-
Notifications
You must be signed in to change notification settings - Fork 2
/
A3CAgentContinuous.py
276 lines (237 loc) · 7.61 KB
/
A3CAgentContinuous.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
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
#! -*- coding: UTF-8 -*-
import tensorflow as tf
import tensorflow_probability as tfp
import numpy as np
from collections import deque
import random
class Actor:
def __init__(self, name, obvSpace_dim, actSpace_dim, actSpace_low, actSpace_high, alpha, training):
self.name = name
self.training = training
# Value that encourages exploration.
# Directly added to sigma (standard deviation value for normal distribution).
self.explore_value = 1e-5
self.entropy_value = 0.005
with tf.variable_scope(self.name):
self.tf_state = tf.placeholder(tf.float32, shape=(None,)+obvSpace_dim)
# Forward.
tf_hidden_1 = tf.layers.dense(
self.tf_state,
units=512,
activation=tf.nn.relu6,
kernel_initializer=tf.contrib.layers.xavier_initializer()
)
tf_hidden_out = tf.layers.dense(
tf_hidden_1,
units=256,
activation=tf.nn.relu6,
kernel_initializer=tf.contrib.layers.xavier_initializer()
)
self.tf_mu = tf.layers.dense(
tf_hidden_out,
units=actSpace_dim,
activation=tf.nn.tanh,
kernel_initializer=tf.contrib.layers.xavier_initializer()
) * actSpace_high
self.tf_sigma = tf.layers.dense(
tf_hidden_out,
units=actSpace_dim,
activation=tf.nn.softplus,
kernel_initializer=tf.contrib.layers.xavier_initializer()
)
if self.training:
self.tf_sigma = self.tf_sigma + self.explore_value
self.tf_normal_distribution = tfp.distributions.Normal(
loc=self.tf_mu,
scale=self.tf_sigma
)
self.tf_output_sample = tf.squeeze(self.tf_normal_distribution.sample(1), axis=0)
self.tf_output = tf.clip_by_value(
self.tf_output_sample,
actSpace_low,
actSpace_high
)
# Backward.
if self.training:
self.tf_advantage = tf.placeholder(tf.float32, shape=(None, 1))
self.tf_action = tf.placeholder(tf.float32, shape=(None, actSpace_dim))
self.tf_log_loss = self.tf_normal_distribution.log_prob(self.tf_action) * self.tf_advantage
self.tf_entropy = self.entropy_value * self.tf_normal_distribution.entropy()
self.tf_loss = tf.reduce_mean(-(self.tf_log_loss + self.tf_entropy))
self.tf_optimizer = tf.train.RMSPropOptimizer(alpha)
# Calculate gradients on worker's weights.
self.tf_gradients = tf.gradients(
self.tf_loss,
tf.trainable_variables(self.name)
)
# Apply gradients to main network.
self.tf_train = self.tf_optimizer.apply_gradients(
zip(
self.tf_gradients,
tf.trainable_variables("_".join(["Main", "Actor"]))
)
)
class Critic:
def __init__(self, name, obvSpace_dim, actSpace_dim, alpha, training):
self.name = name
self.training = training
with tf.variable_scope(self.name):
self.tf_state = tf.placeholder(tf.float32, shape=(None,)+obvSpace_dim)
# Forward.
tf_hidden_1 = tf.layers.dense(
self.tf_state,
units=512,
activation=tf.nn.relu6,
kernel_initializer=tf.contrib.layers.xavier_initializer()
)
tf_hidden_out = tf.layers.dense(
tf_hidden_1,
units=256,
activation=tf.nn.relu6,
kernel_initializer=tf.contrib.layers.xavier_initializer()
)
self.tf_output = tf.layers.dense(
tf_hidden_out,
units=1,
activation=None,
kernel_initializer=tf.contrib.layers.xavier_initializer()
)
# Backward.
if self.training:
self.tf_value_target = tf.placeholder(tf.float32, shape=(None, 1))
self.tf_loss = tf.reduce_mean(
tf.square(
self.tf_value_target - self.tf_output
)
)
self.tf_optimizer = tf.train.RMSPropOptimizer(alpha)
# Calculate gradients on worker's weights.
self.tf_gradients = tf.gradients(
self.tf_loss,
tf.trainable_variables(self.name)
)
# Apply gradients to main network.
self.tf_train = self.tf_optimizer.apply_gradients(
zip(
self.tf_gradients,
tf.trainable_variables("_".join(["Main", "Critic"]))
)
)
class A3CAgentContinuous:
def __init__(self, name, sess, obvSpace_dim, actSpace_dim, actSpace_low, actSpace_high, gamma, alpha_actor, alpha_critic, training=True):
self.name = name
self.sess = sess
self.obvSpace_dim = obvSpace_dim
self.actSpace_dim = actSpace_dim
self.actSpace_low = actSpace_low
self.actSpace_high = actSpace_high
self.gamma = gamma
self.training = training
self.name_actor = "_".join([self.name, "Actor"])
self.name_critic = "_".join([self.name, "Critic"])
self.memory_max_size = 1000000
self.done = False
# Build networks.
self.actor = Actor(self.name_actor, obvSpace_dim, actSpace_dim, actSpace_low, actSpace_high, alpha_actor, self.training)
if self.training:
self.critic = Critic(self.name_critic, obvSpace_dim, actSpace_dim, alpha_critic, self.training)
# Memory stores only current episode's experiences.
# At end of the train function, its set to be empty.
self.Memory = deque(maxlen=self.memory_max_size)
# Op that pulls weights from global network.
if self.name != "Main":
self.tf_main_weights_update = [
tar_v.assign(src_v)
for src_v, tar_v in zip(
tf.trainable_variables("Main"),
tf.trainable_variables(self.name)
)
]
def act(self, state):
# With given state, calculates output of policy network.
with self.sess.as_default(), self.sess.graph.as_default():
state = np.expand_dims(state, axis=0)
output, sigma = self.sess.run(
[self.actor.tf_output, self.actor.tf_sigma],
feed_dict={
self.actor.tf_state:state
}
)
return output[0], np.mean(sigma[0])
# Remember given experience on memory.
def remember(self, state, action, reward, next_state, done):
self.Memory.append(
[
state,
action,
reward,
next_state
]
)
self.done = done
# Sample from memory and train the model on it.
def train(self):
mini_batch = self.Memory
state = np.array([np.array(b[0]) for b in mini_batch])
action = np.array([np.array(b[1]) for b in mini_batch])
reward = np.array([np.array(b[2]) for b in mini_batch])
next_state = np.array([np.array(b[3]) for b in mini_batch])
with self.sess.as_default(), self.sess.graph.as_default():
# Target value for critic.
if self.done:
next_value = 0.0
else:
next_value = self.sess.run(
self.critic.tf_output,
feed_dict={
self.critic.tf_state:np.expand_dims(next_state[-1], axis=0)
}
)[0][0]
value_target = []
for r in reward[::-1]:
next_value = r + self.gamma * next_value
value_target.append(next_value)
value_target.reverse()
value_target = np.expand_dims(np.array(value_target), axis=1)
value_current = self.sess.run(
self.critic.tf_output,
feed_dict={
self.critic.tf_state:state
}
)
advantage = (value_target - value_current)
### Train critic.
self.sess.run(
self.critic.tf_train,
feed_dict={
self.critic.tf_state:state,
self.critic.tf_value_target:value_target
}
)
### Train actor.
self.sess.run(
self.actor.tf_train,
feed_dict={
self.actor.tf_state:state,
self.actor.tf_action:action,
self.actor.tf_advantage:advantage
}
)
# Set memory back to empty.
self.Memory = deque(maxlen=self.memory_max_size)
def weights_update(self):
if self.name != "Main":
with self.sess.as_default(), self.sess.graph.as_default():
self.sess.run(
self.tf_main_weights_update
)
def increment_global_episode(self, global_episode_increment_name):
with self.sess.as_default(), self.sess.graph.as_default():
self.sess.run(
tf.get_default_graph().get_tensor_by_name(global_episode_increment_name)
)
def get_global_episode(self, global_episode_name):
with self.sess.as_default(), self.sess.graph.as_default():
return self.sess.run(
tf.get_default_graph().get_tensor_by_name(global_episode_name)
)