# coding=utf-8
# Copyright 2020 The Google Research Authors.
#
# 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.
# Lint as: python3
"""Multi Q-Network DQN agent."""
import copy
import os
from batch_rl.multi_head import atari_helpers
from dopamine.agents.dqn import dqn_agent
import gin
import tensorflow.compat.v1 as tf
@gin.configurable
class MultiNetworkDQNAgent(dqn_agent.DQNAgent):
"""DQN agent with multiple heads."""
def __init__(self,
sess,
num_actions,
num_networks=1,
transform_strategy='IDENTITY',
num_convex_combinations=1,
network=atari_helpers.MulitNetworkQNetwork,
init_checkpoint_dir=None,
use_deep_exploration=False,
**kwargs):
"""Initializes the agent and constructs the components of its graph.
Args:
sess: tf.Session, for executing ops.
num_actions: int, number of actions the agent can take at any state.
num_networks: int, Number of different Q-functions.
transform_strategy: str, Possible options include (1) 'STOCHASTIC' for
multiplication with a left stochastic matrix. (2) 'IDENTITY', in which
case the heads are not transformed.
num_convex_combinations: If transform_strategy is 'STOCHASTIC',
then this argument specifies the number of random
convex combinations to be created. If None, `num_heads` convex
combinations are created.
network: tf.Keras.Model. A call to this object will return an
instantiation of the network provided. The network returned can be run
with different inputs to create different outputs. See
atari_helpers.MultiNetworkQNetwork as an example.
init_checkpoint_dir: str, directory from which initial checkpoint before
training is loaded if there doesn't exist any checkpoint in the current
agent directory. If None, no initial checkpoint is loaded.
use_deep_exploration: Adaptation of Bootstrapped DQN for REM exploration.
**kwargs: Arbitrary keyword arguments.
"""
tf.logging.info('Creating MultiNetworkDQNAgent with following parameters:')
tf.logging.info('\t num_networks: %d', num_networks)
tf.logging.info('\t transform_strategy: %s', transform_strategy)
tf.logging.info('\t num_convex_combinations: %d', num_convex_combinations)
tf.logging.info('\t init_checkpoint_dir: %s', init_checkpoint_dir)
tf.logging.info('\t use_deep_exploration %s', use_deep_exploration)
self.num_networks = num_networks
if init_checkpoint_dir is not None:
self._init_checkpoint_dir = os.path.join(init_checkpoint_dir,
'checkpoints')
else:
self._init_checkpoint_dir = None
# The transform matrix should be created on device specified by tf_device
# if the transform_strategy is UNIFORM_STOCHASTIC or STOCHASTIC
self._q_networks_transform = None
self._num_convex_combinations = num_convex_combinations
self.transform_strategy = transform_strategy
self.use_deep_exploration = use_deep_exploration
super(MultiNetworkDQNAgent, self).__init__(
sess, num_actions, network=network, **kwargs)
def _create_network(self, name):
"""Builds a multi-network Q-network that outputs Q-values for each network.
Args:
name: str, this name is passed to the tf.keras.Model and used to create
variable scope under the hood by the tf.keras.Model.
Returns:
network: tf.keras.Model, the network instantiated by the Keras model.
"""
# Pass the device_fn to place Q-networks on different devices
kwargs = {'device_fn': lambda i: '/gpu:{}'.format(i // 4)}
if self._q_networks_transform is None:
if self.transform_strategy == 'STOCHASTIC':
tf.logging.info('Creating q_networks transformation matrix..')
self._q_networks_transform = atari_helpers.random_stochastic_matrix(
self.num_networks, num_cols=self._num_convex_combinations)
if self._q_networks_transform is not None:
kwargs.update({'transform_matrix': self._q_networks_transform})
return self.network(
num_actions=self.num_actions,
num_networks=self.num_networks,
transform_strategy=self.transform_strategy,
name=name,
**kwargs)
def _build_target_q_op(self):
"""Build an op used as a target for the Q-value.
Returns:
target_q_op: An op calculating the Q-value.
"""
# Get the maximum Q-value across the actions dimension for each head.
replay_next_qt_max = tf.reduce_max(
self._replay_next_target_net_outputs.q_networks, axis=1)
is_non_terminal = 1. - tf.cast(self._replay.terminals, tf.float32)
is_non_terminal = tf.expand_dims(is_non_terminal, axis=-1)
rewards = tf.expand_dims(self._replay.rewards, axis=-1)
return rewards + (
self.cumulative_gamma * replay_next_qt_max * is_non_terminal)
def begin_episode(self, observation):
"""Returns the agent's first action for this episode.
Args:
observation: numpy array, the environment's initial observation.
Returns:
int, the selected action.
"""
if self.use_deep_exploration:
# Randomly pick a Q-function from all possible Q-functions for data
# collection each episode for online experiments, similar to deep
# exploration strategy proposed by Bootstrapped DQN
self._sess.run(self._update_episode_q_function)
return super(MultiNetworkDQNAgent, self).begin_episode(observation)
def _build_networks(self):
super(MultiNetworkDQNAgent, self)._build_networks()
# q_argmax is only used for picking an action
self._q_argmax_eval = tf.argmax(self._net_outputs.q_values, axis=1)[0]
if self.use_deep_exploration:
if self.transform_strategy.endswith('STOCHASTIC'):
q_transform = atari_helpers.random_stochastic_matrix(
self.num_networks, num_cols=1)
self._q_episode_transform = tf.get_variable(
trainable=False,
dtype=tf.float32,
shape=q_transform.get_shape().as_list(),
name='q_episode_transform')
self._update_episode_q_function = self._q_episode_transform.assign(
q_transform)
episode_q_function = tf.tensordot(
self._net_outputs.unordered_q_networks,
self._q_episode_transform, axes=[[2], [0]])
self._q_argmax_train = tf.argmax(episode_q_function[:, :, 0], axis=1)[0]
elif self.transform_strategy == 'IDENTITY':
self._q_function_index = tf.Variable(
initial_value=0,
trainable=False,
dtype=tf.int32,
shape=(),
name='q_head_episode')
self._update_episode_q_function = self._q_function_index.assign(
tf.random.uniform(
shape=(), maxval=self.num_networks, dtype=tf.int32))
q_function = self._net_outputs.unordered_q_networks[
:, :, self._q_function_index]
# This is only used for picking an action
self._q_argmax_train = tf.argmax(q_function, axis=1)[0]
else:
self._q_argmax_train = self._q_argmax_eval
def _select_action(self):
if self.eval_mode:
self._q_argmax = self._q_argmax_eval
else:
self._q_argmax = self._q_argmax_train
return super(MultiNetworkDQNAgent, self)._select_action()
def _build_train_op(self):
"""Builds a training op.
Returns:
train_op: An op performing one step of training from replay data.
"""
actions = self._replay.actions
indices = tf.stack([tf.range(actions.shape[0]), actions], axis=-1)
replay_chosen_q = tf.gather_nd(
self._replay_net_outputs.q_networks, indices=indices)
target = tf.stop_gradient(self._build_target_q_op())
loss = tf.losses.huber_loss(
target, replay_chosen_q, reduction=tf.losses.Reduction.NONE)
q_head_losses = tf.reduce_mean(loss, axis=0)
final_loss = tf.reduce_mean(q_head_losses)
if self.summary_writer is not None:
with tf.variable_scope('Losses'):
tf.summary.scalar('HuberLoss', final_loss)
self.optimizers = [copy.deepcopy(self.optimizer) for _ in
range(self.num_networks)]
train_ops = []
for i in range(self.num_networks):
var_list = tf.trainable_variables(scope='Online/subnet_{}'.format(i))
train_op = self.optimizers[i].minimize(final_loss, var_list=var_list)
train_ops.append(train_op)
return tf.group(*train_ops, name='merged_train_op')
