-
Notifications
You must be signed in to change notification settings - Fork 169
/
player.py
287 lines (236 loc) · 10.9 KB
/
player.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
from _asyncio import Future
from asyncio.queues import Queue
from collections import defaultdict, namedtuple
from logging import getLogger
import asyncio
import numpy as np
from numpy.random import random
from reversi_zero.agent.api import ReversiModelAPI
from reversi_zero.config import Config
from reversi_zero.env.reversi_env import ReversiEnv, Player, Winner
from reversi_zero.lib.bitboard import find_correct_moves, bit_to_array, flip_vertical, rotate90
CounterKey = namedtuple("CounterKey", "black white next_player")
QueueItem = namedtuple("QueueItem", "state future")
HistoryItem = namedtuple("HistoryItem", "action policy values visit enemy_values enemy_visit")
logger = getLogger(__name__)
class ReversiPlayer:
def __init__(self, config: Config, model, play_config=None, enable_resign=True):
"""
:param config:
:param reversi_zero.agent.model.ReversiModel model:
"""
self.config = config
self.model = model
self.play_config = play_config or self.config.play
self.enable_resign = enable_resign
self.api = ReversiModelAPI(self.config, self.model)
# key=(own, enemy, action)
self.var_n = defaultdict(lambda: np.zeros((64,)))
self.var_w = defaultdict(lambda: np.zeros((64,)))
self.var_q = defaultdict(lambda: np.zeros((64,)))
self.var_u = defaultdict(lambda: np.zeros((64,)))
self.var_p = defaultdict(lambda: np.zeros((64,)))
self.expanded = set()
self.now_expanding = set()
self.prediction_queue = Queue(self.play_config.prediction_queue_size)
self.sem = asyncio.Semaphore(self.play_config.parallel_search_num)
self.moves = []
self.loop = asyncio.get_event_loop()
self.running_simulation_num = 0
self.thinking_history = {} # for fun
self.resigned = False
def action(self, own, enemy):
"""
:param own: BitBoard
:param enemy: BitBoard
:return: action: move pos=0 ~ 63 (0=top left, 7 top right, 63 bottom right)
"""
env = ReversiEnv().update(own, enemy, Player.black)
key = self.counter_key(env)
for tl in range(self.play_config.thinking_loop):
if tl > 0 and self.play_config.logging_thinking:
logger.debug(f"continue thinking: policy move=({action % 8}, {action // 8}), "
f"value move=({action_by_value % 8}, {action_by_value // 8})")
self.search_moves(own, enemy)
policy = self.calc_policy(own, enemy)
action = int(np.random.choice(range(64), p=policy))
action_by_value = int(np.argmax(self.var_q[key] + (self.var_n[key] > 0)*100))
if action == action_by_value or env.turn < self.play_config.change_tau_turn or env.turn <= 1:
break
# this is for play_gui, not necessary when training.
next_key = self.get_next_key(own, enemy, action)
self.thinking_history[(own, enemy)] = HistoryItem(action, policy, list(self.var_q[key]), list(self.var_n[key]),
list(self.var_q[next_key]), list(self.var_n[next_key]))
if self.play_config.resign_threshold is not None and \
np.max(self.var_q[key] - (self.var_n[key] == 0)*10) <= self.play_config.resign_threshold:
self.resigned = True
if self.enable_resign:
return None # means resign
self.moves.append([(own, enemy), list(policy)])
return action
def get_next_key(self, own, enemy, action):
env = ReversiEnv().update(own, enemy, Player.black)
env.step(action)
return self.counter_key(env)
def ask_thought_about(self, own, enemy) -> HistoryItem:
return self.thinking_history.get((own, enemy))
def search_moves(self, own, enemy):
loop = self.loop
self.running_simulation_num = 0
coroutine_list = []
for it in range(self.play_config.simulation_num_per_move):
cor = self.start_search_my_move(own, enemy)
coroutine_list.append(cor)
coroutine_list.append(self.prediction_worker())
loop.run_until_complete(asyncio.gather(*coroutine_list))
async def start_search_my_move(self, own, enemy):
self.running_simulation_num += 1
with await self.sem: # reduce parallel search number
env = ReversiEnv().update(own, enemy, Player.black)
leaf_v = await self.search_my_move(env, is_root_node=True)
self.running_simulation_num -= 1
return leaf_v
async def search_my_move(self, env: ReversiEnv, is_root_node=False):
"""
Q, V is value for this Player(always black).
P is value for the player of next_player (black or white)
:param env:
:param is_root_node:
:return:
"""
if env.done:
if env.winner == Winner.black:
return 1
elif env.winner == Winner.white:
return -1
else:
return 0
key = self.counter_key(env)
while key in self.now_expanding:
await asyncio.sleep(self.config.play.wait_for_expanding_sleep_sec)
# is leaf?
if key not in self.expanded: # reach leaf node
leaf_v = await self.expand_and_evaluate(env)
if env.next_player == Player.black:
return leaf_v # Value for black
else:
return -leaf_v # Value for white == -Value for black
action_t = self.select_action_q_and_u(env, is_root_node)
_, _ = env.step(action_t)
virtual_loss = self.config.play.virtual_loss
self.var_n[key][action_t] += virtual_loss
self.var_w[key][action_t] -= virtual_loss
leaf_v = await self.search_my_move(env) # next move
# on returning search path
# update: N, W, Q, U
n = self.var_n[key][action_t] = self.var_n[key][action_t] - virtual_loss + 1
w = self.var_w[key][action_t] = self.var_w[key][action_t] + virtual_loss + leaf_v
self.var_q[key][action_t] = w / n
return leaf_v
async def expand_and_evaluate(self, env):
"""expand new leaf
update var_p, return leaf_v
:param ReversiEnv env:
:return: leaf_v
"""
key = self.counter_key(env)
self.now_expanding.add(key)
black, white = env.board.black, env.board.white
# (di(p), v) = fθ(di(sL))
# rotation and flip. flip -> rot.
is_flip_vertical = random() < 0.5
rotate_right_num = int(random() * 4)
if is_flip_vertical:
black, white = flip_vertical(black), flip_vertical(white)
for i in range(rotate_right_num):
black, white = rotate90(black), rotate90(white) # rotate90: rotate bitboard RIGHT 1 time
black_ary = bit_to_array(black, 64).reshape((8, 8))
white_ary = bit_to_array(white, 64).reshape((8, 8))
state = [black_ary, white_ary] if env.next_player == Player.black else [white_ary, black_ary]
future = await self.predict(np.array(state)) # type: Future
await future
leaf_p, leaf_v = future.result()
# reverse rotate and flip about leaf_p
if rotate_right_num > 0 or is_flip_vertical: # reverse rotation and flip. rot -> flip.
leaf_p = leaf_p.reshape((8, 8))
if rotate_right_num > 0:
leaf_p = np.rot90(leaf_p, k=rotate_right_num) # rot90: rotate matrix LEFT k times
if is_flip_vertical:
leaf_p = np.flipud(leaf_p)
leaf_p = leaf_p.reshape((64, ))
self.var_p[key] = leaf_p # P is value for next_player (black or white)
self.expanded.add(key)
self.now_expanding.remove(key)
return float(leaf_v)
async def prediction_worker(self):
"""For better performance, queueing prediction requests and predict together in this worker.
speed up about 45sec -> 15sec for example.
:return:
"""
q = self.prediction_queue
margin = 10 # avoid finishing before other searches starting.
while self.running_simulation_num > 0 or margin > 0:
if q.empty():
if margin > 0:
margin -= 1
await asyncio.sleep(self.config.play.prediction_worker_sleep_sec)
continue
item_list = [q.get_nowait() for _ in range(q.qsize())] # type: list[QueueItem]
# logger.debug(f"predicting {len(item_list)} items")
data = np.array([x.state for x in item_list])
policy_ary, value_ary = self.api.predict(data)
for p, v, item in zip(policy_ary, value_ary, item_list):
item.future.set_result((p, v))
async def predict(self, x):
future = self.loop.create_future()
item = QueueItem(x, future)
await self.prediction_queue.put(item)
return future
def finish_game(self, z):
"""
:param z: win=1, lose=-1, draw=0
:return:
"""
for move in self.moves: # add this game winner result to all past moves.
move += [z]
def calc_policy(self, own, enemy):
"""calc π(a|s0)
:param own:
:param enemy:
:return:
"""
pc = self.play_config
env = ReversiEnv().update(own, enemy, Player.black)
key = self.counter_key(env)
if env.turn < pc.change_tau_turn:
return self.var_n[key] / np.sum(self.var_n[key]) # tau = 1
else:
action = np.argmax(self.var_n[key]) # tau = 0
ret = np.zeros(64)
ret[action] = 1
return ret
@staticmethod
def counter_key(env: ReversiEnv):
return CounterKey(env.board.black, env.board.white, env.next_player.value)
def select_action_q_and_u(self, env, is_root_node):
key = self.counter_key(env)
if env.next_player == Player.black:
legal_moves = find_correct_moves(key.black, key.white)
else:
legal_moves = find_correct_moves(key.white, key.black)
# noinspection PyUnresolvedReferences
xx_ = np.sqrt(np.sum(self.var_n[key])) # SQRT of sum(N(s, b); for all b)
xx_ = max(xx_, 1) # avoid u_=0 if N is all 0
p_ = self.var_p[key]
if is_root_node: # Is it correct?? -> (1-e)p + e*Dir(0.03)
p_ = (1 - self.play_config.noise_eps) * p_ + \
self.play_config.noise_eps * np.random.dirichlet([self.play_config.dirichlet_alpha] * 64)
u_ = self.play_config.c_puct * p_ * xx_ / (1 + self.var_n[key])
if env.next_player == Player.black:
v_ = (self.var_q[key] + u_ + 1000) * bit_to_array(legal_moves, 64)
else:
# When enemy's selecting action, flip Q-Value.
v_ = (-self.var_q[key] + u_ + 1000) * bit_to_array(legal_moves, 64)
# noinspection PyTypeChecker
action_t = int(np.argmax(v_))
return action_t