def actor( self, replay: reverb.Client, variable_source: acme.VariableSource, counter: counting.Counter, ): """The actor process.""" action_spec = self._environment_spec.actions observation_spec = self._environment_spec.observations # Create environment and behavior networks environment = self._environment_factory(False) agent_networks = self._network_factory(action_spec) # Create behavior network by adding some random dithering. behavior_network = snt.Sequential([ agent_networks.get('observation', tf.identity), agent_networks.get('policy'), networks.ClippedGaussian(self._sigma), ]) # Ensure network variables are created. tf2_utils.create_variables(behavior_network, [observation_spec]) variables = {'policy': behavior_network.variables} # Create the variable client responsible for keeping the actor up-to-date. variable_client = tf2_variable_utils.VariableClient( variable_source, variables, update_period=self._variable_update_period) # Make sure not to use a random policy after checkpoint restoration by # assigning variables before running the environment loop. variable_client.update_and_wait() # Component to add things into replay. adder = adders.NStepTransitionAdder(client=replay, n_step=self._n_step, discount=self._discount) # Create the agent. actor = actors.FeedForwardActor(behavior_network, adder=adder, variable_client=variable_client) # Create logger and counter; actors will not spam bigtable. counter = counting.Counter(counter, 'actor') logger = loggers.make_default_logger('actor', save_data=False, time_delta=self._log_every, steps_key='actor_steps') # Create the loop to connect environment and agent. return acme.EnvironmentLoop(environment, actor, counter, logger)
def make_policy( self, environment_spec: specs.EnvironmentSpec, sigma: float = 0.0, ) -> snt.Module: """Create a single network which evaluates the policy.""" # Stack the observation and policy networks. stack = [ self.observation_network, self.policy_network, ] # If a stochastic/non-greedy policy is requested, add Gaussian noise on # top to enable a simple form of exploration. # TODO(mwhoffman): Refactor this to remove it from the class. if sigma > 0.0: stack += [ network_utils.ClippedGaussian(sigma), network_utils.ClipToSpec(environment_spec.actions), ] # Return a network which sequentially evaluates everything in the stack. return snt.Sequential(stack)
def __init__(self, environment_spec: specs.EnvironmentSpec, policy_network: snt.Module, critic_network: snt.Module, observation_network: types.TensorTransformation = tf.identity, discount: float = 0.99, batch_size: int = 256, prefetch_size: int = 4, target_update_period: int = 100, min_replay_size: int = 1000, max_replay_size: int = 1000000, samples_per_insert: float = 32.0, n_step: int = 5, sigma: float = 0.3, clipping: bool = True, logger: loggers.Logger = None, counter: counting.Counter = None, checkpoint: bool = True, replay_table_name: str = adders.DEFAULT_PRIORITY_TABLE): """Initialize the agent. Args: environment_spec: description of the actions, observations, etc. policy_network: the online (optimized) policy. critic_network: the online critic. observation_network: optional network to transform the observations before they are fed into any network. discount: discount to use for TD updates. batch_size: batch size for updates. prefetch_size: size to prefetch from replay. target_update_period: number of learner steps to perform before updating the target networks. min_replay_size: minimum replay size before updating. max_replay_size: maximum replay size. samples_per_insert: number of samples to take from replay for every insert that is made. n_step: number of steps to squash into a single transition. sigma: standard deviation of zero-mean, Gaussian exploration noise. clipping: whether to clip gradients by global norm. logger: logger object to be used by learner. counter: counter object used to keep track of steps. checkpoint: boolean indicating whether to checkpoint the learner. replay_table_name: string indicating what name to give the replay table. """ # Create a replay server to add data to. This uses no limiter behavior in # order to allow the Agent interface to handle it. replay_table = reverb.Table( name=replay_table_name, sampler=reverb.selectors.Uniform(), remover=reverb.selectors.Fifo(), max_size=max_replay_size, rate_limiter=reverb.rate_limiters.MinSize(1), signature=adders.NStepTransitionAdder.signature(environment_spec)) self._server = reverb.Server([replay_table], port=None) # The adder is used to insert observations into replay. address = f'localhost:{self._server.port}' adder = adders.NStepTransitionAdder( priority_fns={replay_table_name: lambda x: 1.}, client=reverb.Client(address), n_step=n_step, discount=discount) # The dataset provides an interface to sample from replay. dataset = datasets.make_reverb_dataset( table=replay_table_name, client=reverb.TFClient(address), environment_spec=environment_spec, batch_size=batch_size, prefetch_size=prefetch_size, transition_adder=True) # Get observation and action specs. act_spec = environment_spec.actions obs_spec = environment_spec.observations emb_spec = tf2_utils.create_variables(observation_network, [obs_spec]) # pytype: disable=wrong-arg-types # Make sure observation network is a Sonnet Module. observation_network = tf2_utils.to_sonnet_module(observation_network) # Create target networks. target_policy_network = copy.deepcopy(policy_network) target_critic_network = copy.deepcopy(critic_network) target_observation_network = copy.deepcopy(observation_network) # Create the behavior policy. behavior_network = snt.Sequential([ observation_network, policy_network, networks.ClippedGaussian(sigma), networks.ClipToSpec(act_spec), ]) # Create variables. tf2_utils.create_variables(policy_network, [emb_spec]) tf2_utils.create_variables(critic_network, [emb_spec, act_spec]) tf2_utils.create_variables(target_policy_network, [emb_spec]) tf2_utils.create_variables(target_critic_network, [emb_spec, act_spec]) tf2_utils.create_variables(target_observation_network, [obs_spec]) # Create the actor which defines how we take actions. actor = actors.FeedForwardActor(behavior_network, adder=adder) # Create optimizers. policy_optimizer = snt.optimizers.Adam(learning_rate=1e-4) critic_optimizer = snt.optimizers.Adam(learning_rate=1e-4) # The learner updates the parameters (and initializes them). learner = learning.DDPGLearner( policy_network=policy_network, critic_network=critic_network, observation_network=observation_network, target_policy_network=target_policy_network, target_critic_network=target_critic_network, target_observation_network=target_observation_network, policy_optimizer=policy_optimizer, critic_optimizer=critic_optimizer, clipping=clipping, discount=discount, target_update_period=target_update_period, dataset=dataset, counter=counter, logger=logger, checkpoint=checkpoint, ) super().__init__(actor=actor, learner=learner, min_observations=max(batch_size, min_replay_size), observations_per_step=float(batch_size) / samples_per_insert)
def __init__(self, visualRadius, action_size, action_spec, exploration_sigma): super(ActorNetwork, self).__init__(name="commons-actor") self.policy_network = PolicyNetwork( visualRadius, action_size, action_spec) self.behavior_network = self.policy_network + snt.Sequential([networks.ClippedGaussian(exploration_sigma), networks.ClipToSpec(action_spec)])
# Create the target networks target_policy_network = copy.deepcopy(policy_network) target_critic_network = copy.deepcopy(critic_network) target_observation_network = copy.deepcopy(observation_network) # Get observation and action specs. act_spec = environment_spec.actions obs_spec = environment_spec.observations emb_spec = tf2_utils.create_variables(observation_network, [obs_spec]) # Create the behavior policy. behavior_network = snt.Sequential([ observation_network, policy_network, networks.ClippedGaussian(0.3), #sigma = 0.3 networks.ClipToSpec(act_spec), ]) # We must create the variables in the networks before passing them to learner. # Create variables. tf2_utils.create_variables(policy_network, [emb_spec]) tf2_utils.create_variables(critic_network, [emb_spec, act_spec]) tf2_utils.create_variables(target_policy_network, [emb_spec]) tf2_utils.create_variables(target_critic_network, [emb_spec, act_spec]) tf2_utils.create_variables(target_observation_network, [obs_spec]) actor = actors.FeedForwardActor(behavior_network, adder=adder) learner = d4pg.D4PGLearner(policy_network=policy_network, critic_network=critic_network,
def make_networks( environment_spec: mava_specs.MAEnvironmentSpec, policy_networks_layer_sizes: Union[Dict[str, Sequence], Sequence] = ( 256, 256, 256, ), critic_networks_layer_sizes: Union[Dict[str, Sequence], Sequence] = (512, 512, 256), shared_weights: bool = True, sigma: float = 0.3, ) -> Mapping[str, types.TensorTransformation]: """Creates networks used by the agents.""" specs = environment_spec.get_agent_specs() # Create agent_type specs if shared_weights: type_specs = {key.split("_")[0]: specs[key] for key in specs.keys()} specs = type_specs if isinstance(policy_networks_layer_sizes, Sequence): policy_networks_layer_sizes = { key: policy_networks_layer_sizes for key in specs.keys() } if isinstance(critic_networks_layer_sizes, Sequence): critic_networks_layer_sizes = { key: critic_networks_layer_sizes for key in specs.keys() } observation_networks = {} policy_networks = {} critic_networks = {} for key in specs.keys(): # Get total number of action dimensions from action spec. num_dimensions = np.prod(specs[key].actions.shape, dtype=int) # Create the shared observation network; here simply a state-less operation. observation_network = tf2_utils.to_sonnet_module(tf.identity) # Create the policy network. policy_network = snt.Sequential( [ networks.LayerNormMLP( policy_networks_layer_sizes[key], activate_final=True ), networks.NearZeroInitializedLinear(num_dimensions), networks.TanhToSpec(specs[key].actions), networks.ClippedGaussian(sigma), networks.ClipToSpec(specs[key].actions), ] ) # Create the critic network. critic_network = snt.Sequential( [ # The multiplexer concatenates the observations/actions. networks.CriticMultiplexer(), networks.LayerNormMLP( critic_networks_layer_sizes[key], activate_final=False ), snt.Linear(1), ] ) observation_networks[key] = observation_network policy_networks[key] = policy_network critic_networks[key] = critic_network return { "policies": policy_networks, "critics": critic_networks, "observations": observation_networks, }
def make_default_networks( environment_spec: mava_specs.MAEnvironmentSpec, policy_networks_layer_sizes: Union[Dict[str, Sequence], Sequence] = (256, 256, 256), critic_networks_layer_sizes: Union[Dict[str, Sequence], Sequence] = (512, 512, 256), shared_weights: bool = True, sigma: float = 0.3, archecture_type: ArchitectureType = ArchitectureType.feedforward, ) -> Mapping[str, types.TensorTransformation]: """Default networks for maddpg. Args: environment_spec (mava_specs.MAEnvironmentSpec): description of the action and observation spaces etc. for each agent in the system. policy_networks_layer_sizes (Union[Dict[str, Sequence], Sequence], optional): size of policy networks. Defaults to (256, 256, 256). critic_networks_layer_sizes (Union[Dict[str, Sequence], Sequence], optional): size of critic networks. Defaults to (512, 512, 256). shared_weights (bool, optional): whether agents should share weights or not. Defaults to True. sigma (float, optional): hyperparameters used to add Gaussian noise for simple exploration. Defaults to 0.3. archecture_type (ArchitectureType, optional): archecture used for agent networks. Can be feedforward or recurrent. Defaults to ArchitectureType.feedforward. Returns: Mapping[str, types.TensorTransformation]: returned agent networks. """ # Set Policy function and layer size if archecture_type == ArchitectureType.feedforward: policy_network_func = snt.Sequential elif archecture_type == ArchitectureType.recurrent: policy_networks_layer_sizes = (128, 128) policy_network_func = snt.DeepRNN specs = environment_spec.get_agent_specs() # Create agent_type specs if shared_weights: type_specs = {key.split("_")[0]: specs[key] for key in specs.keys()} specs = type_specs if isinstance(policy_networks_layer_sizes, Sequence): policy_networks_layer_sizes = { key: policy_networks_layer_sizes for key in specs.keys() } if isinstance(critic_networks_layer_sizes, Sequence): critic_networks_layer_sizes = { key: critic_networks_layer_sizes for key in specs.keys() } observation_networks = {} policy_networks = {} critic_networks = {} for key in specs.keys(): # TODO (dries): Make specs[key].actions # return a list of specs for hybrid action space # Get total number of action dimensions from action spec. agent_act_spec = specs[key].actions if type(specs[key].actions) == DiscreteArray: num_actions = agent_act_spec.num_values minimum = [-1.0] * num_actions maximum = [1.0] * num_actions agent_act_spec = BoundedArray( shape=(num_actions, ), minimum=minimum, maximum=maximum, dtype="float32", name="actions", ) # Get total number of action dimensions from action spec. num_dimensions = np.prod(agent_act_spec.shape, dtype=int) # An optional network to process observations observation_network = tf2_utils.to_sonnet_module(tf.identity) # Create the policy network. if archecture_type == ArchitectureType.feedforward: policy_network = [ networks.LayerNormMLP(policy_networks_layer_sizes[key], activate_final=True), ] elif archecture_type == ArchitectureType.recurrent: policy_network = [ networks.LayerNormMLP(policy_networks_layer_sizes[key][:-1], activate_final=True), snt.LSTM(policy_networks_layer_sizes[key][-1]), ] policy_network += [ networks.NearZeroInitializedLinear(num_dimensions), networks.TanhToSpec(agent_act_spec), ] # Add Gaussian noise for simple exploration. if sigma and sigma > 0.0: policy_network += [ networks.ClippedGaussian(sigma), networks.ClipToSpec(agent_act_spec), ] policy_network = policy_network_func(policy_network) # Create the critic network. critic_network = snt.Sequential([ # The multiplexer concatenates the observations/actions. networks.CriticMultiplexer(), networks.LayerNormMLP(list(critic_networks_layer_sizes[key]) + [1], activate_final=False), ]) observation_networks[key] = observation_network policy_networks[key] = policy_network critic_networks[key] = critic_network return { "policies": policy_networks, "critics": critic_networks, "observations": observation_networks, }