-
Notifications
You must be signed in to change notification settings - Fork 5
/
brain.py
245 lines (187 loc) · 8.63 KB
/
brain.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
import csv
import scipy as sp
import scipy.stats
import tensorflow as tf
import gym
import numpy as np
from collections import deque
import random
class Model:
def __init__(self, env, learning_rate, memory):
self.env = env
self.num_actions = env.action_space.n
self.num_env_space = len(env.observation_space.high)
self.memory = memory
self.input = tf.placeholder(tf.float32, shape=[None, self.num_env_space], name='input')
self.actions = tf.placeholder(tf.float32, shape=[None, self.num_actions], name='actions')
self.rewards = tf.placeholder(tf.float32, shape=[None], name='rewards')
init = tf.truncated_normal_initializer()
# create the network
net = self.input
net = tf.layers.dense(inputs=net, units=100, activation=tf.nn.relu, kernel_initializer=init, name='dense1')
net = tf.layers.dense(inputs=net, units=100, activation=tf.nn.relu, kernel_initializer=init, name='dense2')
net = tf.layers.dense(inputs=net, units=100, activation=tf.nn.relu, kernel_initializer=init, name='dense3')
net = tf.layers.dense(inputs=net, units=self.num_actions, activation=None, kernel_initializer=init)
self.output = net
q_reward = tf.reduce_sum(tf.multiply(self.output, self.actions), 1)
loss = tf.reduce_mean(tf.squared_difference(self.rewards, q_reward))
self.optimiser = tf.train.AdamOptimizer(learning_rate).minimize(loss)
self.session = tf.Session()
self.session.run(tf.global_variables_initializer())
self.saver = tf.train.Saver()
def train(self):
if len(self.memory.memory) < self.memory.batch_size:
return
states, actions, rewards = self.memory.get_batch(self, self.memory.batch_size)
self.session.run(self.optimiser, feed_dict={self.input: states, self.actions: actions, self.rewards: rewards})
def save_model(self, game_number):
print 'saving model'
self.saver.save(self.session, "./model_save_mc_v0/mountaincar-v0-game-", global_step=game_number)
def read_model(self):
checkpoint = tf.train.get_checkpoint_state('./model_save_mc_v0/')
if checkpoint and checkpoint.model_checkpoint_path:
saver = self.saver
sess = self.session
saver.restore(sess, checkpoint.model_checkpoint_path)
print("Successfully loaded:", checkpoint.model_checkpoint_path)
else:
print("Could not find old network weights")
def get_qvalues(self, state_list):
qValues = self.session.run(self.output, feed_dict={self.input: state_list})
return qValues
def get_action(self, state, epsilon):
qValues = self.session.run(self.output, feed_dict={self.input: [state]})[0]
if random.random() < epsilon:
action = self.env.action_space.sample()
else:
action = np.argmax(qValues)
return action, qValues
def get_batch_action(self, states):
return self.session.run(self.output, feed_dict={self.input: states})
class ReplayMemory:
def __init__(self, env, batch_size, max_memory_size, gamma):
self.env = env
self.batch_size = batch_size
self.memory = deque(maxlen=max_memory_size)
self.gamma = gamma
def add(self, state, action, reward, done, next_state):
actions = np.zeros(self.env.action_space.n)
actions[action] = 1
self.memory.append([state, actions, reward, done, next_state])
def get_batch(self, model, batch_size=50):
mini_batch = random.sample(self.memory, batch_size)
states = [item[0] for item in mini_batch]
actions = [item[1] for item in mini_batch]
rewards = [item[2] for item in mini_batch]
done = [item[3] for item in mini_batch]
next_states = [item[4] for item in mini_batch]
q_values = model.get_batch_action(next_states)
y_batch = []
for i in range(batch_size):
if done[i]:
y_batch.append(rewards[i])
else:
y_batch.append(rewards[i] + self.gamma * np.max(q_values[i]))
return states, actions, y_batch
class Agent:
def __init__(self):
# Hyper-parameters
self.gamma = 0.97
self.learning_rate = 1e-3
self.epsilon = 1.
self.final_epsilon = .05
self.epsilon_decay = .995
# Memory parameters
self.batch_size = 50
self.max_memory_size = 10000
self.env = gym.make('MountainCar-v0')
self.memory = ReplayMemory(self.env, self.batch_size, self.max_memory_size, self.gamma)
self.model = Model(self.env, self.learning_rate, self.memory)
self.max_episodes = 1000
self.render = False
def train(self):
for i in range(self.max_episodes):
current_state = self.env.reset()
done = False
count = 0
total_reward = 0
observation_list = []
while not done:
if self.render:
self.env.render()
action, _ = self.model.get_action(current_state, self.epsilon)
next_state, reward, done, _ = self.env.step(action)
observation_list.append(current_state.tolist() + [action])
self.memory.add(current_state, action, reward, done, next_state)
current_state = next_state
total_reward += reward
self.model.train()
count += 1
print('TRAIN: The episode ' + str(i) + ' lasted for ' + str(
count) + ' time steps with epsilon ' + str(self.epsilon))
self.save_csv_all_observations(observation_list, './record_training_observations.csv')
# if i > 200:
if self.epsilon > self.final_epsilon:
self.epsilon *= self.epsilon_decay
# self.model.save_model(i)
def test(self):
self.model.read_model()
test_epsilon = 0
total_reward_list = []
for i in range(100):
observation = self.env.reset()
done = False
count = 0
total_reward = 0
record_transition = []
while not done:
if self.render:
self.env.render()
action, qValues = self.model.get_action(observation, test_epsilon)
newObservation, reward, done, _ = self.env.step(action)
if done:
pass
observation_str = ''
for feature in observation:
observation_str += str(feature) + '$'
observation_str = observation_str[:-1]
newObservation_str = ''
for feature in newObservation:
newObservation_str += str(feature) + '$'
newObservation_str = newObservation_str[:-1]
record_transition.append(
{'observation': observation_str, 'action': action, 'qValue': qValues[action], 'reward': reward,
'newObservation': newObservation_str})
observation = newObservation
total_reward += reward
count += 1
total_reward_list.append(total_reward)
print('TRAIN: The episode ' + str(i) + ' lasted for ' + str(
count) + ' time steps with epsilon ' + str(test_epsilon))
m, h = self.mean_confidence_interval(total_reward_list)
print 'DRL mean:{0}, interval:{1}'.format(str(m), str(h))
# self.save_csv_all_correlations(record_transition,
# './save_all_transition_e{1}/record_moutaincar_transition_game{0}.csv'.format(
# int(i), str(test_epsilon)))
# def save_csv_all_correlations(self, record_transition, csv_name):
# with open(csv_name, 'a') as csvfile:
# writer = csv.DictWriter(csvfile, fieldnames=record_transition[0].keys())
# writer.writeheader()
#
# for row_dict in record_transition:
# writer.writerow(row_dict)
def save_csv_all_observations(self, record_observation, csv_name):
with open(csv_name, "a") as output:
writer = csv.writer(output, lineterminator='\n')
for val in record_observation:
writer.writerow(val)
def mean_confidence_interval(self, data, confidence=0.95):
a = 1.0 * np.array(data)
n = len(a)
m, se = np.mean(a), scipy.stats.sem(a)
h = se * sp.stats.t._ppf((1 + confidence) / 2., n - 1)
return m, h
if __name__ == "__main__":
agent = Agent()
agent.train()
# agent.test()