forked from Ceruleanacg/Personae
-
Notifications
You must be signed in to change notification settings - Fork 1
/
DDPG.py
212 lines (172 loc) · 8.74 KB
/
DDPG.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
# coding=utf-8
import tensorflow as tf
import numpy as np
import os
from algorithm import config
from base.env.market import Market
from checkpoints import CHECKPOINTS_DIR
from base.algorithm.model import BaseRLTFModel
from helper.args_parser import model_launcher_parser
from helper.data_logger import generate_algorithm_logger, generate_market_logger
class Algorithm(BaseRLTFModel):
def __init__(self, session, env, a_space, s_space, **options):
super(Algorithm, self).__init__(session, env, a_space, s_space, **options)
self.actor_loss, self.critic_loss = .0, .0
# Initialize buffer.
self.buffer = np.zeros((self.buffer_size, self.s_space * 2 + 1 + 1))
self.buffer_length = 0
self._init_input()
self._init_nn()
self._init_op()
self._init_saver()
self._init_summary_writer()
def _init_input(self):
self.s = tf.placeholder(tf.float32, [None, self.s_space], 'state')
self.r = tf.placeholder(tf.float32, [None, 1], 'reward')
self.s_next = tf.placeholder(tf.float32, [None, self.s_space], 'state_next')
def _init_nn(self):
# Initialize predict actor and critic.
self.a_predict = self.__build_actor_nn(self.s, "predict/actor", trainable=True)
self.q_predict = self.__build_critic(self.s, self.a_predict, "predict/critic", trainable=True)
# Initialize target actor and critic.
self.a_next = self.__build_actor_nn(self.s_next, "target/actor", trainable=False)
self.q_next = self.__build_critic(self.s_next, self.a_next, "target/critic", trainable=False)
# Save scopes
self.scopes = ["predict/actor", "target/actor", "predict/critic", "target/critic"]
def _init_op(self):
# Get actor and critic parameters.
params = [tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope) for scope in self.scopes]
zipped_a_params, zipped_c_params = zip(params[0], params[1]), zip(params[2], params[3])
# Initialize update actor and critic op.
self.update_a = [tf.assign(t_a, (1 - self.tau) * t_a + self.tau * p_a) for p_a, t_a in zipped_a_params]
self.update_c = [tf.assign(t_c, (1 - self.tau) * t_c + self.tau * p_c) for p_c, t_c in zipped_c_params]
# Initialize actor loss and train op.
with tf.variable_scope('actor_loss'):
self.a_loss = -tf.reduce_mean(self.q_predict)
with tf.variable_scope('actor_train'):
self.a_train_op = tf.train.RMSPropOptimizer(self.learning_rate).minimize(self.a_loss, var_list=params[0])
# Initialize critic loss and train op.
self.q_target = self.r + self.gamma * self.q_next
with tf.variable_scope('critic_loss'):
self.c_loss = tf.losses.mean_squared_error(self.q_target, self.q_predict)
with tf.variable_scope('critic_train'):
self.c_train_op = tf.train.RMSPropOptimizer(self.learning_rate * 2).minimize(self.c_loss, var_list=params[2])
# Initialize variables.
self.session.run(tf.global_variables_initializer())
def run(self):
if self.mode != 'train':
self.restore()
else:
for episode in range(self.episodes):
self.log_loss(episode)
s = self.env.reset(self.mode)
while True:
c, a, a_index = self.predict(s)
s_next, r, status, info = self.env.forward(c, a)
self.save_transition(s, a_index, r, s_next)
self.train()
s = s_next
if status == self.env.Done:
self.env.trader.log_asset(episode)
break
if self.enable_saver and episode % 10 == 0:
self.save(episode)
def train(self):
if self.buffer_length < self.buffer_size:
return
self.session.run([self.update_a, self.update_c])
s, a, r, s_next = self.get_transition_batch()
self.critic_loss, _ = self.session.run([self.c_loss, self.c_train_op], {self.s: s, self.a_predict: a, self.r: r, self.s_next: s_next})
self.actor_loss, _ = self.session.run([self.a_loss, self.a_train_op], {self.s: s})
def predict(self, s):
a = self.session.run(self.a_predict, {self.s: s})[0][0]
return self.get_stock_code_and_action(a, use_greedy=True, use_prob=True if self.mode == 'train' else False)
def save_transition(self, s, a, r, s_next):
transition = np.hstack((s, [[a]], [[r]], s_next))
self.buffer[self.buffer_length % self.buffer_size, :] = transition
self.buffer_length += 1
def get_transition_batch(self):
indices = np.random.choice(self.buffer_size, size=self.batch_size)
batch = self.buffer[indices, :]
s = batch[:, :self.s_space]
a = batch[:, self.s_space: self.s_space + 1]
r = batch[:, -self.s_space - 1: -self.s_space]
s_next = batch[:, -self.s_space:]
return s, a, r, s_next
def log_loss(self, episode):
self.logger.warning("Episode: {0} | Actor Loss: {1:.2f} | Critic Loss: {2:.2f}".format(episode,
self.actor_loss,
self.critic_loss))
def __build_actor_nn(self, state, scope, trainable=True):
w_init, b_init = tf.random_normal_initializer(.0, .001), tf.constant_initializer(.1)
with tf.variable_scope(scope):
# state is ? * code_count * data_dim.
first_dense = tf.layers.dense(state,
64,
tf.nn.relu,
kernel_initializer=w_init,
bias_initializer=b_init,
trainable=trainable)
action = tf.layers.dense(first_dense,
1,
tf.nn.sigmoid,
kernel_initializer=w_init,
bias_initializer=b_init,
trainable=trainable)
return tf.multiply(action, self.a_space - 1)
@staticmethod
def __build_critic(state, action, scope, trainable=True):
w_init, b_init = tf.random_normal_initializer(.0, .3), tf.constant_initializer(.1)
with tf.variable_scope(scope):
s_first_dense = tf.layers.dense(state,
32,
tf.nn.relu,
kernel_initializer=w_init,
bias_initializer=b_init,
trainable=trainable)
a_first_dense = tf.layers.dense(action,
32,
tf.nn.relu,
kernel_initializer=w_init,
bias_initializer=b_init,
trainable=trainable)
q_value = tf.layers.dense(tf.nn.relu(s_first_dense + a_first_dense),
1,
kernel_initializer=w_init,
bias_initializer=b_init,
trainable=trainable)
return q_value
def main(args):
mode = args.mode
# mode = 'test'
codes = args.codes
# codes = ["AU88", "RB88", "CU88", "AL88"]
# codes = ["T9999"]
market = args.market
# market = 'future'
episode = args.episode
# episode = 2000
# training_data_ratio = 0.5
training_data_ratio = args.training_data_ratio
model_name = os.path.basename(__file__).split('.')[0]
env = Market(codes, start_date="2012-01-01", end_date="2018-01-01", **{
"market": market,
# "use_sequence": True,
"logger": generate_market_logger(model_name),
"training_data_ratio": training_data_ratio,
})
algorithm = Algorithm(tf.Session(config=config), env, env.trader.action_space, env.data_dim, **{
"mode": mode,
"episodes": episode,
"enable_saver": True,
"learning_rate": 0.003,
"enable_summary_writer": True,
"logger": generate_algorithm_logger(model_name),
"save_path": os.path.join(CHECKPOINTS_DIR, "RL", model_name, market, "model"),
"summary_path": os.path.join(CHECKPOINTS_DIR, "RL", model_name, market, "summary"),
})
algorithm.run()
algorithm.eval()
algorithm.plot()
if __name__ == '__main__':
main(model_launcher_parser.parse_args())