-
Notifications
You must be signed in to change notification settings - Fork 0
/
Q_Learning_Eric3.py
318 lines (275 loc) · 13.2 KB
/
Q_Learning_Eric3.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
306
307
308
309
310
311
312
313
314
315
316
317
318
from SOM import SOM
import cv2
import numpy as np
import gym
import universe
import math
import scipy.misc
import random
import queue
import time
# track reward probably
import SOM
class QLearner():
def __init__(self, width, height, action_list, random_threshold=1.0, exploration_threshold=60, decay_rate=.95, learning_rate_constant=10.0, random_floor=0.05, learning_floor=0.1):
self.width = width
self.height = height
self.action_list = action_list
self.random_threshold = random_threshold
self.exploration_threshold = exploration_threshold
self.random_floor = random_floor
self.learning_rate_constant = learning_rate_constant
self.decay_rate = decay_rate
self.q_table = [[[0.0 for q in range(len(action_list))] for i in range(height)] for j in range(width)]
self.learning_rate = 1.0
self.discount_factor = .5
self.times_chosen = [[[0 for q in range(len(action_list))] for i in range(height)] for j in range(width)]
self.learning_floor = learning_floor
def select_action(self, x, y):
saved_index = 0
max_reward = 0
# exploration
if random.random() < self.random_threshold:
return random.choice(self.action_list)
for times_chosen in range(len(self.times_chosen[x][y])):
if self.times_chosen[x][y][times_chosen] < self.exploration_threshold:
return self.action_list[times_chosen]
# exploitation
for action_reward_index in range(len(self.q_table[x][y])):
if self.q_table[x][y][action_reward_index] > max_reward:
max_reward = self.q_table[x][y][action_reward_index]
saved_index = action_reward_index
return self.action_list[saved_index]
def update_qtable(self, action, prev_state_w, prev_state_h, state_w, state_h, prev_reward):
action_i = self.action_list.index(action)
learn_amount = 1.0 / (1 + (self.times_chosen[prev_state_w][prev_state_h][action_i]) / self.learning_rate_constant)
if learn_amount < self.learning_floor:
learning_rate = self.learning_floor
else:
learning_rate = learn_amount
# learning_rate = 1.0 / (1 + (self.times_chosen[prev_state_w][prev_state_h][action_i]) / self.learning_rate_constant) # higher the constant, higher the learning rate
self.q_table[prev_state_w][prev_state_h][action_i] = (1.0 - learning_rate) * \
self.q_table[prev_state_w][prev_state_h][action_i] + \
learning_rate * (self.discount_factor *
self.max_reward(state_w, state_h) + prev_reward)
self.times_chosen[prev_state_w][prev_state_h][action_i] += 1
if self.random_threshold == self.random_floor or self.random_threshold < self.random_floor:
self.random_threshold = self.random_floor
else:
self.random_threshold *= self.decay_rate
def max_reward(self, state_w, state_h):
return max(self.q_table[state_w][state_h])
def shorten_buffers_to_one(list_of_queues):
for buffer_i in range(len(list_of_queues)):
list_of_queues[buffer_i] = queue.Queue(2)
def main():
# setup
env = gym.make('internet.SlitherIO-v0')
env.configure(remotes=1, fps=5)
observation_n = env.reset()
tiny_image_h = 20
tiny_image_w = 30 # for tiny image
SOM_WIDTH = 9
SOM_HEIGHT = 9
buffer_threshold = 10
som = SOM.SOM(SOM_WIDTH, SOM_HEIGHT, tiny_image_w, tiny_image_h, learning_rate=1, decay_rate=.95, radius=3)
time_to_switch_to_live = 60 * 5
live_bool = False
## Define Actions on keyboard
left = [universe.spaces.PointerEvent(30,240,0)]
right = [universe.spaces.PointerEvent(515, 240, 0)]
up = [universe.spaces.PointerEvent(275, 95, 0)]
down = [universe.spaces.PointerEvent(275,380,0)]
boost_left = [universe.spaces.PointerEvent(30, 240, 1)]
boost_right = [universe.spaces.PointerEvent(515, 240, 1)]
boost_up = [universe.spaces.PointerEvent(275, 95, 1)]
boost_down = [universe.spaces.PointerEvent(275, 380, 1)]
# left_boost = [('KeyEvent', 'ArrowUp', True), ('KeyEvent', 'ArrowLeft', True), ('KeyEvent', 'ArrowRight', False)]
# left = [('KeyEvent', 'ArrowUp', False), ('KeyEvent', 'ArrowLeft', True), ('KeyEvent', 'ArrowRight', False)]
# right = [('KeyEvent', 'ArrowUp', False), ('KeyEvent', 'ArrowLeft', False), ('KeyEvent', 'ArrowRight', True)]
# right_boost = [('KeyEvent', 'ArrowUp', True), ('KeyEvent', 'ArrowLeft', False), ('KeyEvent', 'ArrowRight', True)]
# forward = [('KeyEvent', 'ArrowUp', True), ('KeyEvent', 'ArrowLeft', False), ('KeyEvent', 'ArrowRight', False)]
# still = [('KeyEvent', 'ArrowUp', False), ('KeyEvent', 'ArrowLeft', False), ('KeyEvent', 'ArrowRight', False)]
# possible_actions.append(still)
# possible_actions.append(forward)
possible_actions = [left, right, up, down, boost_down, boost_right, boost_up, boost_left]
bad_actions = [boost_right, boost_left, boost_up, boost_left]
# possible_actions.append(right_boost)
# possible_actions.append(left_boost)
# bad_actions = [left_boost, right_boost, forward]
q_learner = QLearner(SOM_WIDTH, SOM_HEIGHT, possible_actions)
frame_buffer = queue.Queue(buffer_threshold * 2)
action_buffer =queue.Queue(buffer_threshold * 2)
rewards_buffer = queue.Queue(buffer_threshold * 2)
list_of_buffers = [frame_buffer,action_buffer,rewards_buffer]
trials = []
reward_total = 0.0
action = up
state_w = 0
prev_state_w = None
prev_state_h = None
last_image = None
start_time = time.time()
reward = 0
reward_decay = 0.8
while True:
if observation_n[0] != None :
if time.time() - start_time > time_to_switch_to_live and not live_bool:
live_bool = True
shorten_buffers_to_one(list_of_buffers)
buffer_threshold = 1
if info['n'][0]["env_status.env_state"] == "running" and reward_n[0] is not None:
frame_buffer.put_nowait(crop(observation_n[0]["vision"]))
action_buffer.put_nowait(action)
rewards_buffer.put_nowait(reward_n[0])
reward_total += reward_n[0]
if frame_buffer.qsize() > buffer_threshold:
current_frame = frame_buffer.get_nowait()
current_action = action_buffer.get_nowait()
current_reward = rewards_buffer.get_nowait()
action_n = [action for ob in observation_n]
last_image = current_frame
image_for_som = process_image(last_image, tiny_image_w, tiny_image_h)
state_w, state_h = som.train(image_for_som)
if prev_state_w is None:
prev_state_w = state_w
prev_state_h = state_h
action = q_learner.select_action(state_w, state_h)
else:
reward *= reward_decay
reward += reward_n[0]
default_reward = reward
# if current_action in bad_actions:
# default_reward = reward_n[0]
# else:
# default_reward = 5 + reward_n[0]
'''if current_action in bad_actions:
if current_reward < 0:
default_reward = current_reward
else:
default_reward = 0'''
q_learner.update_qtable(current_action, prev_state_w, prev_state_h, state_w, state_h, default_reward)
action = q_learner.select_action(state_w, state_h)
prev_state_w = state_w
prev_state_h = state_h
cv2.imshow("Frame Training", cv2.cvtColor(last_image,cv2.COLOR_RGB2BGR))
else:
# do stuff with death here
if reward_total > 0.0:
trials.append(reward_total)
print(trials)
if frame_buffer.qsize() > 0:
current_frame = frame_buffer.get_nowait()
current_action = action_buffer.get_nowait()
current_reward = rewards_buffer.get_nowait()
last_image = current_frame
image_for_som = process_image(last_image, tiny_image_w, tiny_image_h)
state_w, state_h = som.train(image_for_som)
# The following assumes that prev_state has been defined
q_learner.update_qtable(current_action, prev_state_w, prev_state_h, state_w, state_h, -500)
action = q_learner.select_action(state_w, state_h)
prev_state_w = state_w
prev_state_h = state_h
cv2.imshow("Frame Training", cv2.cvtColor(last_image, cv2.COLOR_RGB2BGR))
else:
action_n = [up for ob in observation_n]
observation_n, reward_n, done_n, info = env.step(action_n)
# print(done_n)
print_som(som)
env.render()
def print_som(som):
columns = []
for w in range(som.width):
array = []
for h in range(som.height):
array.append(som.get(w,h).array)
column = np.vstack(array)
columns.append(column)
actual_array = np.hstack(columns)
# print(actual_array)
enlarged_image = scipy.misc.imresize(actual_array,200,"cubic")
cv2.imshow("SOM.jpg", enlarged_image)
cv2.waitKey(5)
def crop(observation):
top_left_x = 20
top_left_y = 85
bottom_right_x = 520
bottom_right_y = 385
photo_array = crop_photo(observation, top_left_x, top_left_y, bottom_right_x,
bottom_right_y)
return photo_array
def process_image(observation, tiny_image_w, tiny_image_h):
# shrunken_image = scipy.misc.imresize(observation, 0.5, "nearest")
# shrunken_image = cv2.cvtColor(shrunken_image, cv2.COLOR_RGB2BGR)
# shrunken_image2 = cv2.cvtColor(shrunken_image, cv2.COLOR_BGR2GRAY)
# tiny = cv2.resize(fill_contour(shrunken_image2), (tiny_image_w, tiny_image_h))
# return tiny
bgr = cv2.cvtColor(observation, cv2.COLOR_RGB2BGR)
grayscale = cv2.cvtColor(bgr, cv2.COLOR_BGR2GRAY)
filled_orbs = fill_orbs(grayscale.copy())
# filled_orbs = cv2.resize(filled_orbs, (0,0), fx=.5, fy=.5)
# shrunken_image = cv2.resize(grayscale, (0, 0), fx=.5, fy=.5)
filled_orbs = scipy.misc.imresize(filled_orbs, 0.5, "nearest")
shrunken_image = scipy.misc.imresize(grayscale, 0.5, "nearest")
filled_orbs = cv2.cvtColor(filled_orbs, cv2.COLOR_RGB2BGR)
# shrunken_image = cv2.cvtColor(shrunken_image,
tiny = cv2.resize(fill_snake(shrunken_image, filled_orbs), (tiny_image_w, tiny_image_h))
return tiny
def crop_photo(array,top_left_x, top_left_y, bottom_right_x, bottom_right_y):
return array[top_left_y:bottom_right_y,top_left_x:bottom_right_x]
def fill_snake(orig, to_draw):
img = orig.copy()
img = cv2.medianBlur(img, 5)
ret, img = cv2.threshold(img, 60, 255, cv2.THRESH_BINARY)
ret, contour, hier = cv2.findContours(img, cv2.RETR_CCOMP, cv2.CHAIN_APPROX_SIMPLE)
for cnt in contour:
cv2.drawContours(img, [cnt], 0, 255, -1)
detector = cv2.MSER_create()
fs = detector.detect(img)
fs.sort(key=lambda x: -x.size)
sfs = [x for x in fs if not supress(x, fs)]
h, w = orig.shape
final_img = np.zeros((h, w, 3), np.uint8)
for f in sfs:
cv2.circle(to_draw, (int(f.pt[0]), int(f.pt[1])), int(f.size / 2), (0, 255, 0), cv2.FILLED)
cv2.imshow("binary", img)
cv2.imshow("contours", to_draw)
return to_draw
def fill_orbs(orig):
img = orig.copy()
img2 = cv2.medianBlur(img, 5)
ret, img2 = cv2.threshold(img2, 63, 255, cv2.THRESH_BINARY)
h, w = img2.shape
bw_image = np.zeros((h, w, 3), np.uint8)
ret, contour, hier = cv2.findContours(img2, cv2.RETR_CCOMP, cv2.CHAIN_APPROX_SIMPLE)
for cnt in contour:
if len(cnt) >= 5:
ellipse = cv2.fitEllipse(cnt)
(x, y), (MA, ma), angle = ellipse
big = 0.0
small = 0.0
if MA > ma:
big = MA
small = ma
else:
big = ma
small = MA
if small / big > .7 and (math.pi * MA * ma <= 200 and math.pi * MA * ma >= 20):
ellipse = ((x,y),(MA * 4,ma * 4),angle)
cv2.ellipse(bw_image, ellipse, (255, 0, 0), cv2.FILLED)
cv2.ellipse(img, ellipse, (0, 0, 255), 2)
return bw_image
def supress(x, fs):
# leaderboard
if (x.pt[0] > 165.0 and x.pt[0] < 208.0 and x.pt[1] < 15.0):
return True
# mini map
if (x.pt[0] > 195.0 and x.pt[0] < 230.0 and x.pt[1] < 120.0 and x.pt[1] > 98.0):
return True
for f in fs:
distx = f.pt[0] - x.pt[0]
disty = f.pt[1] - x.pt[1]
dist = math.sqrt(distx * distx + disty * disty)
if (f.size > x.size) and (dist < f.size / 2):
return True
main()