-
Notifications
You must be signed in to change notification settings - Fork 0
/
QNetwork.py
259 lines (181 loc) · 6.96 KB
/
QNetwork.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
import random
import numpy as np
from PIL import Image
from QGraph import QGraph
from Memory import Memory
from Minibatch import MiniBatch
from QTargetGraph import QTargetGraph
class QNetwork(object):
def __init__(self, bTrain):
# Settings
self.directory = '/tmp/TrainedQNetwork'
self.num_actions = 9
self.im_height = 84
self.im_width = 84
self.discount_factor = 0.99
self.minibatch_size = 32
self.initial_epsilon = 1.0
self.final_epsilon = 0.1
self.epsilon_frames = 1000000
self.replay_start_size = 50000
self.policy_start_size = self.replay_start_size
self.k = 4 # action repeat (frame skipping)
self.u = 4 # update frequency
self.m = 4 # number of frames to include in sequence
self.c = 10000 # number of actions selected before updating the network used to generate the targets
# Internal Variables
self.bTrain = bTrain
self.ki = 0
self.ui = 0
self.mi = 0
self.frame = 0
self.ci = 0
self.sequence = []
self.prev_phi = np.array([])
self.phi = np.array([])
self.epsilon_increment = (self.initial_epsilon - self.final_epsilon) / self.epsilon_frames
self.epsilon = self.initial_epsilon
self.action = 0
self.reward = 0
self.memory = Memory()
self.minibatch = MiniBatch()
self.targets = np.zeros(self.minibatch_size)
self.bTrial_over = False
self.bStartLearning = False
self.bStartPolicy = False
self.ti = 0
random.seed(0)
# Construct tensorflow graphs
self.q_graph = QGraph(self.im_width, self.im_height, self.m, self.num_actions, self.directory)
if (self.bTrain):
self.q_graph.SaveGraphAndVariables()
self.q_graph_targets = QTargetGraph(self.im_width, self.im_height, self.m, self.num_actions, self.directory)
else:
self.q_graph.LoadGraphAndVariables()
return
def Update(self, pxarray, bTrial_over, reward):
self.frame += 1
self.ki +=1
self.bTrial_over = bTrial_over
if(self.bTrain):
self.Train(pxarray=pxarray, bTrial_over=bTrial_over, reward=reward)
else:
self.Test(pxarray=pxarray, bTrial_over=bTrial_over)
return self.action
def Train(self, pxarray, bTrial_over, reward):
if (self.frame == self.replay_start_size):
print('Starting Learning...')
self.bStartLearning = True
if (self.frame == self.policy_start_size):
print('Starting Policy...')
self.bStartPolicy = True
self.reward += reward
if (self.ki >= self.k):
self.PreprocessSequence(pxarray)
self.mi += 1
self.ki = 0
if(self.mi == self.m):
self.prev_phi = np.copy(self.phi)
self.reward = 0
elif(self.mi >= self.m + 1):
self.ui += 1
self.StoreExperience()
if (self.bStartLearning):
if(self.ui >= self.u):
self.SampleRandomMinibatch()
self.GenerateTargets()
self.GradientDescentStep()
self.ui = 0
if (not self.bStartPolicy):
self.SelectRandomAction()
else:
self.UpdateAction()
self.prev_phi = np.copy(self.phi)
self.reward = 0
if(bTrial_over):
self.mi = 0
self.ki = 0
self.ui = 0
if (self.bStartPolicy):
if(self.epsilon > self.final_epsilon):
self.epsilon -= self.epsilon_increment
return
def Test(self, pxarray, bTrial_over):
if (self.ki >= self.k):
self.PreprocessSequence(pxarray)
self.mi += 1
self.ki = 0
if (self.mi >= self.m):
self.SelectAction()
if (bTrial_over):
self.mi = 0
self.ki = 0
return
def PreprocessSequence(self, pxarray):
img = Image.fromarray(pxarray)
img = img.resize([self.im_width, self.im_height])
img_grey = img.convert('LA')
img_grey = np.array(img_grey).astype(np.uint8)
img = img_grey[:, :, 0]
self.sequence.append(img)
if (self.sequence.__len__() > self.m):
del self.sequence[0]
self.phi = np.stack(self.sequence, axis=-1)
self.phi = np.expand_dims(self.phi, axis=0)
return
def UpdateAction(self):
# e-greedy policy
if (random.random() <= self.epsilon):
self.SelectRandomAction()
else:
self.SelectAction()
return
def SelectAction(self):
action_values = np.array(self.q_graph.GetActionValues(self.phi))
num = np.argmax(action_values)
self.action = num
return
def SelectRandomAction(self):
num = random.randint(0, self.num_actions)
self.action = num
return
def StoreExperience(self):
self.memory.RecordExperience(self.prev_phi, self.phi, self.action, self.reward, self.bTrial_over)
return
def SampleRandomMinibatch(self):
self.minibatch = self.memory.GetMinibatch(self.minibatch_size)
return
def GenerateTargets(self):
self.ci += 1
if (self.ti == 0):
self.example_minibatch = self.minibatch
# refresh target network
if(self.ci >= self.c):
print('Loading New Target Graph...')
self.ci = 0
self.q_graph.SaveGraphAndVariables()
self.q_graph_targets = QTargetGraph(self.im_width, self.im_height, self.m, self.num_actions, self.directory)
self.targets = self.GetTargets(self.minibatch)
if(self.ti % 2500 == 0):
print('Example Actions:')
print(self.example_minibatch.actions)
print('Example Targets:')
print(self.GetTargets(self.example_minibatch))
print('Phi Example Action Values Q:')
print(self.q_graph.GetActionValues(self.example_minibatch.prev_phis))
self.ti += 1
return
def GetTargets(self, minibatch):
targets = np.zeros(minibatch.rewards.__len__())
action_values = np.array(self.q_graph_targets.GetActionValues(minibatch.phis))
action_values = np.squeeze(action_values)
max_action_values = np.amax(action_values, axis=1)
for i in range(targets.__len__()):
if (not minibatch.bTrial_over[i]):
targets[i] = minibatch.rewards[i] + (max_action_values[i] * self.discount_factor)
else:
targets[i] = minibatch.rewards[i]
return targets
def GradientDescentStep(self):
self.q_graph.GradientDescentStep(self.minibatch.prev_phis, self.minibatch.actions, self.targets)
return