TD3 (#338)

rl_coach/agents/ddpg_agent.py

@@ -41,14 +41,15 @@ class DDPGCriticNetworkParameters(NetworkParameters):
         self.middleware_parameters = FCMiddlewareParameters()
         self.heads_parameters = [DDPGVHeadParameters()]
         self.optimizer_type = 'Adam'
-        self.adam_optimizer_beta2 = 0.999
-        self.optimizer_epsilon = 1e-8
         self.batch_size = 64
         self.async_training = False
         self.learning_rate = 0.001
+        self.adam_optimizer_beta2 = 0.999
+        self.optimizer_epsilon = 1e-8
         self.create_target_network = True
         self.shared_optimizer = True
         self.scale_down_gradients_by_number_of_workers_for_sync_training = False
+        # self.l2_regularization = 1e-2
 
 
 class DDPGActorNetworkParameters(NetworkParameters):
@@ -58,9 +59,9 @@ class DDPGActorNetworkParameters(NetworkParameters):
         self.middleware_parameters = FCMiddlewareParameters(batchnorm=True)
         self.heads_parameters = [DDPGActorHeadParameters()]
         self.optimizer_type = 'Adam'
+        self.batch_size = 64
         self.adam_optimizer_beta2 = 0.999
         self.optimizer_epsilon = 1e-8
-        self.batch_size = 64
         self.async_training = False
         self.learning_rate = 0.0001
         self.create_target_network = True
@@ -217,4 +218,4 @@ class DDPGAgent(ActorCriticAgent):
         action_info = ActionInfo(action=action,
                                  action_value=q_value)
 
-        return action_info
+        return action_info

rl_coach/agents/ddqn_bcq_agent.py

@@ -90,7 +90,7 @@ class DDQNBCQAgent(DQNAgent):
                 if self.ap.algorithm.action_drop_method_parameters.use_state_embedding_instead_of_state:
                     return self.networks['reward_model'].online_network.predict(
                         states,
-                        outputs=[self.networks['reward_model'].online_network.state_embedding])
+                        outputs=[self.networks['reward_model'].online_network.state_embedding[0]])
                 else:
                     return states['observation']
             self.embedding = to_embedding
@@ -189,7 +189,7 @@ class DDQNBCQAgent(DQNAgent):
             if self.ap.algorithm.action_drop_method_parameters.use_state_embedding_instead_of_state:
                 self.knn_trees = [AnnoyDictionary(
                     dict_size=knn_size,
-                    key_width=int(self.networks['reward_model'].online_network.state_embedding.shape[-1]),
+                    key_width=int(self.networks['reward_model'].online_network.state_embedding[0].shape[-1]),
                     batch_size=knn_size)
                     for _ in range(len(self.spaces.action.actions))]
             else:

rl_coach/agents/nec_agent.py

@@ -194,7 +194,7 @@ class NECAgent(ValueOptimizationAgent):
         )
         if self.phase != RunPhase.TEST:
             # store the state embedding for inserting it to the DND later
-            self.current_episode_state_embeddings.append(embedding.squeeze())
+            self.current_episode_state_embeddings.append(embedding[0].squeeze())
         actions_q_values = actions_q_values[0][0]
         return actions_q_values
 

rl_coach/agents/td3_agent.py

@@ -0,0 +1,223 @@
+#
+# Copyright (c) 2019 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import copy
+from typing import Union
+from collections import OrderedDict
+
+import numpy as np
+
+from rl_coach.agents.agent import Agent
+from rl_coach.agents.ddpg_agent import DDPGAgent
+from rl_coach.architectures.embedder_parameters import InputEmbedderParameters
+from rl_coach.architectures.head_parameters import DDPGActorHeadParameters, TD3VHeadParameters
+from rl_coach.architectures.middleware_parameters import FCMiddlewareParameters
+from rl_coach.base_parameters import NetworkParameters, AlgorithmParameters, \
+    AgentParameters, EmbedderScheme
+from rl_coach.core_types import ActionInfo, TrainingSteps, Transition
+from rl_coach.exploration_policies.additive_noise import AdditiveNoiseParameters
+from rl_coach.memories.episodic.episodic_experience_replay import EpisodicExperienceReplayParameters
+from rl_coach.spaces import BoxActionSpace, GoalsSpace
+
+
+class TD3CriticNetworkParameters(NetworkParameters):
+    def __init__(self, num_q_networks):
+        super().__init__()
+        self.input_embedders_parameters = {'observation': InputEmbedderParameters(),
+                                            'action': InputEmbedderParameters(scheme=EmbedderScheme.Shallow)}
+        self.middleware_parameters = FCMiddlewareParameters(num_streams=num_q_networks)
+        self.heads_parameters = [TD3VHeadParameters()]
+        self.optimizer_type = 'Adam'
+        self.adam_optimizer_beta2 = 0.999
+        self.optimizer_epsilon = 1e-8
+        self.batch_size = 100
+        self.async_training = False
+        self.learning_rate = 0.001
+        self.create_target_network = True
+        self.shared_optimizer = True
+        self.scale_down_gradients_by_number_of_workers_for_sync_training = False
+
+
+class TD3ActorNetworkParameters(NetworkParameters):
+    def __init__(self):
+        super().__init__()
+        self.input_embedders_parameters = {'observation': InputEmbedderParameters()}
+        self.middleware_parameters = FCMiddlewareParameters()
+        self.heads_parameters = [DDPGActorHeadParameters(batchnorm=False)]
+        self.optimizer_type = 'Adam'
+        self.adam_optimizer_beta2 = 0.999
+        self.optimizer_epsilon = 1e-8
+        self.batch_size = 100
+        self.async_training = False
+        self.learning_rate = 0.001
+        self.create_target_network = True
+        self.shared_optimizer = True
+        self.scale_down_gradients_by_number_of_workers_for_sync_training = False
+
+
+class TD3AlgorithmParameters(AlgorithmParameters):
+    """
+    :param num_steps_between_copying_online_weights_to_target: (StepMethod)
+        The number of steps between copying the online network weights to the target network weights.
+
+    :param rate_for_copying_weights_to_target: (float)
+        When copying the online network weights to the target network weights, a soft update will be used, which
+        weight the new online network weights by rate_for_copying_weights_to_target
+
+    :param num_consecutive_playing_steps: (StepMethod)
+        The number of consecutive steps to act between every two training iterations
+
+    :param use_target_network_for_evaluation: (bool)
+        If set to True, the target network will be used for predicting the actions when choosing actions to act.
+        Since the target network weights change more slowly, the predicted actions will be more consistent.
+
+    :param action_penalty: (float)
+        The amount by which to penalize the network on high action feature (pre-activation) values.
+        This can prevent the actions features from saturating the TanH activation function, and therefore prevent the
+        gradients from becoming very low.
+
+    :param clip_critic_targets: (Tuple[float, float] or None)
+        The range to clip the critic target to in order to prevent overestimation of the action values.
+
+    :param use_non_zero_discount_for_terminal_states: (bool)
+        If set to True, the discount factor will be used for terminal states to bootstrap the next predicted state
+        values. If set to False, the terminal states reward will be taken as the target return for the network.
+    """
+    def __init__(self):
+        super().__init__()
+        self.rate_for_copying_weights_to_target = 0.005
+        self.use_target_network_for_evaluation = False
+        self.action_penalty = 0
+        self.clip_critic_targets = None  # expected to be a tuple of the form (min_clip_value, max_clip_value) or None
+        self.use_non_zero_discount_for_terminal_states = False
+        self.act_for_full_episodes = True
+        self.update_policy_every_x_episode_steps = 2
+        self.num_steps_between_copying_online_weights_to_target = TrainingSteps(self.update_policy_every_x_episode_steps)
+        self.policy_noise = 0.2
+        self.noise_clipping = 0.5
+        self.num_q_networks = 2
+
+
+class TD3AgentExplorationParameters(AdditiveNoiseParameters):
+    def __init__(self):
+        super().__init__()
+        self.noise_as_percentage_from_action_space = False
+
+
+class TD3AgentParameters(AgentParameters):
+    def __init__(self):
+        td3_algorithm_params = TD3AlgorithmParameters()
+        super().__init__(algorithm=td3_algorithm_params,
+                         exploration=TD3AgentExplorationParameters(),
+                         memory=EpisodicExperienceReplayParameters(),
+                         networks=OrderedDict([("actor", TD3ActorNetworkParameters()),
+                                               ("critic",
+                                                TD3CriticNetworkParameters(td3_algorithm_params.num_q_networks))]))
+
+    @property
+    def path(self):
+        return 'rl_coach.agents.td3_agent:TD3Agent'
+
+
+# Twin Delayed DDPG - https://arxiv.org/pdf/1802.09477.pdf
+class TD3Agent(DDPGAgent):
+    def __init__(self, agent_parameters, parent: Union['LevelManager', 'CompositeAgent']=None):
+        super().__init__(agent_parameters, parent)
+
+        self.q_values = self.register_signal("Q")
+        self.TD_targets_signal = self.register_signal("TD targets")
+        self.action_signal = self.register_signal("actions")
+
+    def learn_from_batch(self, batch):
+        actor = self.networks['actor']
+        critic = self.networks['critic']
+
+        actor_keys = self.ap.network_wrappers['actor'].input_embedders_parameters.keys()
+        critic_keys = self.ap.network_wrappers['critic'].input_embedders_parameters.keys()
+
+        # TD error = r + discount*max(q_st_plus_1) - q_st
+        next_actions, actions_mean = actor.parallel_prediction([
+            (actor.target_network, batch.next_states(actor_keys)),
+            (actor.online_network, batch.states(actor_keys))
+        ])
+
+        # add noise to the next_actions
+        noise = np.random.normal(0, self.ap.algorithm.policy_noise, next_actions.shape).clip(
+            -self.ap.algorithm.noise_clipping, self.ap.algorithm.noise_clipping)
+        next_actions = self.spaces.action.clip_action_to_space(next_actions + noise)
+
+        critic_inputs = copy.copy(batch.next_states(critic_keys))
+        critic_inputs['action'] = next_actions
+        q_st_plus_1 = critic.target_network.predict(critic_inputs)[2]  # output #2 is the min (Q1, Q2)
+
+        # calculate the bootstrapped TD targets while discounting terminal states according to
+        # use_non_zero_discount_for_terminal_states
+        if self.ap.algorithm.use_non_zero_discount_for_terminal_states:
+            TD_targets = batch.rewards(expand_dims=True) + self.ap.algorithm.discount * q_st_plus_1
+        else:
+            TD_targets = batch.rewards(expand_dims=True) + \
+                         (1.0 - batch.game_overs(expand_dims=True)) * self.ap.algorithm.discount * q_st_plus_1
+
+        # clip the TD targets to prevent overestimation errors
+        if self.ap.algorithm.clip_critic_targets:
+            TD_targets = np.clip(TD_targets, *self.ap.algorithm.clip_critic_targets)
+
+        self.TD_targets_signal.add_sample(TD_targets)
+
+        # train the critic
+        critic_inputs = copy.copy(batch.states(critic_keys))
+        critic_inputs['action'] = batch.actions(len(batch.actions().shape) == 1)
+        result = critic.train_and_sync_networks(critic_inputs, TD_targets)
+        total_loss, losses, unclipped_grads = result[:3]
+
+        if self.training_iteration % self.ap.algorithm.update_policy_every_x_episode_steps == 0:
+            # get the gradients of output #3 (=mean of Q1 network) w.r.t the action
+            critic_inputs = copy.copy(batch.states(critic_keys))
+            critic_inputs['action'] = actions_mean
+            action_gradients = critic.online_network.predict(critic_inputs,
+                                                             outputs=critic.online_network.gradients_wrt_inputs[3]['action'])
+
+            # apply the gradients from the critic to the actor
+            initial_feed_dict = {actor.online_network.gradients_weights_ph[0]: -action_gradients}
+            gradients = actor.online_network.predict(batch.states(actor_keys),
+                                                     outputs=actor.online_network.weighted_gradients[0],
+                                                     initial_feed_dict=initial_feed_dict)
+
+            if actor.has_global:
+                actor.apply_gradients_to_global_network(gradients)
+                actor.update_online_network()
+            else:
+                actor.apply_gradients_to_online_network(gradients)
+
+        return total_loss, losses, unclipped_grads
+
+    def train(self):
+        self.ap.algorithm.num_consecutive_training_steps = self.current_episode_steps_counter
+        return Agent.train(self)
+
+    def update_transition_before_adding_to_replay_buffer(self, transition: Transition) -> Transition:
+        """
+        Allows agents to update the transition just before adding it to the replay buffer.
+        Can be useful for agents that want to tweak the reward, termination signal, etc.
+
+        :param transition: the transition to update
+        :return: the updated transition
+        """
+        transition.game_over = False if self.current_episode_steps_counter ==\
+                                        self.parent_level_manager.environment.env._max_episode_steps\
+            else transition.game_over
+
+        return transition

rl_coach/architectures/head_parameters.py

@@ -221,3 +221,14 @@ class SACQHeadParameters(HeadParameters):
         super().__init__(parameterized_class_name='SACQHead', activation_function=activation_function, name=name,
                          dense_layer=dense_layer)
         self.network_layers_sizes = layers_sizes
+
+
+class TD3VHeadParameters(HeadParameters):
+    def __init__(self, activation_function: str ='relu', name: str='td3_v_head_params',
+                 num_output_head_copies: int = 1, rescale_gradient_from_head_by_factor: float = 1.0,
+                 loss_weight: float = 1.0, dense_layer=None, initializer='xavier'):
+        super().__init__(parameterized_class_name="TD3VHead", activation_function=activation_function, name=name,
+                         dense_layer=dense_layer, num_output_head_copies=num_output_head_copies,
+                         rescale_gradient_from_head_by_factor=rescale_gradient_from_head_by_factor,
+                         loss_weight=loss_weight)
+        self.initializer = initializer

rl_coach/architectures/middleware_parameters.py

@@ -41,10 +41,11 @@ class FCMiddlewareParameters(MiddlewareParameters):
     def __init__(self, activation_function='relu',
                  scheme: Union[List, MiddlewareScheme] = MiddlewareScheme.Medium,
                  batchnorm: bool = False, dropout_rate: float = 0.0,
-                 name="middleware_fc_embedder", dense_layer=None, is_training=False):
+                 name="middleware_fc_embedder", dense_layer=None, is_training=False, num_streams=1):
         super().__init__(parameterized_class_name="FCMiddleware", activation_function=activation_function,
                          scheme=scheme, batchnorm=batchnorm, dropout_rate=dropout_rate, name=name, dense_layer=dense_layer,
                          is_training=is_training)
+        self.num_streams = num_streams
 
 
 class LSTMMiddlewareParameters(MiddlewareParameters):

rl_coach/architectures/tensorflow_components/architecture.py

@@ -203,7 +203,6 @@ class TensorFlowArchitecture(Architecture):
             self._create_gradient_accumulators()
 
         # gradients of the outputs w.r.t. the inputs
-        # at the moment, this is only used by ddpg
         self.gradients_wrt_inputs = [{name: tf.gradients(output, input_ph) for name, input_ph in
                                       self.inputs.items()} for output in self.outputs]
         self.gradients_weights_ph = [tf.placeholder('float32', self.outputs[i].shape, 'output_gradient_weights')

rl_coach/architectures/tensorflow_components/heads/__init__.py

@@ -16,6 +16,7 @@ from .sac_head import SACPolicyHead
 from .sac_q_head import SACQHead
 from .classification_head import ClassificationHead
 from .cil_head import RegressionHead
+from .td3_v_head import TD3VHead
 from .ddpg_v_head import DDPGVHead
 
 __all__ = [
@@ -37,5 +38,6 @@ __all__ = [
     'SACQHead',
     'ClassificationHead',
     'RegressionHead',
+    'TD3VHead'
     'DDPGVHead'
 ]

rl_coach/architectures/tensorflow_components/heads/head.py

@@ -13,7 +13,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 #
-from typing import Type
 
 import numpy as np
 import tensorflow as tf
@@ -22,7 +21,7 @@ from rl_coach.architectures.tensorflow_components.layers import Dense, convert_l
 from rl_coach.base_parameters import AgentParameters
 from rl_coach.spaces import SpacesDefinition
 from rl_coach.utils import force_list
-
+from rl_coach.architectures.tensorflow_components.utils import squeeze_tensor
 
 # Used to initialize weights for policy and value output layers
 def normalized_columns_initializer(std=1.0):
@@ -72,8 +71,9 @@ class Head(object):
         :param input_layer: the input to the graph
         :return: the output of the last layer and the target placeholder
         """
+
         with tf.variable_scope(self.get_name(), initializer=tf.contrib.layers.xavier_initializer()):
-            self._build_module(input_layer)
+            self._build_module(squeeze_tensor(input_layer))
 
             self.output = force_list(self.output)
             self.target = force_list(self.target)

rl_coach/architectures/tensorflow_components/heads/td3_v_head.py

@@ -0,0 +1,67 @@
+#
+# Copyright (c) 2019 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import tensorflow as tf
+
+from rl_coach.architectures.tensorflow_components.layers import Dense
+from rl_coach.architectures.tensorflow_components.heads.head import Head, normalized_columns_initializer
+from rl_coach.base_parameters import AgentParameters
+from rl_coach.core_types import VStateValue
+from rl_coach.spaces import SpacesDefinition
+
+
+class TD3VHead(Head):
+    def __init__(self, agent_parameters: AgentParameters, spaces: SpacesDefinition, network_name: str,
+                 head_idx: int = 0, loss_weight: float = 1., is_local: bool = True, activation_function: str='relu',
+                 dense_layer=Dense, initializer='xavier'):
+        super().__init__(agent_parameters, spaces, network_name, head_idx, loss_weight, is_local, activation_function,
+                         dense_layer=dense_layer)
+        self.name = 'td3_v_values_head'
+        self.return_type = VStateValue
+        self.loss_type = []
+        self.initializer = initializer
+        self.loss = []
+        self.output = []
+
+    def _build_module(self, input_layer):
+        # Standard V Network
+        q_outputs = []
+        self.target = tf.placeholder(tf.float32, shape=(None, 1), name="q_networks_min_placeholder")
+
+        for i in range(input_layer.shape[0]): # assuming that the actual size is 2, as there are two critic networks
+            if self.initializer == 'normalized_columns':
+                q_outputs.append(self.dense_layer(1)(input_layer[i], name='q_output_{}'.format(i + 1),
+                                                     kernel_initializer=normalized_columns_initializer(1.0)))
+            elif self.initializer == 'xavier' or self.initializer is None:
+                q_outputs.append(self.dense_layer(1)(input_layer[i], name='q_output_{}'.format(i + 1)))
+
+            self.output.append(q_outputs[i])
+            self.loss.append(tf.reduce_mean((self.target-q_outputs[i])**2))
+
+        self.output.append(tf.reduce_min(q_outputs, axis=0))
+        self.output.append(tf.reduce_mean(self.output[0]))
+        self.loss = sum(self.loss)
+        tf.losses.add_loss(self.loss)
+
+    def __str__(self):
+        result = [
+            "Q1 Action-Value Stream",
+            "\tDense (num outputs = 1)",
+            "Q2 Action-Value Stream",
+            "\tDense (num outputs = 1)",
+            "Min (Q1, Q2)"
+        ]
+        return '\n'.join(result)

rl_coach/architectures/tensorflow_components/middlewares/fc_middleware.py

@@ -28,23 +28,28 @@ class FCMiddleware(Middleware):
     def __init__(self, activation_function=tf.nn.relu,
                  scheme: MiddlewareScheme = MiddlewareScheme.Medium,
                  batchnorm: bool = False, dropout_rate: float = 0.0,
-                 name="middleware_fc_embedder", dense_layer=Dense, is_training=False):
+                 name="middleware_fc_embedder", dense_layer=Dense, is_training=False, num_streams: int = 1):
         super().__init__(activation_function=activation_function, batchnorm=batchnorm,
                          dropout_rate=dropout_rate, scheme=scheme, name=name, dense_layer=dense_layer,
                          is_training=is_training)
         self.return_type = Middleware_FC_Embedding
-        self.layers = []
+
+        assert(isinstance(num_streams, int) and num_streams >= 1)
+        self.num_streams = num_streams
 
     def _build_module(self):
-        self.layers.append(self.input)
+        self.output = []
 
-        for idx, layer_params in enumerate(self.layers_params):
-            self.layers.extend(force_list(
-                layer_params(self.layers[-1], name='{}_{}'.format(layer_params.__class__.__name__, idx),
-                             is_training=self.is_training)
-            ))
+        for stream_idx in range(self.num_streams):
+            layers = [self.input]
 
-        self.output = self.layers[-1]
+            for idx, layer_params in enumerate(self.layers_params):
+                layers.extend(force_list(
+                    layer_params(layers[-1], name='{}_{}'.format(layer_params.__class__.__name__,
+                                                                 idx + stream_idx * len(self.layers_params)),
+                                 is_training=self.is_training)
+                ))
+            self.output.append((layers[-1]))
 
     @property
     def schemes(self):
@@ -72,3 +77,15 @@ class FCMiddleware(Middleware):
                 ]
         }
 
+    def __str__(self):
+        stream = [str(l) for l in self.layers_params]
+        if self.layers_params:
+            if self.num_streams > 1:
+                stream = [''] + ['\t' + l for l in stream]
+                result = stream * self.num_streams
+                result[0::len(stream)] = ['Stream {}'.format(i) for i in range(self.num_streams)]
+            else:
+                result = stream
+            return '\n'.join(result)
+        else:
+            return 'No layers'

rl_coach/architectures/tensorflow_components/utils.py

@@ -38,3 +38,10 @@ def get_activation_function(activation_function_string: str):
         "Activation function must be one of the following {}. instead it was: {}" \
             .format(activation_functions.keys(), activation_function_string)
     return activation_functions[activation_function_string]
+
+
+def squeeze_tensor(tensor):
+    if tensor.shape[0] == 1:
+        return tensor[0]
+    else:
+        return tensor

rl_coach/exploration_policies/additive_noise.py

@@ -17,7 +17,6 @@
 from typing import List
 
 import numpy as np
-import scipy.stats
 
 from rl_coach.core_types import RunPhase, ActionType
 from rl_coach.exploration_policies.exploration_policy import ContinuousActionExplorationPolicy, ExplorationParameters
@@ -31,8 +30,9 @@ from rl_coach.spaces import ActionSpace, BoxActionSpace
 class AdditiveNoiseParameters(ExplorationParameters):
     def __init__(self):
         super().__init__()
-        self.noise_percentage_schedule = LinearSchedule(0.1, 0.1, 50000)
-        self.evaluation_noise_percentage = 0.05
+        self.noise_schedule = LinearSchedule(0.1, 0.1, 50000)
+        self.evaluation_noise = 0.05
+        self.noise_as_percentage_from_action_space = True
 
     @property
     def path(self):
@@ -48,17 +48,19 @@ class AdditiveNoise(ContinuousActionExplorationPolicy):
     2. Specified by the agents action. In case the agents action is a list with 2 values, the 1st one is assumed to
     be the mean of the action, and 2nd is assumed to be its standard deviation.
     """
-    def __init__(self, action_space: ActionSpace, noise_percentage_schedule: Schedule,
-                 evaluation_noise_percentage: float):
+    def __init__(self, action_space: ActionSpace, noise_schedule: Schedule,
+                 evaluation_noise: float, noise_as_percentage_from_action_space: bool = True):
         """
         :param action_space: the action space used by the environment
-        :param noise_percentage_schedule: the schedule for the noise variance percentage relative to the absolute range
-                                          of the action space
-        :param evaluation_noise_percentage: the noise variance percentage that will be used during evaluation phases
+        :param noise_schedule: the schedule for the noise
+        :param evaluation_noise: the noise variance that will be used during evaluation phases
+        :param noise_as_percentage_from_action_space: a bool deciding whether the noise is absolute or as a percentage
+                                                      from the action space
         """
         super().__init__(action_space)
-        self.noise_percentage_schedule = noise_percentage_schedule
-        self.evaluation_noise_percentage = evaluation_noise_percentage
+        self.noise_schedule = noise_schedule
+        self.evaluation_noise = evaluation_noise
+        self.noise_as_percentage_from_action_space = noise_as_percentage_from_action_space
 
         if not isinstance(action_space, BoxActionSpace):
             raise ValueError("Additive noise exploration works only for continuous controls."
@@ -68,19 +70,20 @@ class AdditiveNoise(ContinuousActionExplorationPolicy):
                 or not np.all(-np.inf < action_space.low) or not np.all(action_space.low < np.inf):
             raise ValueError("Additive noise exploration requires bounded actions")
 
-        # TODO: allow working with unbounded actions by defining the noise in terms of range and not percentage
-
     def get_action(self, action_values: List[ActionType]) -> ActionType:
         # TODO-potential-bug consider separating internally defined stdev and externally defined stdev into 2 policies
 
-        # set the current noise percentage
+        # set the current noise
         if self.phase == RunPhase.TEST:
-            current_noise_precentage = self.evaluation_noise_percentage
+            current_noise = self.evaluation_noise
         else:
-            current_noise_precentage = self.noise_percentage_schedule.current_value
+            current_noise = self.noise_schedule.current_value
 
         # scale the noise to the action space range
-        action_values_std = current_noise_precentage * (self.action_space.high - self.action_space.low)
+        if self.noise_as_percentage_from_action_space:
+            action_values_std = current_noise * (self.action_space.high - self.action_space.low)
+        else:
+            action_values_std = current_noise
 
         # extract the mean values
         if isinstance(action_values, list):
@@ -92,15 +95,18 @@ class AdditiveNoise(ContinuousActionExplorationPolicy):
 
         # step the noise schedule
         if self.phase is not RunPhase.TEST:
-            self.noise_percentage_schedule.step()
+            self.noise_schedule.step()
             # the second element of the list is assumed to be the standard deviation
             if isinstance(action_values, list) and len(action_values) > 1:
                 action_values_std = action_values[1].squeeze()
 
         # add noise to the action means
-        action = np.random.normal(action_values_mean, action_values_std)
+        if self.phase is not RunPhase.TEST:
+            action = np.random.normal(action_values_mean, action_values_std)
+        else:
+            action = action_values_mean
 
-        return action
+        return np.atleast_1d(action)
 
     def get_control_param(self):
-        return np.ones(self.action_space.shape)*self.noise_percentage_schedule.current_value
+        return np.ones(self.action_space.shape)*self.noise_schedule.current_value

rl_coach/exploration_policies/e_greedy.py

@@ -32,7 +32,7 @@ class EGreedyParameters(ExplorationParameters):
         self.epsilon_schedule = LinearSchedule(0.5, 0.01, 50000)
         self.evaluation_epsilon = 0.05
         self.continuous_exploration_policy_parameters = AdditiveNoiseParameters()
-        self.continuous_exploration_policy_parameters.noise_percentage_schedule = LinearSchedule(0.1, 0.1, 50000)
+        self.continuous_exploration_policy_parameters.noise_schedule = LinearSchedule(0.1, 0.1, 50000)
         # for continuous control -
         # (see http://www.cs.ubc.ca/~van/papers/2017-TOG-deepLoco/2017-TOG-deepLoco.pdf)
 

rl_coach/exploration_policies/truncated_normal.py

@@ -28,10 +28,11 @@ from rl_coach.spaces import ActionSpace, BoxActionSpace
 class TruncatedNormalParameters(ExplorationParameters):
     def __init__(self):
         super().__init__()
-        self.noise_percentage_schedule = LinearSchedule(0.1, 0.1, 50000)
-        self.evaluation_noise_percentage = 0.05
+        self.noise_schedule = LinearSchedule(0.1, 0.1, 50000)
+        self.evaluation_noise = 0.05
         self.clip_low = 0
         self.clip_high = 1
+        self.noise_as_percentage_from_action_space = True
 
     @property
     def path(self):
@@ -49,17 +50,20 @@ class TruncatedNormal(ContinuousActionExplorationPolicy):
     When the sampled action is outside of the action bounds given by the user, it is sampled again and again, until it
     is within the bounds.
     """
-    def __init__(self, action_space: ActionSpace, noise_percentage_schedule: Schedule,
-                 evaluation_noise_percentage: float, clip_low: float, clip_high: float):
+    def __init__(self, action_space: ActionSpace, noise_schedule: Schedule,
+                 evaluation_noise: float, clip_low: float, clip_high: float,
+                 noise_as_percentage_from_action_space: bool = True):
         """
         :param action_space: the action space used by the environment
-        :param noise_percentage_schedule: the schedule for the noise variance percentage relative to the absolute range
-                                          of the action space
-        :param evaluation_noise_percentage: the noise variance percentage that will be used during evaluation phases
+        :param noise_schedule: the schedule for the noise variance
+        :param evaluation_noise: the noise variance that will be used during evaluation phases
+        :param noise_as_percentage_from_action_space: whether to consider the noise as a percentage of the action space
+                                                        or absolute value
         """
         super().__init__(action_space)
-        self.noise_percentage_schedule = noise_percentage_schedule
-        self.evaluation_noise_percentage = evaluation_noise_percentage
+        self.noise_schedule = noise_schedule
+        self.evaluation_noise = evaluation_noise
+        self.noise_as_percentage_from_action_space = noise_as_percentage_from_action_space
         self.clip_low = clip_low
         self.clip_high = clip_high
 
@@ -71,17 +75,21 @@ class TruncatedNormal(ContinuousActionExplorationPolicy):
                 or not np.all(-np.inf < action_space.low) or not np.all(action_space.low < np.inf):
             raise ValueError("Additive noise exploration requires bounded actions")
 
-        # TODO: allow working with unbounded actions by defining the noise in terms of range and not percentage
-
     def get_action(self, action_values: List[ActionType]) -> ActionType:
-        # set the current noise percentage
+        # set the current noise
         if self.phase == RunPhase.TEST:
-            current_noise_precentage = self.evaluation_noise_percentage
+            current_noise = self.evaluation_noise
         else:
-            current_noise_precentage = self.noise_percentage_schedule.current_value
+            current_noise = self.noise_schedule.current_value
 
         # scale the noise to the action space range
-        action_values_std = current_noise_precentage * (self.action_space.high - self.action_space.low)
+        if self.noise_as_percentage_from_action_space:
+            action_values_std = current_noise * (self.action_space.high - self.action_space.low)
+        else:
+            action_values_std = current_noise
+
+        # scale the noise to the action space range
+        action_values_std = current_noise * (self.action_space.high - self.action_space.low)
 
         # extract the mean values
         if isinstance(action_values, list):
@@ -93,7 +101,7 @@ class TruncatedNormal(ContinuousActionExplorationPolicy):
 
         # step the noise schedule
         if self.phase is not RunPhase.TEST:
-            self.noise_percentage_schedule.step()
+            self.noise_schedule.step()
             # the second element of the list is assumed to be the standard deviation
             if isinstance(action_values, list) and len(action_values) > 1:
                 action_values_std = action_values[1].squeeze()
@@ -107,4 +115,4 @@ class TruncatedNormal(ContinuousActionExplorationPolicy):
         return action
 
     def get_control_param(self):
-        return np.ones(self.action_space.shape)*self.noise_percentage_schedule.current_value
+        return np.ones(self.action_space.shape)*self.noise_schedule.current_value

rl_coach/presets/CARLA_CIL.py

@@ -123,8 +123,8 @@ agent_params.input_filter.add_observation_filter(
 
 # no exploration is used
 agent_params.exploration = AdditiveNoiseParameters()
-agent_params.exploration.noise_percentage_schedule = ConstantSchedule(0)
-agent_params.exploration.evaluation_noise_percentage = 0
+agent_params.exploration.noise_schedule = ConstantSchedule(0)
+agent_params.exploration.evaluation_noise = 0
 
 # no playing during the training phase
 agent_params.algorithm.num_consecutive_playing_steps = EnvironmentSteps(0)

rl_coach/presets/CartPole_Dueling_DDQN.py

@@ -53,7 +53,7 @@ env_params = GymVectorEnvironment(level='CartPole-v0')
 preset_validation_params = PresetValidationParameters()
 preset_validation_params.test = True
 preset_validation_params.min_reward_threshold = 150
-preset_validation_params.max_episodes_to_achieve_reward = 250
+preset_validation_params.max_episodes_to_achieve_reward = 300
 
 graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params,
                                     schedule_params=schedule_params, vis_params=VisualizationParameters(),

rl_coach/presets/Fetch_DDPG_HER_baselines.py

@@ -87,9 +87,9 @@ agent_params.memory.shared_memory = True
 agent_params.exploration = EGreedyParameters()
 agent_params.exploration.epsilon_schedule = ConstantSchedule(0.3)
 agent_params.exploration.evaluation_epsilon = 0
-# they actually take the noise_percentage_schedule to be 0.2 * max_abs_range which is 0.1 * total_range
-agent_params.exploration.continuous_exploration_policy_parameters.noise_percentage_schedule = ConstantSchedule(0.1)
-agent_params.exploration.continuous_exploration_policy_parameters.evaluation_noise_percentage = 0
+# they actually take the noise_schedule to be 0.2 * max_abs_range which is 0.1 * total_range
+agent_params.exploration.continuous_exploration_policy_parameters.noise_schedule = ConstantSchedule(0.1)
+agent_params.exploration.continuous_exploration_policy_parameters.evaluation_noise = 0
 
 agent_params.input_filter = InputFilter()
 agent_params.input_filter.add_observation_filter('observation', 'clipping', ObservationClippingFilter(-200, 200))

rl_coach/presets/Mujoco_DDPG.py

@@ -15,7 +15,7 @@ schedule_params = ScheduleParameters()
 schedule_params.improve_steps = EnvironmentSteps(2000000)
 schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(20)
 schedule_params.evaluation_steps = EnvironmentEpisodes(1)
-schedule_params.heatup_steps = EnvironmentSteps(1000)
+schedule_params.heatup_steps = EnvironmentSteps(10000)
 
 #########
 # Agent #
@@ -38,7 +38,7 @@ env_params = GymVectorEnvironment(level=SingleLevelSelection(mujoco_v2))
 preset_validation_params = PresetValidationParameters()
 preset_validation_params.test = True
 preset_validation_params.min_reward_threshold = 400
-preset_validation_params.max_episodes_to_achieve_reward = 1000
+preset_validation_params.max_episodes_to_achieve_reward = 3000
 preset_validation_params.reward_test_level = 'inverted_pendulum'
 preset_validation_params.trace_test_levels = ['inverted_pendulum', 'hopper']
 

rl_coach/presets/Mujoco_TD3.py

@@ -0,0 +1,49 @@
+from rl_coach.agents.td3_agent import TD3AgentParameters
+from rl_coach.architectures.layers import Dense
+from rl_coach.base_parameters import VisualizationParameters, PresetValidationParameters, EmbedderScheme
+from rl_coach.core_types import EnvironmentEpisodes, EnvironmentSteps
+from rl_coach.environments.environment import SingleLevelSelection
+from rl_coach.environments.gym_environment import GymVectorEnvironment, mujoco_v2
+from rl_coach.graph_managers.basic_rl_graph_manager import BasicRLGraphManager
+from rl_coach.graph_managers.graph_manager import ScheduleParameters
+
+####################
+# Graph Scheduling #
+####################
+
+schedule_params = ScheduleParameters()
+schedule_params.improve_steps = EnvironmentSteps(1000000)
+schedule_params.steps_between_evaluation_periods = EnvironmentSteps(5000)
+schedule_params.evaluation_steps = EnvironmentEpisodes(10)
+schedule_params.heatup_steps = EnvironmentSteps(10000)
+
+#########
+# Agent #
+#########
+agent_params = TD3AgentParameters()
+agent_params.network_wrappers['actor'].input_embedders_parameters['observation'].scheme = [Dense(400)]
+agent_params.network_wrappers['actor'].middleware_parameters.scheme = [Dense(300)]
+
+agent_params.network_wrappers['critic'].input_embedders_parameters['observation'].scheme = EmbedderScheme.Empty
+agent_params.network_wrappers['critic'].input_embedders_parameters['action'].scheme = EmbedderScheme.Empty
+agent_params.network_wrappers['critic'].middleware_parameters.scheme = [Dense(400), Dense(300)]
+
+
+###############
+# Environment #
+###############
+env_params = GymVectorEnvironment(level=SingleLevelSelection(mujoco_v2))
+
+########
+# Test #
+########
+preset_validation_params = PresetValidationParameters()
+preset_validation_params.test = True
+preset_validation_params.min_reward_threshold = 500
+preset_validation_params.max_episodes_to_achieve_reward = 1100
+preset_validation_params.reward_test_level = 'hopper'
+preset_validation_params.trace_test_levels = ['inverted_pendulum', 'hopper']
+
+graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params,
+                                    schedule_params=schedule_params, vis_params=VisualizationParameters(),
+                                    preset_validation_params=preset_validation_params)

rl_coach/presets/Starcraft_CollectMinerals_A3C.py

@@ -37,9 +37,9 @@ agent_params.network_wrappers['main'].input_embedders_parameters = {
 }
 
 agent_params.exploration = AdditiveNoiseParameters()
-agent_params.exploration.noise_percentage_schedule = ConstantSchedule(0.05)
-# agent_params.exploration.noise_percentage_schedule = LinearSchedule(0.4, 0.05, 100000)
-agent_params.exploration.evaluation_noise_percentage = 0.05
+agent_params.exploration.noise_schedule = ConstantSchedule(0.05)
+# agent_params.exploration.noise_schedule = LinearSchedule(0.4, 0.05, 100000)
+agent_params.exploration.evaluation_noise = 0.05
 
 agent_params.network_wrappers['main'].batch_size = 64
 agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5

rl_coach/run_multiple_seeds.py

@@ -53,10 +53,14 @@ if __name__ == "__main__":
                              "the preset name, followed by the environment level",
                         default='',
                         type=str)
-    parser.add_argument('-sd', '--level_as_sub_dir',
+    parser.add_argument('-lsd', '--level_as_sub_dir',
                         help="(flag) Store each level in it's own sub directory where the root directory name matches "
                              "the preset name",
                         action='store_true')
+    parser.add_argument('-ssd', '--seed_as_sub_dir',
+                        help="(flag) Store each seed in it's own sub directory where the root directory name matches "
+                             "the preset name",
+                        action='store_true')
     parser.add_argument('-ew', '--evaluation_worker',
                         help="(flag) Start an additional worker that will only do evaluation",
                         action='store_true')
@@ -108,6 +112,8 @@ if __name__ == "__main__":
                 command.append("-c")
             if args.evaluation_worker:
                 command.append("-ew")
+            if args.seed_as_sub_dir:
+                seed = ''
             if level is not None:
                 command.extend(['-lvl', '{}'.format(level)])
                 if level_as_sub_dir: