[rllib] example and docs on how to use parametric actions with DQN / …

…PG algorithms (#3384)

doc/source/rllib-algorithms.rst

@@ -133,10 +133,10 @@ Tuned examples: `Pendulum-v0 <https://github.com/ray-project/ray/blob/master/pyt
    :start-after: __sphinx_doc_begin__
    :end-before: __sphinx_doc_end__
 
-Deep Q Networks (DQN, Rainbow)
-------------------------------
+Deep Q Networks (DQN, Rainbow, Parametric DQN)
+----------------------------------------------
 `[paper] <https://arxiv.org/abs/1312.5602>`__ `[implementation] <https://github.com/ray-project/ray/blob/master/python/ray/rllib/agents/dqn/dqn.py>`__
-RLlib DQN is implemented using the SyncReplayOptimizer. The algorithm can be scaled by increasing the number of workers, using the AsyncGradientsOptimizer for async DQN, or using Ape-X. Memory usage is reduced by compressing samples in the replay buffer with LZ4. All of the DQN improvements evaluated in `Rainbow <https://arxiv.org/abs/1710.02298>`__ are available, though not all are enabled by default.
+RLlib DQN is implemented using the SyncReplayOptimizer. The algorithm can be scaled by increasing the number of workers, using the AsyncGradientsOptimizer for async DQN, or using Ape-X. Memory usage is reduced by compressing samples in the replay buffer with LZ4. All of the DQN improvements evaluated in `Rainbow <https://arxiv.org/abs/1710.02298>`__ are available, though not all are enabled by default. See also how to use `parametric-actions in DQN <rllib-models.html#variable-length-parametric-action-spaces>`__.
 
 Tuned examples: `PongDeterministic-v4 <https://github.com/ray-project/ray/blob/master/python/ray/rllib/tuned_examples/pong-dqn.yaml>`__, `Rainbow configuration <https://github.com/ray-project/ray/blob/master/python/ray/rllib/tuned_examples/pong-rainbow.yaml>`__, `{BeamRider,Breakout,Qbert,SpaceInvaders}NoFrameskip-v4 <https://github.com/ray-project/ray/blob/master/python/ray/rllib/tuned_examples/atari-basic-dqn.yaml>`__, `with Dueling and Double-Q <https://github.com/ray-project/ray/blob/master/python/ray/rllib/tuned_examples/atari-duel-ddqn.yaml>`__, `with Distributional DQN <https://github.com/ray-project/ray/blob/master/python/ray/rllib/tuned_examples/atari-dist-dqn.yaml>`__.
 

doc/source/rllib-env.rst

@@ -7,20 +7,22 @@ RLlib works with several different types of environments, including `OpenAI Gym
 
 **Compatibility matrix**:
 
-=============  ================  ==================  ===========  ==================
-Algorithm      Discrete Actions  Continuous Actions  Multi-Agent  Recurrent Policies
-=============  ================  ==================  ===========  ==================
-A2C, A3C        **Yes**           **Yes**             **Yes**      **Yes**
-PPO             **Yes**           **Yes**             **Yes**      **Yes**
-PG              **Yes**           **Yes**             **Yes**      **Yes**
-IMPALA          **Yes**           No                  **Yes**      **Yes**
-DQN, Rainbow    **Yes**           No                  **Yes**      No
-DDPG, TD3       No                **Yes**             **Yes**      No
-APEX-DQN        **Yes**           No                  **Yes**      No
-APEX-DDPG       No                **Yes**             **Yes**      No
-ES              **Yes**           **Yes**             No           No
-ARS             **Yes**           **Yes**             No           No
-=============  ================  ==================  ===========  ==================
+=============  =======================  ==================  ===========  ==================
+Algorithm      Discrete Actions         Continuous Actions  Multi-Agent  Recurrent Policies
+=============  =======================  ==================  ===========  ==================
+A2C, A3C        **Yes** `+parametric`_  **Yes**             **Yes**      **Yes**
+PPO             **Yes** `+parametric`_  **Yes**             **Yes**      **Yes**
+PG              **Yes** `+parametric`_  **Yes**             **Yes**      **Yes**
+IMPALA          **Yes** `+parametric`_  No                  **Yes**      **Yes**
+DQN, Rainbow    **Yes** `+parametric`_  No                  **Yes**      No
+DDPG, TD3       No                      **Yes**             **Yes**      No
+APEX-DQN        **Yes** `+parametric`_  No                  **Yes**      No
+APEX-DDPG       No                      **Yes**             **Yes**      No
+ES              **Yes**                 **Yes**             No           No
+ARS             **Yes**                 **Yes**             No           No
+=============  =======================  ==================  ===========  ==================
+
+.. _`+parametric`: rllib-models.html#variable-length-parametric-action-spaces
 
 In the high-level agent APIs, environments are identified with string names. By default, the string will be interpreted as a gym `environment name <https://gym.openai.com/envs>`__, however you can also register custom environments by name:
 

doc/source/rllib-models.rst

@@ -110,6 +110,43 @@ Custom models should subclass the common RLlib `model class <https://github.com/
 
 For a full example of a custom model in code, see the `Carla RLlib model <https://github.com/ray-project/ray/blob/master/python/ray/rllib/examples/carla/models.py>`__ and associated `training scripts <https://github.com/ray-project/ray/tree/master/python/ray/rllib/examples/carla>`__. You can also reference the `unit tests <https://github.com/ray-project/ray/blob/master/python/ray/rllib/test/test_nested_spaces.py>`__ for Tuple and Dict spaces, which show how to access nested observation fields.
 
+Custom Recurrent Models
+~~~~~~~~~~~~~~~~~~~~~~~
+
+Instead of using the ``use_lstm: True`` option, it can be preferable use a custom recurrent model. This provides more control over postprocessing of the LSTM output and can also allow the use of multiple LSTM cells to process different portions of the input. The only difference from a normal custom model is that you have to define ``self.state_init``, ``self.state_in``, and ``self.state_out``. You can refer to the existing `lstm.py <https://github.com/ray-project/ray/blob/master/python/ray/rllib/models/lstm.py>`__ model as an example to implement your own model:
+
+
+.. code-block:: python
+
+    class MyCustomLSTM(Model):
+        def _build_layers_v2(self, input_dict, num_outputs, options):
+            # Some initial layers to process inputs, shape [BATCH, OBS...].
+            features = some_hidden_layers(input_dict["obs"])
+
+            # Add back the nested time dimension for tf.dynamic_rnn, new shape
+            # will be [BATCH, MAX_SEQ_LEN, OBS...].
+            last_layer = add_time_dimension(features, self.seq_lens)
+
+            # Setup the LSTM cell (see lstm.py for an example)
+            lstm = rnn.BasicLSTMCell(256, state_is_tuple=True)
+            self.state_init = ...
+            self.state_in = ...
+            lstm_out, lstm_state = tf.nn.dynamic_rnn(
+                lstm,
+                last_layer,
+                initial_state=...,
+                sequence_length=self.seq_lens,
+                time_major=False,
+                dtype=tf.float32)
+            self.state_out = list(lstm_state)
+
+            # Drop the time dimension again so back to shape [BATCH, OBS...].
+            # Note that we retain the zero padding (see issue #2992).
+            last_layer = tf.reshape(lstm_out, [-1, cell_size])
+            logits = linear(last_layer, num_outputs, "action",
+                            normc_initializer(0.01))
+            return logits, last_layer
+
 Custom Preprocessors
 --------------------
 
@@ -188,6 +225,53 @@ Then, you can create an agent with your custom policy graph by:
 
 In this example we overrode existing methods of the existing DDPG policy graph, i.e., `_build_q_network`, `_build_p_network`, `_build_action_network`, `_build_actor_critic_loss`, but you can also replace the entire graph class entirely.
 
+Variable-length / Parametric Action Spaces
+------------------------------------------
+
+Custom models can be used to work with environments where (1) the set of valid actions varies per step, and/or (2) the number of valid actions is very large, as in `OpenAI Five <https://neuro.cs.ut.ee/the-use-of-embeddings-in-openai-five/>`__ and `Horizon <https://arxiv.org/abs/1811.00260>`__. The general idea is that the meaning of actions can be completely conditioned on the observation, that is, the ``a`` in ``Q(s, a)`` is just a token in ``[0, MAX_AVAIL_ACTIONS)`` that only has meaning in the context of ``s``. This works with algorithms in the `DQN and policy-gradient families <rllib-env.html>`__ and can be implemented as follows:
+
+1. The environment should return a mask and/or list of valid action embeddings as part of the observation for each step. To enable batching, the number of actions can be allowed to vary from 1 to some max number:
+
+.. code-block:: python
+
+   class MyParamActionEnv(gym.Env):
+       def __init__(self, max_avail_actions):
+           self.action_space = Discrete(max_avail_actions)
+           self.observation_space = Dict({
+               "action_mask": Box(0, 1, shape=(max_avail_actions, )),
+               "avail_actions": Box(-1, 1, shape=(max_avail_actions, action_embedding_sz)),
+               "real_obs": ...,
+           })
+
+2. A custom model can be defined that can interpret the ``action_mask`` and ``avail_actions`` portions of the observation. Here the model computes the action logits via the dot product of some network output and each action embedding. Invalid actions can be masked out of the softmax by scaling the probability to zero:
+
+.. code-block:: python
+
+    class MyParamActionModel(Model):
+        def _build_layers_v2(self, input_dict, num_outputs, options):
+            avail_actions = input_dict["obs"]["avail_actions"]
+            action_mask = input_dict["obs"]["action_mask"]
+
+            output = FullyConnectedNetwork(
+                input_dict["obs"]["real_obs"], num_outputs=action_embedding_sz)
+
+            # Expand the model output to [BATCH, 1, EMBED_SIZE]. Note that the
+            # avail actions tensor is of shape [BATCH, MAX_ACTIONS, EMBED_SIZE].
+            intent_vector = tf.expand_dims(output, 1)
+
+            # Shape of logits is [BATCH, MAX_ACTIONS].
+            action_logits = tf.reduce_sum(avail_actions * intent_vector, axis=2)
+
+            # Mask out invalid actions (use tf.float32.min for stability)
+            inf_mask = tf.maximum(tf.log(action_mask), tf.float32.min)
+            masked_logits = inf_mask + action_logits
+
+            return masked_logits, last_layer
+
+
+Depending on your use case it may make sense to use just the masking, just action embeddings, or both. For a runnable example of this in code, check out `parametric_action_cartpole.py <https://github.com/ray-project/ray/blob/master/python/ray/rllib/examples/parametric_action_cartpole.py>`__. Note that since masking introduces ``tf.float32.min`` values into the model output, this technique might not work with all algorithm options. For example, algorithms might crash if they incorrectly process the ``tf.float32.min`` values. The cartpole example has working configurations for DQN and several policy gradient algorithms.
+
+
 Model-Based Rollouts
 --------------------
 

doc/source/rllib-training.rst

@@ -73,13 +73,13 @@ In an example below, we train A2C by specifying 8 workers through the config fla
     python ray/python/ray/rllib/train.py --env=PongDeterministic-v4 \
         --run=A2C --config '{"num_workers": 8}'
 
-.. image:: rllib-config.svg
-
 Specifying Resources
 ~~~~~~~~~~~~~~~~~~~~
 
 You can control the degree of parallelism used by setting the ``num_workers`` hyperparameter for most agents. The number of GPUs the driver should use can be set via the ``num_gpus`` option. Similarly, the resource allocation to workers can be controlled via ``num_cpus_per_worker``, ``num_gpus_per_worker``, and ``custom_resources_per_worker``. The number of GPUs can be a fractional quantity to allocate only a fraction of a GPU. For example, with DQN you can pack five agents onto one GPU by setting ``num_gpus: 0.2``. Note that in Ray < 0.6.0 fractional GPU support requires setting the environment variable ``RAY_USE_XRAY=1``.
 
+.. image:: rllib-config.svg
+
 Common Parameters
 ~~~~~~~~~~~~~~~~~
 

doc/source/rllib.rst

@@ -56,7 +56,7 @@ Algorithms
 
    -  `Deep Deterministic Policy Gradients (DDPG, TD3) <rllib-algorithms.html#deep-deterministic-policy-gradients-ddpg-td3>`__
 
-   -  `Deep Q Networks (DQN, Rainbow) <rllib-algorithms.html#deep-q-networks-dqn-rainbow>`__
+   -  `Deep Q Networks (DQN, Rainbow, Parametric DQN) <rllib-algorithms.html#deep-q-networks-dqn-rainbow-parametric-dqn>`__
 
    -  `Policy Gradients <rllib-algorithms.html#policy-gradients>`__
 
@@ -75,6 +75,7 @@ Models and Preprocessors
 * `Custom Models <rllib-models.html#custom-models>`__
 * `Custom Preprocessors <rllib-models.html#custom-preprocessors>`__
 * `Customizing Policy Graphs <rllib-models.html#customizing-policy-graphs>`__
+* `Variable-length / Parametric Action Spaces <rllib-models.html#variable-length-parametric-action-spaces>`__
 * `Model-Based Rollouts <rllib-models.html#model-based-rollouts>`__
 
 RLlib Concepts

python/ray/rllib/agents/dqn/dqn_policy_graph.py

@@ -30,28 +30,35 @@ def __init__(self,
                  sigma0=0.5):
         self.model = model
         with tf.variable_scope("action_value"):
-            action_out = model.last_layer
-            for i in range(len(hiddens)):
-                if use_noisy:
-                    action_out = self.noisy_layer("hidden_%d" % i, action_out,
-                                                  hiddens[i], sigma0)
-                else:
-                    action_out = layers.fully_connected(
-                        action_out,
-                        num_outputs=hiddens[i],
-                        activation_fn=tf.nn.relu)
+            if hiddens:
+                action_out = model.last_layer
+                for i in range(len(hiddens)):
+                    if use_noisy:
+                        action_out = self.noisy_layer(
+                            "hidden_%d" % i, action_out, hiddens[i], sigma0)
+                    else:
+                        action_out = layers.fully_connected(
+                            action_out,
+                            num_outputs=hiddens[i],
+                            activation_fn=tf.nn.relu)
+            else:
+                # Avoid postprocessing the outputs. This enables custom models
+                # to be used for parametric action DQN.
+                action_out = model.outputs
             if use_noisy:
                 action_scores = self.noisy_layer(
                     "output",
                     action_out,
                     num_actions * num_atoms,
                     sigma0,
                     non_linear=False)
-            else:
+            elif hiddens:
                 action_scores = layers.fully_connected(
                     action_out,
                     num_outputs=num_actions * num_atoms,
                     activation_fn=None)
+            else:
+                action_scores = model.outputs
             if num_atoms > 1:
                 # Distributional Q-learning uses a discrete support z
                 # to represent the action value distribution
@@ -107,7 +114,7 @@ def __init__(self,
                 self.logits = support_logits_per_action
                 self.dist = support_prob_per_action
             else:
-                action_scores_mean = tf.reduce_mean(action_scores, 1)
+                action_scores_mean = _reduce_mean_ignore_inf(action_scores, 1)
                 action_scores_centered = action_scores - tf.expand_dims(
                     action_scores_mean, 1)
                 self.value = state_score + action_scores_centered
@@ -176,11 +183,15 @@ class QValuePolicy(object):
     def __init__(self, q_values, observations, num_actions, stochastic, eps):
         deterministic_actions = tf.argmax(q_values, axis=1)
         batch_size = tf.shape(observations)[0]
-        random_actions = tf.random_uniform(
-            tf.stack([batch_size]),
-            minval=0,
-            maxval=num_actions,
-            dtype=tf.int64)
+
+        # Special case masked out actions (q_value ~= -inf) so that we don't
+        # even consider them for exploration.
+        random_valid_action_logits = tf.where(
+            tf.equal(q_values, tf.float32.min),
+            tf.ones_like(q_values) * tf.float32.min, tf.ones_like(q_values))
+        random_actions = tf.squeeze(
+            tf.multinomial(random_valid_action_logits, 1), axis=1)
+
         chose_random = tf.random_uniform(
             tf.stack([batch_size]), minval=0, maxval=1, dtype=tf.float32) < eps
         stochastic_actions = tf.where(chose_random, random_actions,
@@ -368,8 +379,8 @@ def _build_q_network(self, obs, space):
         qnet = QNetwork(
             ModelCatalog.get_model({
                 "obs": obs
-            }, space, 1, self.config["model"]), self.num_actions,
-            self.config["dueling"], self.config["hiddens"],
+            }, space, self.num_actions, self.config["model"]),
+            self.num_actions, self.config["dueling"], self.config["hiddens"],
             self.config["noisy"], self.config["num_atoms"],
             self.config["v_min"], self.config["v_max"], self.config["sigma0"])
         return qnet.value, qnet.logits, qnet.dist, qnet.model
@@ -507,6 +518,14 @@ def _postprocess_dqn(policy_graph, sample_batch):
     return batch
 
 
+def _reduce_mean_ignore_inf(x, axis):
+    """Same as tf.reduce_mean() but ignores -inf values."""
+    mask = tf.not_equal(x, tf.float32.min)
+    x_zeroed = tf.where(mask, x, tf.zeros_like(x))
+    return (tf.reduce_sum(x_zeroed, axis) / tf.reduce_sum(
+        tf.cast(mask, tf.float32), axis))
+
+
 def _huber_loss(x, delta=1.0):
     """Reference: https://en.wikipedia.org/wiki/Huber_loss"""
     return tf.where(

python/ray/rllib/agents/ppo/ppo.py

@@ -110,6 +110,11 @@ def _validate_config(self):
                 and not self.config["simple_optimizer"]):
             logger.warn("forcing simple_optimizer=True in multi-agent mode")
             self.config["simple_optimizer"] = True
+        if self.config["observation_filter"] != "NoFilter":
+            # TODO(ekl): consider setting the default to be NoFilter
+            logger.warn(
+                "By default, observations will be normalized with {}".format(
+                    self.config["observation_filter"]))
 
     def _train(self):
         prev_steps = self.optimizer.num_steps_sampled