DDPG Critic Head Bug Fix (#344)

* A bug fix for DDPG, where the update to the policy network was based on the sum of the critic's Q predictions on the batch instead of their mean

rl_coach/agents/ddpg_agent.py

@@ -23,7 +23,7 @@ import numpy as np
 from rl_coach.agents.actor_critic_agent import ActorCriticAgent
 from rl_coach.agents.agent import Agent
 from rl_coach.architectures.embedder_parameters import InputEmbedderParameters
-from rl_coach.architectures.head_parameters import DDPGActorHeadParameters, VHeadParameters
+from rl_coach.architectures.head_parameters import DDPGActorHeadParameters, DDPGVHeadParameters
 from rl_coach.architectures.middleware_parameters import FCMiddlewareParameters
 from rl_coach.base_parameters import NetworkParameters, AlgorithmParameters, \
     AgentParameters, EmbedderScheme
@@ -39,8 +39,10 @@ class DDPGCriticNetworkParameters(NetworkParameters):
         self.input_embedders_parameters = {'observation': InputEmbedderParameters(batchnorm=True),
                                             'action': InputEmbedderParameters(scheme=EmbedderScheme.Shallow)}
         self.middleware_parameters = FCMiddlewareParameters()
-        self.heads_parameters = [VHeadParameters()]
+        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
@@ -56,6 +58,8 @@ class DDPGActorNetworkParameters(NetworkParameters):
         self.middleware_parameters = FCMiddlewareParameters(batchnorm=True)
         self.heads_parameters = [DDPGActorHeadParameters()]
         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 = 0.0001
@@ -140,7 +144,7 @@ class DDPGAgent(ActorCriticAgent):
 
         critic_inputs = copy.copy(batch.next_states(critic_keys))
         critic_inputs['action'] = next_actions
-        q_st_plus_1 = critic.target_network.predict(critic_inputs)
+        q_st_plus_1 = critic.target_network.predict(critic_inputs)[0]
 
         # calculate the bootstrapped TD targets while discounting terminal states according to
         # use_non_zero_discount_for_terminal_states
@@ -160,7 +164,7 @@ class DDPGAgent(ActorCriticAgent):
         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[0]['action'])
+                                                         outputs=critic.online_network.gradients_wrt_inputs[1]['action'])
 
         # train the critic
         critic_inputs = copy.copy(batch.states(critic_keys))

rl_coach/architectures/head_parameters.py

@@ -57,6 +57,17 @@ class VHeadParameters(HeadParameters):
         self.initializer = initializer
 
 
+class DDPGVHeadParameters(HeadParameters):
+    def __init__(self, activation_function: str ='relu', name: str='ddpg_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='normalized_columns'):
+        super().__init__(parameterized_class_name="DDPGVHead", 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
+
+
 class CategoricalQHeadParameters(HeadParameters):
     def __init__(self, activation_function: str ='relu', name: str='categorical_q_head_params',
                  num_output_head_copies: int = 1, rescale_gradient_from_head_by_factor: float = 1.0,

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 .ddpg_v_head import DDPGVHead
 
 __all__ = [
     'CategoricalQHead',
@@ -35,5 +36,6 @@ __all__ = [
     'SACPolicyHead',
     'SACQHead',
     'ClassificationHead',
-    'RegressionHead'
+    'RegressionHead',
+    'DDPGVHead'
 ]

rl_coach/architectures/tensorflow_components/heads/ddpg_v_head.py

@@ -0,0 +1,39 @@
+#
+# Copyright (c) 2017 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.heads import VHead
+from rl_coach.architectures.tensorflow_components.layers import Dense
+from rl_coach.base_parameters import AgentParameters
+from rl_coach.spaces import SpacesDefinition
+
+
+class DDPGVHead(VHead):
+    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='normalized_columns'):
+        super().__init__(agent_parameters, spaces, network_name, head_idx, loss_weight, is_local, activation_function,
+                         dense_layer=dense_layer, initializer=initializer)
+
+    def _build_module(self, input_layer):
+        super()._build_module(input_layer)
+        self.output = [self.output, tf.reduce_mean(self.output)]
+
+    def __str__(self):
+        result = [
+            "Dense (num outputs = 1)"
+        ]
+        return '\n'.join(result)