobjective_scalarized_predicted_rewards_mean=( scalarized_reward if policy_utilities.InfoFields. MULTIOBJECTIVE_SCALARIZED_PREDICTED_REWARDS_MEAN in self._emit_policy_info else ()), bandit_policy_type=( bandit_policy_values if policy_utilities.InfoFields.BANDIT_POLICY_TYPE in self._emit_policy_info else ())) return policy_step.PolicyStep( tfp.distributions.Deterministic(loc=actions), policy_state, policy_info)