-
Notifications
You must be signed in to change notification settings - Fork 1
/
dqn.py
221 lines (184 loc) · 8.22 KB
/
dqn.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
'''
[DQN](https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf) for CartPole env, Atari env (maybe) upcoming
Notes:
- `observation` from atari is of shape (210,160,3), for CartPole is (4,)
- for Breakout, preprocess to 84x84x4
Architecture:
- for CartPole, 2 layers of feedforward with 24 units each
To run on a fresh instance (with numpy + tensorflow + gym installed)
```
export LC_ALL="en_US.UTF-8"
export LC_CTYPE="en_US.UTF-8"
TF_CPP_MIN_LOG_LEVEL="3" python3 dqn.py --env=CartPole-v1
```
'''
import getopt
import random
import sys
from collections import deque, namedtuple
import gym
import numpy as np
import tensorflow as tf
# hyper parameters
REPLAY_MEMORY_SIZE = 100000
REPLAY_START_SIZE = 64
BATCH_SIZE = 64
EPISODES = 1000
EPSILON_DECAY = 0.96
STARTING_EPSILON = 1
FINAL_EPSILON = 0.05
GAMMA = 0.99
EPOCHS_PER_STEP = 1
MODEL_DIR = 'tf_processing/dqn'
VALIDATION_DIR = 'tf_processing/dqn_validation'
Experience = namedtuple('Experience', ['state', 'action', 'reward', 'next_state', 'terminal'])
class ExperienceMemory:
def __init__(self, max_size: int):
self.experiences = deque()
self._max_size = max_size
def store(self, experience: Experience):
self.experiences.append(experience)
if len(self.experiences) > self._max_size:
self.experiences.popleft()
def size(self):
return len(self.experiences)
def sample(self, batch_size: int):
return random.sample(self.experiences, k=batch_size)
class DQNAgent:
def __init__(self, env_name: str):
self.env = gym.make(env_name)
self.experience_memory = ExperienceMemory(REPLAY_MEMORY_SIZE)
self._classifier = tf.estimator.Estimator(
model_fn=self.model_fn_nn,
model_dir=MODEL_DIR,
config=tf.estimator.RunConfig(session_config=tf.ConfigProto(log_device_placement=False)))
if env_name == 'CartPole-v1':
self.OBSERVATION_SPACE_N = 4
else:
self.OBSERVATION_SPACE_N = None
@staticmethod
def model_fn_nn(features, labels, mode):
'''
Simple feedforward model for CartPole
'''
input_layer = tf.reshape(features['x'], [-1, 4]) # hard code 4 is ok since this is only used for CartPole
dense1 = tf.layers.dense(inputs=input_layer, units=24, activation=tf.nn.relu)
dense2 = tf.layers.dense(inputs=dense1, units=24, activation=tf.nn.relu)
qs = tf.layers.dense(inputs=dense2, units=2)
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode=mode, predictions=qs)
if mode == tf.estimator.ModeKeys.TRAIN:
loss = tf.losses.mean_squared_error(labels=labels, predictions=qs)
tf.summary.scalar('loss', loss)
optimiser = tf.train.AdamOptimizer(learning_rate=0.0005)
train_op = optimiser.minimize(loss=loss, global_step=tf.train.get_global_step())
return tf.estimator.EstimatorSpec(mode=mode, loss=loss,
train_op=train_op)
def predict_q_values(self, states):
input_fn = tf.estimator.inputs.numpy_input_fn(
x={'x': states}, batch_size=states.shape[0],
shuffle=False)
return self._classifier.predict(input_fn)
def action(self, state, epsilon):
q_values = list(self.predict_q_values(np.reshape(state, (1, self.OBSERVATION_SPACE_N))))
assert len(q_values) == 1
if random.random() <= epsilon:
return random.choice(range(0, self.env.action_space.n))
else:
assert q_values[0].shape[0] == 2
return np.argmax(q_values[0])
def train(self, x, y):
input_fn = tf.estimator.inputs.numpy_input_fn(x={'x': x}, y=y, batch_size=BATCH_SIZE,
num_epochs=EPOCHS_PER_STEP,
shuffle=True)
self._classifier.train(input_fn, steps=1)
def main(args):
ops, _ = getopt.getopt(args[1:], '', longopts=['env='])
env_name = ops[0][1]
assert env_name in ['CartPole-v1', 'Breakout-v0']
agent = DQNAgent(env_name)
# first we accumulate experience under uniformly random policy
state = agent.env.reset()
while agent.experience_memory.size() < REPLAY_START_SIZE:
action = agent.action(state, epsilon=1)
next_state, reward, is_done, info = agent.env.step(action)
reward_to_record = -999 if is_done else reward
experience = Experience(state, action, reward_to_record, next_state, is_done)
agent.experience_memory.store(experience)
state = next_state
if is_done:
state = agent.env.reset()
print("Starting experience collected")
epsilon = STARTING_EPSILON
# then we start training
for i in range(1, EPISODES + 1):
state = agent.env.reset()
is_done = False
rewards = 0
if i % 10 == 0:
# validation episode
total_rewards = 0
for j in range(100):
state = agent.env.reset()
is_done = False
rewards = 0
while not is_done:
action = agent.action(state, epsilon=FINAL_EPSILON)
next_state, reward, is_done, info = agent.env.step(action)
rewards += reward
state = next_state
total_rewards += rewards
avg_reward = total_rewards / 5.
summary_writer = tf.summary.FileWriter(VALIDATION_DIR)
summary = tf.Summary()
summary.value.add(tag='validation_reward', simple_value=avg_reward)
summary_writer.add_summary(summary, global_step=i)
summary_writer.flush()
print("Validation episode %d avg reward %.2f" % (i, avg_reward))
epsilon = max(FINAL_EPSILON, epsilon * EPSILON_DECAY)
if avg_reward > 195:
print("Success! Stopping..")
sys.exit()
else:
while not is_done:
action = agent.action(state, epsilon=epsilon)
next_state, reward, is_done, info = agent.env.step(action)
rewards += reward
reward_to_record = -999 if is_done else reward
experience = Experience(state, action, reward_to_record, next_state, is_done)
agent.experience_memory.store(experience)
state = next_state
# SGD
batch = agent.experience_memory.sample(BATCH_SIZE)
# prepare x and y
# t = time.monotonic()
x = np.reshape([e.state for e in batch], (BATCH_SIZE, agent.OBSERVATION_SPACE_N))
next_states = np.reshape([e.next_state for e in batch], (BATCH_SIZE, agent.OBSERVATION_SPACE_N))
next_states_q_values = list(agent.predict_q_values(next_states))
curr_state_q_values = list(agent.predict_q_values(x))
assert (len(next_states_q_values) == BATCH_SIZE)
assert (len(curr_state_q_values) == BATCH_SIZE)
y = []
for idx, e in enumerate(batch):
next_qs = next_states_q_values[idx]
curr_qs = curr_state_q_values[idx]
# debug to make sure batch predict preserves input order
# print(next_qs)
# print(next(agent.predict_q_values(np.reshape(e.next_state, (1,4)))))
# print(curr_qs)
# print(next(agent.predict_q_values(np.reshape(e.state, (1, 4)))))
# IMPORTANT!
if e.terminal:
curr_qs[e.action] = e.reward
else:
curr_qs[e.action] = e.reward + GAMMA * np.max(next_qs)
y.append(curr_qs)
y = np.reshape(np.asarray(y), (BATCH_SIZE, 2))
# print("====== step")
agent.train(x, y)
print("Episode %d: total rewards = %d, epsilon = %.2f, size of replay memory %d" % (
i, rewards, epsilon, agent.experience_memory.size()))
with open('logs/dqn.csv', 'a') as f:
f.write("%d,%d,%.2f\n" % (i, rewards, epsilon))
if __name__ == '__main__':
tf.app.run()