-
Notifications
You must be signed in to change notification settings - Fork 41
/
Copy pathdqn_agent_nature.py
305 lines (240 loc) · 11.9 KB
/
dqn_agent_nature.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
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
# -*- coding: utf-8 -*-
"""
Deep Q-network implementation with chainer and rlglue
Copyright (c) 2015 Naoto Yoshida All Right Reserved.
"""
import copy
import pickle
import numpy as np
import scipy.misc as spm
from chainer import cuda, FunctionSet, Variable, optimizers
import chainer.functions as F
from rlglue.agent.Agent import Agent
from rlglue.agent import AgentLoader as AgentLoader
from rlglue.types import Action
class DQN_class:
# Hyper-Parameters
gamma = 0.99 # Discount factor
initial_exploration = 100#10**4 # Initial exploratoin. original: 5x10^4
replay_size = 32 # Replay (batch) size
target_model_update_freq = 10**4 # Target update frequancy. original: 10^4
data_size = 10**5 # Data size of history. original: 10^6
def __init__(self, enable_controller=[0, 3, 4]):
self.num_of_actions = len(enable_controller)
self.enable_controller = enable_controller # Default setting : "Pong"
print "Initializing DQN..."
# Initialization of Chainer 1.1.0 or older.
# print "CUDA init"
# cuda.init()
print "Model Building"
self.model = FunctionSet(
l1=F.Convolution2D(4, 32, ksize=8, stride=4, nobias=False, wscale=np.sqrt(2)),
l2=F.Convolution2D(32, 64, ksize=4, stride=2, nobias=False, wscale=np.sqrt(2)),
l3=F.Convolution2D(64, 64, ksize=3, stride=1, nobias=False, wscale=np.sqrt(2)),
l4=F.Linear(3136, 512, wscale=np.sqrt(2)),
q_value=F.Linear(512, self.num_of_actions,
initialW=np.zeros((self.num_of_actions, 512),
dtype=np.float32))
).to_gpu()
self.model_target = copy.deepcopy(self.model)
print "Initizlizing Optimizer"
self.optimizer = optimizers.RMSpropGraves(lr=0.00025, alpha=0.95, momentum=0.95, eps=0.0001)
self.optimizer.setup(self.model.collect_parameters())
# History Data : D=[s, a, r, s_dash, end_episode_flag]
self.D = [np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros(self.data_size, dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.int8),
np.zeros((self.data_size, 4, 84, 84), dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.bool)]
def forward(self, state, action, Reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
Q = self.Q_func(s) # Get Q-value
# Generate Target Signals
tmp = self.Q_func_target(s_dash) # Q(s',*)
tmp = list(map(np.max, tmp.data.get())) # max_a Q(s',a)
max_Q_dash = np.asanyarray(tmp, dtype=np.float32)
target = np.asanyarray(Q.data.get(), dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = np.sign(Reward[i]) + self.gamma * max_Q_dash[i]
else:
tmp_ = np.sign(Reward[i])
action_index = self.action_to_index(action[i])
target[i, action_index] = tmp_
# TD-error clipping
td = Variable(cuda.to_gpu(target)) - Q # TD error
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td/abs(td_tmp) * (abs(td.data) > 1)
zero_val = Variable(cuda.to_gpu(np.zeros((self.replay_size, self.num_of_actions), dtype=np.float32)))
loss = F.mean_squared_error(td_clip, zero_val)
return loss, Q
def stockExperience(self, time,
state, action, reward, state_dash,
episode_end_flag):
data_index = time % self.data_size
if episode_end_flag is True:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
else:
self.D[0][data_index] = state
self.D[1][data_index] = action
self.D[2][data_index] = reward
self.D[3][data_index] = state_dash
self.D[4][data_index] = episode_end_flag
def experienceReplay(self, time):
if self.initial_exploration < time:
# Pick up replay_size number of samples from the Data
if time < self.data_size: # during the first sweep of the History Data
replay_index = np.random.randint(0, time, (self.replay_size, 1))
else:
replay_index = np.random.randint(0, self.data_size, (self.replay_size, 1))
s_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84), dtype=np.float32)
a_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.uint8)
r_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.float32)
s_dash_replay = np.ndarray(shape=(self.replay_size, 4, 84, 84), dtype=np.float32)
episode_end_replay = np.ndarray(shape=(self.replay_size, 1), dtype=np.bool)
for i in xrange(self.replay_size):
s_replay[i] = np.asarray(self.D[0][replay_index[i]], dtype=np.float32)
a_replay[i] = self.D[1][replay_index[i]]
r_replay[i] = self.D[2][replay_index[i]]
s_dash_replay[i] = np.array(self.D[3][replay_index[i]], dtype=np.float32)
episode_end_replay[i] = self.D[4][replay_index[i]]
s_replay = cuda.to_gpu(s_replay)
s_dash_replay = cuda.to_gpu(s_dash_replay)
# Gradient-based update
self.optimizer.zero_grads()
loss, _ = self.forward(s_replay, a_replay, r_replay, s_dash_replay, episode_end_replay)
loss.backward()
self.optimizer.update()
def Q_func(self, state):
h1 = F.relu(self.model.l1(state / 255.0)) # scale inputs in [0.0 1.0]
h2 = F.relu(self.model.l2(h1))
h3 = F.relu(self.model.l3(h2))
h4 = F.relu(self.model.l4(h3))
Q = self.model.q_value(h4)
return Q
def Q_func_target(self, state):
h1 = F.relu(self.model_target.l1(state / 255.0)) # scale inputs in [0.0 1.0]
h2 = F.relu(self.model_target.l2(h1))
h3 = F.relu(self.model_target.l3(h2))
h4 = F.relu(self.model_target.l4(h3))
Q = self.model_target.q_value(h4)
return Q
def e_greedy(self, state, epsilon):
s = Variable(state)
Q = self.Q_func(s)
Q = Q.data
if np.random.rand() < epsilon:
index_action = np.random.randint(0, self.num_of_actions)
print "RANDOM"
else:
index_action = np.argmax(Q.get())
print "GREEDY"
return self.index_to_action(index_action), Q
def target_model_update(self):
self.model_target = copy.deepcopy(self.model)
def index_to_action(self, index_of_action):
return self.enable_controller[index_of_action]
def action_to_index(self, action):
return self.enable_controller.index(action)
class dqn_agent(Agent): # RL-glue Process
lastAction = Action()
policyFrozen = False
def agent_init(self, taskSpec):
# Some initializations for rlglue
self.lastAction = Action()
self.time = 0
self.epsilon = 1.0 # Initial exploratoin rate
# Pick a DQN from DQN_class
self.DQN = DQN_class() # default is for "Pong".
def agent_start(self, observation):
# Preprocess
tmp = np.bitwise_and(np.asarray(observation.intArray[128:]).reshape([210, 160]), 0b0001111) # Get Intensity from the observation
obs_array = (spm.imresize(tmp, (110, 84)))[110-84-8:110-8, :] # Scaling
# Initialize State
self.state = np.zeros((4, 84, 84), dtype=np.uint8)
self.state[0] = obs_array
state_ = cuda.to_gpu(np.asanyarray(self.state.reshape(1, 4, 84, 84), dtype=np.float32))
# Generate an Action e-greedy
returnAction = Action()
action, Q_now = self.DQN.e_greedy(state_, self.epsilon)
returnAction.intArray = [action]
# Update for next step
self.lastAction = copy.deepcopy(returnAction)
self.last_state = self.state.copy()
self.last_observation = obs_array
return returnAction
def agent_step(self, reward, observation):
# Preproces
tmp = np.bitwise_and(np.asarray(observation.intArray[128:]).reshape([210, 160]), 0b0001111) # Get Intensity from the observation
obs_array = (spm.imresize(tmp, (110, 84)))[110-84-8:110-8, :] # Scaling
obs_processed = np.maximum(obs_array, self.last_observation) # Take maximum from two frames
# Compose State : 4-step sequential observation
self.state = np.asanyarray([self.state[1], self.state[2], self.state[3], obs_processed], dtype=np.uint8)
state_ = cuda.to_gpu(np.asanyarray(self.state.reshape(1, 4, 84, 84), dtype=np.float32))
# Exploration decays along the time sequence
if self.policyFrozen is False: # Learning ON/OFF
if self.DQN.initial_exploration < self.time:
self.epsilon -= 1.0/10**6
if self.epsilon < 0.1:
self.epsilon = 0.1
eps = self.epsilon
else: # Initial Exploation Phase
print "Initial Exploration : %d/%d steps" % (self.time, self.DQN.initial_exploration)
eps = 1.0
else: # Evaluation
print "Policy is Frozen"
eps = 0.05
# Generate an Action by e-greedy action selection
returnAction = Action()
action, Q_now = self.DQN.e_greedy(state_, eps)
returnAction.intArray = [action]
# Learning Phase
if self.policyFrozen is False: # Learning ON/OFF
self.DQN.stockExperience(self.time, self.last_state, self.lastAction.intArray[0], reward, self.state, False)
self.DQN.experienceReplay(self.time)
# Target model update
if self.DQN.initial_exploration < self.time and np.mod(self.time, self.DQN.target_model_update_freq) == 0:
print "########### MODEL UPDATED ######################"
self.DQN.target_model_update()
# Simple text based visualization
print ' Time Step %d / ACTION %d / REWARD %.1f / EPSILON %.6f / Q_max %3f' % (self.time, self.DQN.action_to_index(action), np.sign(reward), eps, np.max(Q_now.get()))
# Updates for next step
self.last_observation = obs_array
if self.policyFrozen is False:
self.lastAction = copy.deepcopy(returnAction)
self.last_state = self.state.copy()
self.time += 1
return returnAction
def agent_end(self, reward): # Episode Terminated
# Learning Phase
if self.policyFrozen is False: # Learning ON/OFF
self.DQN.stockExperience(self.time, self.last_state, self.lastAction.intArray[0], reward, self.last_state, True)
self.DQN.experienceReplay(self.time)
# Target model update
if self.DQN.initial_exploration < self.time and np.mod(self.time, self.DQN.target_model_update_freq) == 0:
print "########### MODEL UPDATED ######################"
self.DQN.target_model_update()
# Simple text based visualization
print ' REWARD %.1f / EPSILON %.5f' % (np.sign(reward), self.epsilon)
# Time count
if self.policyFrozen is False:
self.time += 1
def agent_cleanup(self):
pass
def agent_message(self, inMessage):
if inMessage.startswith("freeze learning"):
self.policyFrozen = True
return "message understood, policy frozen"
if inMessage.startswith("unfreeze learning"):
self.policyFrozen = False
return "message understood, policy unfrozen"
if inMessage.startswith("save model"):
with open('dqn_model.dat', 'w') as f:
pickle.dump(self.DQN.model, f)
return "message understood, model saved"
if __name__ == "__main__":
AgentLoader.loadAgent(dqn_agent())