Repository navigation
Expand file tree
/
Copy pathstrategies.py
More file actions
251 lines (218 loc) · 9.83 KB
/
Copy pathstrategies.py
File metadata and controls
251 lines (218 loc) · 9.83 KB
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
# Copyright 2018 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import copy
import math
import random
import sys
import time
import sgf_wrapper
import coords
import gtp
import numpy as np
from mcts import MCTSNode, MAX_DEPTH
import go
# When to do deterministic move selection. ~30 moves on a 19x19, ~8 on 9x9
TEMPERATURE_CUTOFF = int((go.N * go.N) / 12)
def time_recommendation(move_num, seconds_per_move=5, time_limit=15*60,
decay_factor=0.98):
'''Given current move number and "desired" seconds per move,
return how much time should actually be used. To be used specifically
for CGOS time controls, which are absolute 15 minute time.
The strategy is to spend the maximum time possible using seconds_per_move,
and then switch to an exponentially decaying time usage, calibrated so that
we have enough time for an infinite number of moves.'''
# divide by two since you only play half the moves in a game.
player_move_num = move_num / 2
# sum of geometric series maxes out at endgame_time seconds.
endgame_time = seconds_per_move / (1 - decay_factor)
if endgame_time > time_limit:
# there is so little main time that we're already in "endgame" mode.
base_time = time_limit * (1 - decay_factor)
return base_time * decay_factor ** player_move_num
# leave over endgame_time seconds for the end, and play at seconds_per_move
# for as long as possible
core_time = time_limit - endgame_time
core_moves = core_time / seconds_per_move
if player_move_num < core_moves:
return seconds_per_move
else:
return seconds_per_move * decay_factor ** (player_move_num - core_moves)
class MCTSPlayerMixin:
# If 'simulations_per_move' is nonzero, it will perform that many reads before playing.
# Otherwise, it uses 'seconds_per_move' of wall time'
def __init__(self, network, seconds_per_move=5, simulations_per_move=0,
resign_threshold=-0.90, verbosity=0, two_player_mode=False,
num_parallel=8):
self.network = network
self.seconds_per_move = seconds_per_move
self.simulations_per_move = simulations_per_move
self.verbosity = verbosity
self.two_player_mode = two_player_mode
if two_player_mode:
self.temp_threshold = -1
else:
self.temp_threshold = TEMPERATURE_CUTOFF
self.num_parallel = num_parallel
self.qs = []
self.comments = []
self.searches_pi = []
self.root = None
self.result = 0
self.result_string = None
self.resign_threshold = -abs(resign_threshold)
super().__init__()
def initialize_game(self, position=None):
if position is None:
position = go.Position()
self.root = MCTSNode(position)
self.result = 0
self.result_string = None
self.comments = []
self.searches_pi = []
self.qs = []
def suggest_move(self, position):
''' Used for playing a single game.
For parallel play, use initialize_move, select_leaf,
incorporate_results, and pick_move
'''
start = time.time()
if self.simulations_per_move == 0:
while time.time() - start < self.seconds_per_move:
self.tree_search()
else:
current_readouts = self.root.N
while self.root.N < current_readouts + self.simulations_per_move:
self.tree_search()
if self.verbosity > 0:
print("%d: Searched %d times in %s seconds\n\n" % (
position.n, self.simulations_per_move, time.time() - start), file=sys.stderr)
# print some stats on anything with probability > 1%
if self.verbosity > 2:
print(self.root.describe(), file=sys.stderr)
print('\n\n', file=sys.stderr)
if self.verbosity > 3:
print(self.root.position, file=sys.stderr)
return self.pick_move()
def play_move(self, c):
'''
Notable side effects:
- finalizes the probability distribution according to
this roots visit counts into the class' running tally, `searches_pi`
- Makes the node associated with this move the root, for future
`inject_noise` calls.
'''
if not self.two_player_mode:
self.searches_pi.append(
self.root.children_as_pi(self.root.position.n < self.temp_threshold))
self.qs.append(self.root.Q) # Save our resulting Q.
self.comments.append(self.root.describe())
self.root = self.root.maybe_add_child(coords.flatten_coords(c))
self.position = self.root.position # for showboard
del self.root.parent.children
return True # GTP requires positive result.
def pick_move(self):
'''Picks a move to play, based on MCTS readout statistics.
Highest N is most robust indicator. In the early stage of the game, pick
a move weighted by visit count; later on, pick the absolute max.'''
if self.root.position.n > self.temp_threshold:
fcoord = np.argmax(self.root.child_N)
else:
cdf = self.root.child_N.cumsum()
cdf /= cdf[-1]
selection = random.random()
fcoord = cdf.searchsorted(selection)
assert self.root.child_N[fcoord] != 0
return coords.unflatten_coords(fcoord)
def tree_search(self, num_parallel=None):
if num_parallel is None:
num_parallel = self.num_parallel
leaves = []
failsafe = 0
while len(leaves) < num_parallel and failsafe < num_parallel * 2:
failsafe += 1
leaf = self.root.select_leaf()
if self.verbosity >= 4:
print(self.show_path_to_root(leaf))
# if game is over, override the value estimate with the true score
if leaf.is_done():
value = 1 if leaf.position.score() > 0 else -1
leaf.backup_value(value, up_to=self.root)
continue
leaf.add_virtual_loss(up_to=self.root)
leaves.append(leaf)
if leaves:
move_probs, values = self.network.run_many(
[leaf.position for leaf in leaves])
for leaf, move_prob, value in zip(leaves, move_probs, values):
leaf.revert_virtual_loss(up_to=self.root)
leaf.incorporate_results(move_prob, value, up_to=self.root)
def show_path_to_root(self, node):
pos = node.position
diff = node.position.n - self.root.position.n
if len(pos.recent) == 0:
return
def fmt(move): return "{}-{}".format('b' if move.color == 1 else 'w',
coords.to_human_coord(move.move))
path = " ".join(fmt(move) for move in pos.recent[-diff:])
if node.position.n >= MAX_DEPTH:
path += " (depth cutoff reached) %0.1f" % node.position.score()
elif node.position.is_game_over():
path += " (game over) %0.1f" % node.position.score()
return path
def should_resign(self):
'''Returns true if the player resigned. No further moves should be played'''
return self.root.Q_perspective < self.resign_threshold
def set_result(self, winner, was_resign):
self.result = winner
if was_resign:
string = "B+R" if winner == go.BLACK else "W+R"
else:
string = self.root.position.result_string()
self.result_string = string
def to_sgf(self):
assert self.result_string is not None
pos = self.root.position
if self.comments:
self.comments[0] = ("Resign Threshold: %0.3f\n" %
self.resign_threshold) + self.comments[0]
return sgf_wrapper.make_sgf(pos.recent, self.result_string,
white_name=self.network.name or "Unknown",
black_name=self.network.name or "Unknown",
comments=self.comments)
def extract_data(self):
assert len(self.searches_pi) == self.root.position.n
assert self.result != 0
for pwc, pi in zip(go.replay_position(self.root.position, self.result),
self.searches_pi):
yield pwc.position, pi, pwc.result
def chat(self, msg_type, sender, text):
default_response = "Supported commands are 'winrate', 'nextplay', 'fortune', and 'help'."
if self.root is None or self.root.position.n == 0:
return "I'm not playing right now. " + default_response
if 'winrate' in text.lower():
wr = (abs(self.root.Q) + 1.0) / 2.0
color = "Black" if self.root.Q > 0 else "White"
return "{:s} {:.2f}%".format(color, wr * 100.0)
elif 'nextplay' in text.lower():
return "I'm thinking... " + self.root.most_visited_path()
elif 'fortune' in text.lower():
return "You're feeling lucky!"
elif 'help' in text.lower():
return "I can't help much with go -- try ladders! Otherwise: " + default_response
else:
return default_response
class CGOSPlayerMixin(MCTSPlayerMixin):
def suggest_move(self, position):
self.seconds_per_move = time_recommendation(position.n)
return super().suggest_move(position)