Repository navigation
Expand file tree
/
Copy pathmcts_player.cc
More file actions
448 lines (397 loc) · 14.2 KB
/
Copy pathmcts_player.cc
File metadata and controls
448 lines (397 loc) · 14.2 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
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
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
// 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.
#include "cc/mcts_player.h"
#include <algorithm>
#include <cmath>
#include <sstream>
#include <utility>
#include "absl/memory/memory.h"
#include "absl/strings/str_cat.h"
#include "absl/strings/str_format.h"
#include "absl/strings/str_join.h"
#include "absl/time/clock.h"
#include "cc/logging.h"
#include "cc/random.h"
#include "cc/symmetries.h"
namespace minigo {
std::ostream& operator<<(std::ostream& os, const MctsPlayer::Options& options) {
os << "name:" << options.name << " inject_noise:" << options.inject_noise
<< " soft_pick:" << options.soft_pick
<< " random_symmetry:" << options.random_symmetry
<< " resign_threshold:" << options.resign_threshold
<< " resign_enabled:" << options.resign_enabled
<< " batch_size:" << options.batch_size << " komi:" << options.komi
<< " num_readouts:" << options.num_readouts
<< " seconds_per_move:" << options.seconds_per_move
<< " time_limit:" << options.time_limit
<< " decay_factor:" << options.decay_factor
<< " random_seed:" << options.random_seed;
return os;
}
float TimeRecommendation(int move_num, float seconds_per_move, float time_limit,
float decay_factor) {
// Divide by two since you only play half the moves in a game.
int player_move_num = move_num / 2;
// Sum of geometric series maxes out at endgame_time seconds.
float endgame_time = seconds_per_move / (1.0f - decay_factor);
float base_time;
int core_moves;
if (endgame_time > time_limit) {
// There is so little main time that we're already in 'endgame' mode.
base_time = time_limit * (1.0f - decay_factor);
core_moves = 0;
} else {
// Leave over endgame_time seconds for the end, and play at
// seconds_per_move for as long as possible.
base_time = seconds_per_move;
core_moves = (time_limit - endgame_time) / seconds_per_move;
}
return base_time *
std::pow(decay_factor, std::max(player_move_num - core_moves, 0));
}
MctsPlayer::MctsPlayer(std::unique_ptr<DualNet> network, const Options& options)
: network_(std::move(network)),
game_root_(&root_stats_, {&bv_, &gv_, Color::kBlack}),
rnd_(options.random_seed),
options_(options) {
options_.resign_threshold = -std::abs(options_.resign_threshold);
// When to do deterministic move selection: 30 moves on a 19x19, 6 on 9x9.
// divide 2, multiply 2 guarentees that white and black do even number.
temperature_cutoff_ = !options_.soft_pick ? -1 : (((kN * kN / 12) / 2) * 2);
root_ = &game_root_;
if (options_.verbose) {
MG_LOG(INFO) << "MctsPlayer options: " << options_;
MG_LOG(INFO) << "Random seed used: " << rnd_.seed();
}
InitializeGame({&bv_, &gv_, Color::kBlack});
}
MctsPlayer::~MctsPlayer() {
if (options_.verbose) {
MG_LOG(INFO) << "Inference history:";
for (const auto& info : inferences_) {
MG_LOG(INFO) << info.model << " [" << info.first_move << ", "
<< info.last_move << "]";
}
}
}
void MctsPlayer::InitializeGame(const Position& position) {
root_stats_ = {};
game_root_ = MctsNode(&root_stats_, Position(&bv_, &gv_, position));
ResetRoot();
}
void MctsPlayer::NewGame() {
root_stats_ = {};
game_root_ = MctsNode(&root_stats_, {&bv_, &gv_, Color::kBlack});
ResetRoot();
}
void MctsPlayer::ResetRoot() {
root_ = &game_root_;
history_.clear();
}
bool MctsPlayer::UndoMove() {
if (root_ == &game_root_) {
return false;
}
root_ = root_->parent;
history_.pop_back();
return true;
}
Coord MctsPlayer::SuggestMove() {
auto start = absl::Now();
// In order to correctly count the number of reads performed, the root node
// must be expanded. The root will always be expanded unless this is the first
// time SuggestMove has been called for a game, or PlayMove was called without
// a prior call to SuggestMove.
if (!root_->HasFlag(MctsNode::Flag::kExpanded)) {
tree_search_paths_.clear();
SelectLeaves(root_, 1, &tree_search_paths_);
ProcessLeaves(absl::MakeSpan(tree_search_paths_), options_.random_symmetry);
}
if (options_.inject_noise) {
std::array<float, kNumMoves> noise;
rnd_.Dirichlet(kDirichletAlpha, &noise);
root_->InjectNoise(noise);
}
int current_readouts = root_->N();
if (options_.seconds_per_move > 0) {
// Use time to limit the number of reads.
float seconds_per_move = options_.seconds_per_move;
if (options_.time_limit > 0) {
seconds_per_move =
TimeRecommendation(root_->position.n(), seconds_per_move,
options_.time_limit, options_.decay_factor);
}
while (absl::Now() - start < absl::Seconds(seconds_per_move)) {
TreeSearch();
}
} else {
// Use a fixed number of reads.
while (root_->N() < current_readouts + options_.num_readouts) {
TreeSearch();
}
}
int num_readouts = root_->N() - current_readouts;
auto elapsed = absl::Now() - start;
elapsed = elapsed * 100 / num_readouts;
if (options_.verbose) {
MG_LOG(INFO) << "Milliseconds per 100 reads: "
<< absl::ToInt64Milliseconds(elapsed) << "ms"
<< " over " << num_readouts
<< " readouts (batched: " << options_.batch_size << ")";
MG_LOG(INFO) << root_->CalculateTreeStats();
}
if (ShouldResign()) {
return Coord::kResign;
}
return PickMove();
}
Coord MctsPlayer::PickMove() {
if (root_->position.n() >= temperature_cutoff_) {
Coord c = root_->GetMostVisitedMove();
if (options_.verbose) {
MG_LOG(INFO) << "Picked arg_max " << c;
}
return c;
}
// Select from the first kN * kN moves (instead of kNumMoves) to avoid
// randomly choosing to pass early on in the game.
std::array<float, kN * kN> cdf;
// For moves before the temperature cutoff, exponentiate the probabilities by
// a temperature slightly larger than unity to encourage diversity in early
// play and hopefully to move away from 3-3s.
for (size_t i = 0; i < cdf.size(); ++i) {
cdf[i] = std::pow(root_->child_N(i), kVisitCountSquash);
}
for (size_t i = 1; i < cdf.size(); ++i) {
cdf[i] += cdf[i - 1];
}
float e = rnd_();
Coord c = SearchSorted(cdf, e * cdf.back());
if (options_.verbose) {
MG_LOG(INFO) << "Picked rnd(" << e << ") " << c;
}
MG_DCHECK(root_->child_N(c) != 0);
return c;
}
void MctsPlayer::TreeSearch() {
tree_search_paths_.clear();
SelectLeaves(root_, options_.batch_size, &tree_search_paths_);
ProcessLeaves(absl::MakeSpan(tree_search_paths_), options_.random_symmetry);
}
void MctsPlayer::SelectLeaves(MctsNode* root, int num_leaves,
std::vector<MctsPlayer::TreePath>* paths) {
int max_iterations = num_leaves * 2;
int num_selected = 0;
for (int i = 0; i < max_iterations; ++i) {
auto* leaf = root->SelectLeaf();
if (leaf->game_over() || leaf->at_move_limit()) {
float value = leaf->position.CalculateScore(options_.komi) > 0 ? 1 : -1;
leaf->IncorporateEndGameResult(value, root);
} else {
leaf->AddVirtualLoss(root);
paths->emplace_back(root, leaf);
if (++num_selected == num_leaves) {
break;
}
}
}
}
bool MctsPlayer::ShouldResign() const {
return options_.resign_enabled &&
root_->Q_perspective() < options_.resign_threshold;
}
bool MctsPlayer::PlayMove(Coord c) {
if (root_->game_over()) {
MG_LOG(ERROR) << "can't play move " << c << ", game is over";
return false;
}
// Handle resignations.
if (c == Coord::kResign) {
root_ = root_->MaybeAddChild(c);
if (root_->position.to_play() == Color::kBlack) {
result_ = 1;
result_string_ = "B+R";
} else {
result_ = -1;
result_string_ = "W+R";
}
return true;
}
if (!root_->legal_moves[c]) {
MG_LOG(ERROR) << "Move " << c << " is illegal";
return false;
}
PushHistory(c);
root_ = root_->MaybeAddChild(c);
if (options_.prune_orphaned_nodes) {
// Don't need to keep the parent's children around anymore because we'll
// never revisit them during normal play.
root_->parent->PruneChildren(c);
}
if (options_.verbose) {
MG_LOG(INFO) << absl::StreamFormat("%s Q: %0.5f", name(), root_->Q());
MG_LOG(INFO) << "Played >>" << c;
}
// Handle consecutive passing or termination by move limit.
if (root_->game_over() || root_->at_move_limit()) {
float score = root_->position.CalculateScore(options_.komi);
result_string_ = FormatScore(score);
result_ = score < 0 ? -1 : score > 0 ? 1 : 0;
}
return true;
}
std::string MctsPlayer::FormatScore(float score) const {
return absl::StrFormat("%c+%.1f", score > 0 ? 'B' : 'W', std::abs(score));
}
void MctsPlayer::PushHistory(Coord c) {
history_.emplace_back();
History& history = history_.back();
history.c = c;
history.comment = root_->Describe();
history.node = root_;
if (!inferences_.empty()) {
// Record which model(s) were used when running tree search for this move.
std::vector<std::string> models;
for (auto it = inferences_.rbegin(); it != inferences_.rend(); ++it) {
if (it->last_move < root_->position.n()) {
break;
}
models.push_back(it->model);
}
std::reverse(models.begin(), models.end());
auto model_comment = absl::StrCat("models:", absl::StrJoin(models, ","));
history.comment = absl::StrCat(model_comment, "\n", history.comment);
if (options_.verbose) {
MG_LOG(INFO) << model_comment;
}
}
// Convert child visit counts to a probability distribution, pi.
if (root_->position.n() < temperature_cutoff_) {
// Squash counts before normalizing to match softpick behavior in PickMove.
for (int i = 0; i < kNumMoves; ++i) {
history.search_pi[i] = std::pow(root_->child_N(i), kVisitCountSquash);
}
} else {
for (int i = 0; i < kNumMoves; ++i) {
history.search_pi[i] = root_->child_N(i);
}
}
// Normalize counts.
float sum = 0;
for (int i = 0; i < kNumMoves; ++i) {
sum += history.search_pi[i];
}
for (int i = 0; i < kNumMoves; ++i) {
history.search_pi[i] /= sum;
}
}
void MctsPlayer::ProcessLeaves(absl::Span<TreePath> paths,
bool random_symmetry) {
if (paths.empty()) {
return;
}
// Select symmetry operations to apply.
symmetries_used_.resize(0);
if (random_symmetry) {
symmetries_used_.reserve(paths.size());
for (size_t i = 0; i < paths.size(); ++i) {
symmetries_used_.push_back(static_cast<symmetry::Symmetry>(
rnd_.UniformInt(0, symmetry::kNumSymmetries - 1)));
}
} else {
symmetries_used_.resize(paths.size(), symmetry::kIdentity);
}
// Build input features for each leaf, applying random symmetries if
// requested.
DualNet::BoardFeatures raw_features;
features_.resize(paths.size());
for (size_t i = 0; i < paths.size(); ++i) {
const auto* leaf = paths[i].leaf;
MG_CHECK(leaf->num_virtual_losses_applied > 0)
<< "Don't forget to add a virtual loss before calling ProcessLeaves";
leaf->GetMoveHistory(DualNet::kMoveHistory, &recent_positions_);
DualNet::SetFeatures(recent_positions_, leaf->position.to_play(),
&raw_features);
if (network_->GetInputLayout() == DualNet::InputLayout::kNCHW) {
using OutIter =
symmetry::NchwOutputIterator<kN, DualNet::kNumStoneFeatures, float>;
symmetry::ApplySymmetry<kN, DualNet::kNumStoneFeatures>(
symmetries_used_[i], raw_features.data(),
OutIter(features_[i].data()));
} else {
symmetry::ApplySymmetry<kN, DualNet::kNumStoneFeatures>(
symmetries_used_[i], raw_features.data(), features_[i].data());
}
}
std::vector<const DualNet::BoardFeatures*> feature_ptrs;
feature_ptrs.reserve(features_.size());
for (const auto& feature : features_) {
feature_ptrs.push_back(&feature);
}
outputs_.resize(paths.size());
std::vector<DualNet::Output*> output_ptrs;
output_ptrs.reserve(outputs_.size());
for (auto& output : outputs_) {
output_ptrs.push_back(&output);
}
// Run inference.
network_->RunMany(std::move(feature_ptrs), std::move(output_ptrs), &model_);
// Record some information about the inference.
if (!model_.empty()) {
if (inferences_.empty() || model_ != inferences_.back().model) {
inferences_.emplace_back(model_, root_->position.n());
}
inferences_.back().last_move = root_->position.n();
inferences_.back().total_count += paths.size();
}
// Incorporate the inference outputs back into tree search, undoing any
// previously applied random symmetries.
std::array<float, kNumMoves> raw_policy;
for (size_t i = 0; i < paths.size(); ++i) {
auto* root = paths[i].root;
auto* leaf = paths[i].leaf;
const auto& output = outputs_[i];
symmetry::ApplySymmetry<kN, 1>(symmetry::Inverse(symmetries_used_[i]),
output.policy.data(), raw_policy.data());
raw_policy[Coord::kPass] = output.policy[Coord::kPass];
leaf->IncorporateResults(raw_policy, output.value, root);
leaf->RevertVirtualLoss(root);
}
}
// &q is the bleakest move from the perspective of the winner, i.e., negative.
bool FindBleakestMove(const MctsPlayer& player, int* move, float* q) {
if (player.options().resign_enabled) {
return false;
}
const auto& history = player.history();
if (history.empty()) {
return false;
}
// Find the move at which the game looked the bleakest from the perspective
// of the winner.
float result = player.result();
float bleakest_eval = history[0].node->Q() * result;
size_t bleakest_move = 0;
for (size_t i = 1; i < history.size(); ++i) {
float eval = history[i].node->Q() * result;
if (eval < bleakest_eval) {
bleakest_eval = eval;
bleakest_move = i;
}
}
*move = int(bleakest_move);
*q = bleakest_eval;
return true;
}
} // namespace minigo