-
Notifications
You must be signed in to change notification settings - Fork 70
/
quantile_agent.py
243 lines (211 loc) · 10.2 KB
/
quantile_agent.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
# 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.
"""Distributional RL agent using quantile regression.
This loss is computed as in "Distributional Reinforcement Learning with Quantile
Regression" - Dabney et. al, 2017"
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from batch_rl.multi_head import atari_helpers
from dopamine.agents.dqn import dqn_agent
from dopamine.agents.rainbow import rainbow_agent
import gin
import tensorflow.compat.v1 as tf
@gin.configurable
class QuantileAgent(rainbow_agent.RainbowAgent):
"""An extension of Rainbow to perform quantile regression."""
def __init__(self,
sess,
num_actions,
kappa=1.0,
network=atari_helpers.QuantileNetwork,
num_atoms=200,
gamma=0.99,
update_horizon=1,
min_replay_history=50000,
update_period=4,
target_update_period=10000,
epsilon_fn=dqn_agent.linearly_decaying_epsilon,
epsilon_train=0.1,
epsilon_eval=0.05,
epsilon_decay_period=1000000,
replay_scheme='prioritized',
tf_device='/cpu:0',
optimizer=tf.train.AdamOptimizer(
learning_rate=0.00005, epsilon=0.0003125),
summary_writer=None,
summary_writing_frequency=500,
minq_weight=10.0):
"""Initializes the agent and constructs the Graph.
Args:
sess: A `tf.Session` object for running associated ops.
num_actions: Int, number of actions the agent can take at any state.
kappa: Float, Huber loss cutoff.
network: tf.Keras.Model, expects 3 parameters: num_actions, num_atoms,
network_type. 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.QuantileNetwork
as an example.
num_atoms: Int, the number of buckets for the value function distribution.
gamma: Float, exponential decay factor as commonly used in the RL
literature.
update_horizon: Int, horizon at which updates are performed, the 'n' in
n-step update.
min_replay_history: Int, number of stored transitions for training to
start.
update_period: Int, period between DQN updates.
target_update_period: Int, ppdate period for the target network.
epsilon_fn: Function expecting 4 parameters: (decay_period, step,
warmup_steps, epsilon), and which returns the epsilon value used for
exploration during training.
epsilon_train: Float, final epsilon for training.
epsilon_eval: Float, epsilon during evaluation.
epsilon_decay_period: Int, number of steps for epsilon to decay.
replay_scheme: String, replay memory scheme to be used. Choices are:
uniform - Standard (DQN) replay buffer (Mnih et al., 2015)
prioritized - Prioritized replay buffer (Schaul et al., 2015)
tf_device: Tensorflow device with which the value function is computed
and trained.
optimizer: A `tf.train.Optimizer` object for training the model.
summary_writer: SummaryWriter object for outputting training statistics.
Summary writing disabled if set to None.
summary_writing_frequency: int, frequency with which summaries will be
written. Lower values will result in slower training.
"""
self.kappa = kappa
super(QuantileAgent, self).__init__(
sess=sess,
num_actions=num_actions,
network=network,
num_atoms=num_atoms,
gamma=gamma,
update_horizon=update_horizon,
min_replay_history=min_replay_history,
update_period=update_period,
target_update_period=target_update_period,
epsilon_fn=epsilon_fn,
epsilon_train=epsilon_train,
epsilon_eval=epsilon_eval,
epsilon_decay_period=epsilon_decay_period,
replay_scheme=replay_scheme,
tf_device=tf_device,
optimizer=optimizer,
summary_writer=summary_writer,
summary_writing_frequency=summary_writing_frequency,
min_q_weight=minq_weight)
self.minq_weight = minq_weight
print ('min Q weight (QR-DQN): ', self.minq_weight)
def _create_network(self, name):
"""Builds a Quantile ConvNet.
Equivalent to Rainbow ConvNet, only now the output logits are interpreted
as quantiles.
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.
"""
network = self.network(self.num_actions, self._num_atoms, name=name)
return network
def _build_target_distribution(self):
batch_size = tf.shape(self._replay.rewards)[0]
# size of rewards: batch_size x 1
rewards = self._replay.rewards[:, None]
# size of tiled_support: batch_size x num_atoms
is_terminal_multiplier = 1. - tf.cast(self._replay.terminals, tf.float32)
# Incorporate terminal state to discount factor.
# size of gamma_with_terminal: batch_size x 1
gamma_with_terminal = self.cumulative_gamma * is_terminal_multiplier
gamma_with_terminal = gamma_with_terminal[:, None]
# size of next_qt_argmax: 1 x batch_size
next_qt_argmax = tf.argmax(
self._replay_next_target_net_outputs.q_values, axis=1)[:, None]
batch_indices = tf.range(tf.to_int64(batch_size))[:, None]
# size of next_qt_argmax: batch_size x 2
batch_indexed_next_qt_argmax = tf.concat(
[batch_indices, next_qt_argmax], axis=1)
# size of next_logits (next quantiles): batch_size x num_atoms
next_logits = tf.gather_nd(
self._replay_next_target_net_outputs.logits,
batch_indexed_next_qt_argmax)
return rewards + gamma_with_terminal * next_logits
def _build_train_op(self):
"""Builds a training op.
Returns:
train_op: An op performing one step of training.
"""
target_distribution = tf.stop_gradient(self._build_target_distribution())
# size of indices: batch_size x 1.
indices = tf.range(tf.shape(self._replay_net_outputs.logits)[0])[:, None]
# size of reshaped_actions: batch_size x 2.
reshaped_actions = tf.concat([indices, self._replay.actions[:, None]], 1)
# For each element of the batch, fetch the logits for its selected action.
chosen_action_logits = tf.gather_nd(self._replay_net_outputs.logits,
reshaped_actions)
bellman_errors = (target_distribution[:, None, :] -
chosen_action_logits[:, :, None]) # Input `u' of Eq. 9.
huber_loss = ( # Eq. 9 of paper.
tf.to_float(tf.abs(bellman_errors) <= self.kappa) *
0.5 * bellman_errors ** 2 +
tf.to_float(tf.abs(bellman_errors) > self.kappa) *
self.kappa * (tf.abs(bellman_errors) - 0.5 * self.kappa))
tau_hat = ((tf.range(self._num_atoms, dtype=tf.float32) + 0.5) /
self._num_atoms) # Quantile midpoints. See Lemma 2 of paper.
quantile_huber_loss = ( # Eq. 10 of paper.
tf.abs(tau_hat[None, :, None] - tf.to_float(bellman_errors < 0)) *
huber_loss)
# Sum over tau dimension, average over target value dimension.
loss = tf.reduce_sum(tf.reduce_mean(quantile_huber_loss, 2), 1)
if self._replay_scheme == 'prioritized':
target_priorities = self._replay.tf_get_priority(self._replay.indices)
# The original prioritized experience replay uses a linear exponent
# schedule 0.4 -> 1.0. Comparing the schedule to a fixed exponent of 0.5
# on 5 games (Asterix, Pong, Q*Bert, Seaquest, Space Invaders) suggested
# a fixed exponent actually performs better, except on Pong.
loss_weights = 1.0 / tf.sqrt(target_priorities + 1e-10)
loss_weights /= tf.reduce_max(loss_weights)
# Rainbow and prioritized replay are parametrized by an exponent alpha,
# but in both cases it is set to 0.5 - for simplicity's sake we leave it
# as is here, using the more direct tf.sqrt(). Taking the square root
# "makes sense", as we are dealing with a squared loss.
# Add a small nonzero value to the loss to avoid 0 priority items. While
# technically this may be okay, setting all items to 0 priority will cause
# troubles, and also result in 1.0 / 0.0 = NaN correction terms.
update_priorities_op = self._replay.tf_set_priority(
self._replay.indices, tf.sqrt(loss + 1e-10))
# Weight loss by inverse priorities.
loss = loss_weights * loss
else:
update_priorities_op = tf.no_op()
### Add the CQL loss
replay_action_one_hot = tf.one_hot(
self._replay.actions, self.num_actions, 1., 0., name='action_one_hot')
replay_chosen_q = tf.reduce_sum(
self._replay_net_outputs.q_values * replay_action_one_hot,
reduction_indices=1,
name='replay_chosen_q')
dataset_expec = tf.reduce_mean(replay_chosen_q)
negative_sampling = tf.reduce_mean(tf.reduce_logsumexp(self._replay_net_outputs.q_values, 1))
min_q_loss = (negative_sampling - dataset_expec)
print ('MIN Q WEIGHT: ', self.min_q_weight)
with tf.control_dependencies([update_priorities_op]):
if self.summary_writer is not None:
with tf.variable_scope('Losses'):
tf.summary.scalar('QuantileLoss', tf.reduce_mean(loss))
tf.summary.scalar('minQLoss', tf.reduce_mean(min_q_loss))
tf.summary.scalar('Q_predictions', tf.reduce_mean(replay_chosen_q))
min_q_loss = min_q_loss * self.min_q_weight
return self.optimizer.minimize(tf.reduce_mean(loss) + min_q_loss), loss