Repository navigation
Expand file tree
/
Copy pathUCTSearch.cpp
More file actions
896 lines (760 loc) · 29.8 KB
/
Copy pathUCTSearch.cpp
File metadata and controls
896 lines (760 loc) · 29.8 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
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
/*
This file is part of Leela Zero.
Copyright (C) 2017-2019 Gian-Carlo Pascutto and contributors
Leela Zero is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
Leela Zero is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with Leela Zero. If not, see <http://www.gnu.org/licenses/>.
Additional permission under GNU GPL version 3 section 7
If you modify this Program, or any covered work, by linking or
combining it with NVIDIA Corporation's libraries from the
NVIDIA CUDA Toolkit and/or the NVIDIA CUDA Deep Neural
Network library and/or the NVIDIA TensorRT inference library
(or a modified version of those libraries), containing parts covered
by the terms of the respective license agreement, the licensors of
this Program grant you additional permission to convey the resulting
work.
*/
#include "config.h"
#include "UCTSearch.h"
#include <boost/format.hpp>
#include <cassert>
#include <cmath>
#include <cstddef>
#include <functional>
#include <limits>
#include <memory>
#include <type_traits>
#include <algorithm>
#include "FastBoard.h"
#include "FastState.h"
#include "FullBoard.h"
#include "GTP.h"
#include "GameState.h"
#include "TimeControl.h"
#include "Timing.h"
#include "Training.h"
#include "Utils.h"
#include "OpenCLScheduler.h"
using namespace Utils;
constexpr int UCTSearch::UNLIMITED_PLAYOUTS;
class OutputAnalysisData {
public:
OutputAnalysisData(const std::string& move, int visits,
float winrate, float policy_prior, std::string pv)
: m_move(move), m_visits(visits), m_winrate(winrate),
m_policy_prior(policy_prior), m_pv(pv) {};
std::string get_info_string(int order) const {
auto tmp = "info move " + m_move
+ " visits " + std::to_string(m_visits)
+ " winrate "
+ std::to_string(static_cast<int>(m_winrate * 10000))
+ " prior "
+ std::to_string(static_cast<int>(m_policy_prior * 10000.0f));
if (order >= 0) {
tmp += " order " + std::to_string(order);
}
tmp += " pv " + m_pv;
return tmp;
}
friend bool operator<(const OutputAnalysisData& a,
const OutputAnalysisData& b) {
if (a.m_visits == b.m_visits) {
return a.m_winrate < b.m_winrate;
}
return a.m_visits < b.m_visits;
}
private:
std::string m_move;
int m_visits;
float m_winrate;
float m_policy_prior;
std::string m_pv;
};
UCTSearch::UCTSearch(GameState& g, Network& network)
: m_rootstate(g), m_network(network) {
set_playout_limit(cfg_max_playouts);
set_visit_limit(cfg_max_visits);
m_root = std::make_unique<UCTNode>(FastBoard::PASS, 0.0f);
}
bool UCTSearch::advance_to_new_rootstate() {
if (!m_root || !m_last_rootstate) {
// No current state
return false;
}
if (m_rootstate.get_komi() != m_last_rootstate->get_komi()) {
return false;
}
auto depth =
int(m_rootstate.get_movenum() - m_last_rootstate->get_movenum());
if (depth < 0) {
return false;
}
auto test = std::make_unique<GameState>(m_rootstate);
for (auto i = 0; i < depth; i++) {
test->undo_move();
}
if (m_last_rootstate->board.get_hash() != test->board.get_hash()) {
// m_rootstate and m_last_rootstate don't match
return false;
}
// Make sure that the nodes we destroyed the previous move are
// in fact destroyed.
while (!m_delete_futures.empty()) {
m_delete_futures.front().wait_all();
m_delete_futures.pop_front();
}
// Try to replay moves advancing m_root
for (auto i = 0; i < depth; i++) {
ThreadGroup tg(thread_pool);
test->forward_move();
const auto move = test->get_last_move();
auto oldroot = std::move(m_root);
m_root = oldroot->find_child(move);
// Lazy tree destruction. Instead of calling the destructor of the
// old root node on the main thread, send the old root to a separate
// thread and destroy it from the child thread. This will save a
// bit of time when dealing with large trees.
auto p = oldroot.release();
tg.add_task([p]() { delete p; });
m_delete_futures.push_back(std::move(tg));
if (!m_root) {
// Tree hasn't been expanded this far
return false;
}
m_last_rootstate->play_move(move);
}
assert(m_rootstate.get_movenum() == m_last_rootstate->get_movenum());
if (m_last_rootstate->board.get_hash() != test->board.get_hash()) {
// Can happen if user plays multiple moves in a row by same player
return false;
}
return true;
}
void UCTSearch::update_root() {
// Definition of m_playouts is playouts per search call.
// So reset this count now.
m_playouts = 0;
#ifndef NDEBUG
auto start_nodes = m_root->count_nodes_and_clear_expand_state();
#endif
if (!advance_to_new_rootstate() || !m_root) {
m_root = std::make_unique<UCTNode>(FastBoard::PASS, 0.0f);
}
// Clear last_rootstate to prevent accidental use.
m_last_rootstate.reset(nullptr);
// Check how big our search tree (reused or new) is.
m_nodes = m_root->count_nodes_and_clear_expand_state();
#ifndef NDEBUG
if (m_nodes > 0) {
myprintf("update_root, %d -> %d nodes (%.1f%% reused)\n",
start_nodes, m_nodes.load(), 100.0 * m_nodes.load() / start_nodes);
}
#endif
}
float UCTSearch::get_min_psa_ratio() const {
const auto mem_full = UCTNodePointer::get_tree_size() / static_cast<float>(cfg_max_tree_size);
// If we are halfway through our memory budget, start trimming
// moves with very low policy priors.
if (mem_full > 0.5f) {
// Memory is almost exhausted, trim more aggressively.
if (mem_full > 0.95f) {
// if completely full just stop expansion by returning an impossible number
if (mem_full >= 1.0f) {
return 2.0f;
}
return 0.01f;
}
return 0.001f;
}
return 0.0f;
}
SearchResult UCTSearch::play_simulation(GameState & currstate,
UCTNode* const node) {
const auto color = currstate.get_to_move();
auto result = SearchResult{};
node->virtual_loss();
if (node->expandable()) {
if (currstate.get_passes() >= 2) {
auto score = currstate.final_score();
result = SearchResult::from_score(score);
} else {
float eval;
const auto had_children = node->has_children();
const auto success =
node->create_children(m_network, m_nodes, currstate, eval,
get_min_psa_ratio());
if (!had_children && success) {
result = SearchResult::from_eval(eval);
}
}
}
if (node->has_children() && !result.valid()) {
auto next = node->uct_select_child(color, node == m_root.get());
auto move = next->get_move();
currstate.play_move(move);
if (move != FastBoard::PASS && currstate.superko()) {
next->invalidate();
} else {
result = play_simulation(currstate, next);
}
}
if (result.valid()) {
node->update(result.eval());
}
node->virtual_loss_undo();
return result;
}
void UCTSearch::dump_stats(FastState & state, UCTNode & parent) {
if (cfg_quiet || !parent.has_children()) {
return;
}
const int color = state.get_to_move();
// sort children, put best move on top
parent.sort_children(color);
if (parent.get_first_child()->first_visit()) {
return;
}
int movecount = 0;
for (const auto& node : parent.get_children()) {
// Always display at least two moves. In the case there is
// only one move searched the user could get an idea why.
if (++movecount > 2 && !node->get_visits()) break;
auto move = state.move_to_text(node->get_move());
auto tmpstate = FastState{state};
tmpstate.play_move(node->get_move());
auto pv = move + " " + get_pv(tmpstate, *node);
myprintf("%4s -> %7d (V: %5.2f%%) (N: %5.2f%%) PV: %s\n",
move.c_str(),
node->get_visits(),
node->get_visits() ? node->get_raw_eval(color)*100.0f : 0.0f,
node->get_policy() * 100.0f,
pv.c_str());
}
tree_stats(parent);
}
void UCTSearch::output_analysis(FastState & state, UCTNode & parent) {
// We need to make a copy of the data before sorting
auto sortable_data = std::vector<OutputAnalysisData>();
if (!parent.has_children()) {
return;
}
const auto color = state.get_to_move();
for (const auto& node : parent.get_children()) {
// Send only variations with visits, unless more moves were
// requested explicitly.
if (!node->get_visits()
&& sortable_data.size() >= cfg_analyze_tags.post_move_count()) {
continue;
}
auto move = state.move_to_text(node->get_move());
auto tmpstate = FastState{state};
tmpstate.play_move(node->get_move());
auto rest_of_pv = get_pv(tmpstate, *node);
auto pv = move + (rest_of_pv.empty() ? "" : " " + rest_of_pv);
auto move_eval = node->get_visits() ? node->get_raw_eval(color) : 0.0f;
auto policy = node->get_policy();
// Store data in array
sortable_data.emplace_back(move, node->get_visits(),
move_eval, policy, pv);
}
// Sort array to decide order
std::stable_sort(rbegin(sortable_data), rend(sortable_data));
auto i = 0;
// Output analysis data in gtp stream
for (const auto& node : sortable_data) {
if (i > 0) {
gtp_printf_raw(" ");
}
gtp_printf_raw(node.get_info_string(i).c_str());
i++;
}
gtp_printf_raw("\n");
}
void UCTSearch::tree_stats(const UCTNode& node) {
size_t nodes = 0;
size_t non_leaf_nodes = 0;
size_t depth_sum = 0;
size_t max_depth = 0;
size_t children_count = 0;
std::function<void(const UCTNode& node, size_t)> traverse =
[&](const UCTNode& node, size_t depth) {
nodes += 1;
non_leaf_nodes += node.get_visits() > 1;
depth_sum += depth;
if (depth > max_depth) max_depth = depth;
for (const auto& child : node.get_children()) {
if (child.get_visits() > 0) {
children_count += 1;
traverse(*(child.get()), depth+1);
} else {
nodes += 1;
depth_sum += depth+1;
if (depth >= max_depth) max_depth = depth+1;
}
}
};
traverse(node, 0);
if (nodes > 0) {
myprintf("%.1f average depth, %d max depth\n",
(1.0f*depth_sum) / nodes, max_depth);
myprintf("%d non leaf nodes, %.2f average children\n",
non_leaf_nodes, (1.0f*children_count) / non_leaf_nodes);
}
}
bool UCTSearch::should_resign(passflag_t passflag, float besteval) {
if (passflag & UCTSearch::NORESIGN) {
// resign not allowed
return false;
}
if (cfg_resignpct == 0) {
// resign not allowed
return false;
}
const size_t num_intersections = m_rootstate.board.get_boardsize()
* m_rootstate.board.get_boardsize();
const auto move_threshold = num_intersections / 4;
const auto movenum = m_rootstate.get_movenum();
if (movenum <= move_threshold) {
// too early in game to resign
return false;
}
const auto color = m_rootstate.board.get_to_move();
const auto is_default_cfg_resign = cfg_resignpct < 0;
const auto resign_threshold =
0.01f * (is_default_cfg_resign ? 10 : cfg_resignpct);
if (besteval > resign_threshold) {
// eval > cfg_resign
return false;
}
if ((m_rootstate.get_handicap() > 0)
&& (color == FastBoard::WHITE)
&& is_default_cfg_resign) {
const auto handicap_resign_threshold =
resign_threshold / (1 + m_rootstate.get_handicap());
// Blend the thresholds for the first ~215 moves.
auto blend_ratio = std::min(1.0f, movenum / (0.6f * num_intersections));
auto blended_resign_threshold = blend_ratio * resign_threshold
+ (1 - blend_ratio) * handicap_resign_threshold;
if (besteval > blended_resign_threshold) {
// Allow lower eval for white in handicap games
// where opp may fumble.
return false;
}
}
if (!m_rootstate.is_move_legal(color, FastBoard::RESIGN)) {
return false;
}
return true;
}
int UCTSearch::get_best_move(passflag_t passflag) {
int color = m_rootstate.board.get_to_move();
// Make sure best is first
m_root->sort_children(color);
// Check whether to randomize the best move proportional
// to the playout counts, early game only.
auto movenum = int(m_rootstate.get_movenum());
if (movenum < cfg_random_cnt) {
m_root->randomize_first_proportionally();
}
auto first_child = m_root->get_first_child();
assert(first_child != nullptr);
auto bestmove = first_child->get_move();
auto besteval = first_child->first_visit() ? 0.5f : first_child->get_raw_eval(color);
// do we want to fiddle with the best move because of the rule set?
if (passflag & UCTSearch::NOPASS) {
// were we going to pass?
if (bestmove == FastBoard::PASS) {
UCTNode * nopass = m_root->get_nopass_child(m_rootstate);
if (nopass != nullptr) {
myprintf("Preferring not to pass.\n");
bestmove = nopass->get_move();
if (nopass->first_visit()) {
besteval = 1.0f;
} else {
besteval = nopass->get_raw_eval(color);
}
} else {
myprintf("Pass is the only acceptable move.\n");
}
}
} else if (!cfg_dumbpass) {
const auto relative_score = (color == FastBoard::BLACK ? 1 : -1) * m_rootstate.final_score();
if (bestmove == FastBoard::PASS) {
// Either by forcing or coincidence passing is
// on top...check whether passing loses instantly
// do full count including dead stones.
// In a reinforcement learning setup, it is possible for the
// network to learn that, after passing in the tree, the two last
// positions are identical, and this means the position is only won
// if there are no dead stones in our own territory (because we use
// Trump-Taylor scoring there). So strictly speaking, the next
// heuristic isn't required for a pure RL network, and we have
// a commandline option to disable the behavior during learning.
// On the other hand, with a supervised learning setup, we fully
// expect that the engine will pass out anything that looks like
// a finished game even with dead stones on the board (because the
// training games were using scoring with dead stone removal).
// So in order to play games with a SL network, we need this
// heuristic so the engine can "clean up" the board. It will still
// only clean up the bare necessity to win. For full dead stone
// removal, kgs-genmove_cleanup and the NOPASS mode must be used.
// Do we lose by passing?
if (relative_score < 0.0f) {
myprintf("Passing loses :-(\n");
// Find a valid non-pass move.
UCTNode * nopass = m_root->get_nopass_child(m_rootstate);
if (nopass != nullptr) {
myprintf("Avoiding pass because it loses.\n");
bestmove = nopass->get_move();
if (nopass->first_visit()) {
besteval = 1.0f;
} else {
besteval = nopass->get_raw_eval(color);
}
} else {
myprintf("No alternative to passing.\n");
}
} else if (relative_score > 0.0f) {
myprintf("Passing wins :-)\n");
} else {
myprintf("Passing draws :-|\n");
// Find a valid non-pass move.
const auto nopass = m_root->get_nopass_child(m_rootstate);
if (nopass != nullptr && !nopass->first_visit()) {
const auto nopass_eval = nopass->get_raw_eval(color);
if (nopass_eval > 0.5f) {
myprintf("Avoiding pass because there could be a winning alternative.\n");
bestmove = nopass->get_move();
besteval = nopass_eval;
}
}
if (bestmove == FastBoard::PASS) {
myprintf("No seemingly better alternative to passing.\n");
}
}
} else if (m_rootstate.get_last_move() == FastBoard::PASS) {
// Opponents last move was passing.
// We didn't consider passing. Should we have and
// end the game immediately?
if (!m_rootstate.is_move_legal(color, FastBoard::PASS)) {
myprintf("Passing is forbidden, I'll play on.\n");
// do we lose by passing?
} else if (relative_score < 0.0f) {
myprintf("Passing loses, I'll play on.\n");
} else if (relative_score > 0.0f) {
myprintf("Passing wins, I'll pass out.\n");
bestmove = FastBoard::PASS;
} else {
myprintf("Passing draws, make it depend on evaluation.\n");
if (besteval < 0.5f) {
bestmove = FastBoard::PASS;
}
}
}
}
// if we aren't passing, should we consider resigning?
if (bestmove != FastBoard::PASS) {
if (should_resign(passflag, besteval)) {
myprintf("Eval (%.2f%%) looks bad. Resigning.\n",
100.0f * besteval);
bestmove = FastBoard::RESIGN;
}
}
return bestmove;
}
std::string UCTSearch::get_pv(FastState & state, UCTNode& parent) {
if (!parent.has_children()) {
return std::string();
}
if (parent.expandable()) {
// Not fully expanded. This means someone could expand
// the node while we want to traverse the children.
// Avoid the race conditions and don't go through the rabbit hole
// of trying to print things from this node.
return std::string();
}
auto& best_child = parent.get_best_root_child(state.get_to_move());
if (best_child.first_visit()) {
return std::string();
}
auto best_move = best_child.get_move();
auto res = state.move_to_text(best_move);
state.play_move(best_move);
auto next = get_pv(state, best_child);
if (!next.empty()) {
res.append(" ").append(next);
}
return res;
}
std::string UCTSearch::get_analysis() {
auto playouts = m_playouts.load();
FastState tempstate = m_rootstate;
int color = tempstate.board.get_to_move();
auto pvstring = get_pv(tempstate, *m_root);
float winrate = 100.0f * m_root->get_raw_eval(color);
return str(boost::format("Playouts: %d, Win: %5.2f%%, PV: %s")
% playouts % winrate % pvstring.c_str());
}
bool UCTSearch::is_running() const {
return m_run && UCTNodePointer::get_tree_size() < cfg_max_tree_size;
}
int UCTSearch::est_playouts_left(int elapsed_centis, int time_for_move) const {
auto playouts = m_playouts.load();
const auto playouts_left =
std::max(0, std::min(m_maxplayouts - playouts,
m_maxvisits - m_root->get_visits()));
// Wait for at least 1 second and 100 playouts
// so we get a reliable playout_rate.
if (elapsed_centis < 100 || playouts < 100) {
return playouts_left;
}
const auto playout_rate = 1.0f * playouts / elapsed_centis;
const auto time_left = std::max(0, time_for_move - elapsed_centis);
return std::min(playouts_left,
static_cast<int>(std::ceil(playout_rate * time_left)));
}
size_t UCTSearch::prune_noncontenders(int elapsed_centis, int time_for_move, bool prune) {
auto Nfirst = 0;
// There are no cases where the root's children vector gets modified
// during a multithreaded search, so it is safe to walk it here without
// taking the (root) node lock.
for (const auto& node : m_root->get_children()) {
if (node->valid()) {
Nfirst = std::max(Nfirst, node->get_visits());
}
}
const auto min_required_visits =
Nfirst - est_playouts_left(elapsed_centis, time_for_move);
auto pruned_nodes = size_t{0};
for (const auto& node : m_root->get_children()) {
if (node->valid()) {
const auto has_enough_visits =
node->get_visits() >= min_required_visits;
if (prune) {
node->set_active(has_enough_visits);
}
if (!has_enough_visits) {
++pruned_nodes;
}
}
}
assert(pruned_nodes < m_root->get_children().size());
return pruned_nodes;
}
bool UCTSearch::have_alternate_moves(int elapsed_centis, int time_for_move) {
if (cfg_timemanage == TimeManagement::OFF) {
return true;
}
// For self play use. Disables pruning of non-contenders to not bias the training data.
auto prune = cfg_timemanage != TimeManagement::NO_PRUNING;
auto pruned = prune_noncontenders(elapsed_centis, time_for_move, prune);
if (pruned < m_root->get_children().size() - 1) {
return true;
}
// If we cannot save up time anyway, use all of it. This
// behavior can be overruled by setting "fast" time management,
// which will cause Leela to quickly respond to obvious/forced moves.
// That comes at the cost of some playing strength as she now cannot
// think ahead about her next moves in the remaining time.
auto my_color = m_rootstate.get_to_move();
auto tc = m_rootstate.get_timecontrol();
if (!tc.can_accumulate_time(my_color)
|| m_maxplayouts < UCTSearch::UNLIMITED_PLAYOUTS) {
if (cfg_timemanage != TimeManagement::FAST) {
return true;
}
}
// In a timed search we will essentially always exit because
// the remaining time is too short to let another move win, so
// avoid spamming this message every move. We'll print it if we
// save at least half a second.
if (time_for_move - elapsed_centis > 50) {
myprintf("%.1fs left, stopping early.\n",
(time_for_move - elapsed_centis) / 100.0f);
}
return false;
}
bool UCTSearch::stop_thinking(int elapsed_centis, int time_for_move) const {
return m_playouts >= m_maxplayouts
|| m_root->get_visits() >= m_maxvisits
|| elapsed_centis >= time_for_move;
}
void UCTWorker::operator()() {
do {
auto currstate = std::make_unique<GameState>(m_rootstate);
auto result = m_search->play_simulation(*currstate, m_root);
if (result.valid()) {
m_search->increment_playouts();
}
} while (m_search->is_running());
}
void UCTSearch::increment_playouts() {
m_playouts++;
}
int UCTSearch::think(int color, passflag_t passflag) {
// Start counting time for us
m_rootstate.start_clock(color);
// set up timing info
Time start;
update_root();
// set side to move
m_rootstate.board.set_to_move(color);
auto time_for_move =
m_rootstate.get_timecontrol().max_time_for_move(
m_rootstate.board.get_boardsize(),
color, m_rootstate.get_movenum());
myprintf("Thinking at most %.1f seconds...\n", time_for_move/100.0f);
// create a sorted list of legal moves (make sure we
// play something legal and decent even in time trouble)
m_root->prepare_root_node(m_network, color, m_nodes, m_rootstate);
m_run = true;
int cpus = cfg_num_threads;
ThreadGroup tg(thread_pool);
for (int i = 1; i < cpus; i++) {
tg.add_task(UCTWorker(m_rootstate, this, m_root.get()));
}
auto keeprunning = true;
auto last_update = 0;
auto last_output = 0;
do {
auto currstate = std::make_unique<GameState>(m_rootstate);
auto result = play_simulation(*currstate, m_root.get());
if (result.valid()) {
increment_playouts();
}
Time elapsed;
int elapsed_centis = Time::timediff_centis(start, elapsed);
if (cfg_analyze_tags.interval_centis() &&
elapsed_centis - last_output > cfg_analyze_tags.interval_centis()) {
last_output = elapsed_centis;
output_analysis(m_rootstate, *m_root);
}
// output some stats every few seconds
// check if we should still search
if (!cfg_quiet && elapsed_centis - last_update > 250) {
last_update = elapsed_centis;
myprintf("%s\n", get_analysis().c_str());
}
keeprunning = is_running();
keeprunning &= !stop_thinking(elapsed_centis, time_for_move);
keeprunning &= have_alternate_moves(elapsed_centis, time_for_move);
} while (keeprunning);
// Make sure to post at least once.
if (cfg_analyze_tags.interval_centis() && last_output == 0) {
output_analysis(m_rootstate, *m_root);
}
// Stop the search.
m_run = false;
tg.wait_all();
// Reactivate all pruned root children.
for (const auto& node : m_root->get_children()) {
node->set_active(true);
}
m_rootstate.stop_clock(color);
if (!m_root->has_children()) {
return FastBoard::PASS;
}
// Display search info.
myprintf("\n");
dump_stats(m_rootstate, *m_root);
Training::record(m_network, m_rootstate, *m_root);
Time elapsed;
int elapsed_centis = Time::timediff_centis(start, elapsed);
myprintf("%d visits, %d nodes, %d playouts, %.0f n/s\n\n",
m_root->get_visits(),
m_nodes.load(),
m_playouts.load(),
(m_playouts * 100.0) / (elapsed_centis+1));
#ifdef USE_OPENCL
#ifndef NDEBUG
myprintf("batch stats: %d %d\n",
batch_stats.single_evals.load(),
batch_stats.batch_evals.load()
);
#endif
#endif
int bestmove = get_best_move(passflag);
// Save the explanation.
m_think_output =
str(boost::format("move %d, %c => %s\n%s")
% m_rootstate.get_movenum()
% (color == FastBoard::BLACK ? 'B' : 'W')
% m_rootstate.move_to_text(bestmove).c_str()
% get_analysis().c_str());
// Copy the root state. Use to check for tree re-use in future calls.
m_last_rootstate = std::make_unique<GameState>(m_rootstate);
return bestmove;
}
// Brief output from last think() call.
std::string UCTSearch::explain_last_think() const {
return m_think_output;
}
void UCTSearch::ponder() {
auto disable_reuse = cfg_analyze_tags.has_move_restrictions();
if (disable_reuse) {
m_last_rootstate.reset(nullptr);
}
update_root();
m_root->prepare_root_node(m_network, m_rootstate.board.get_to_move(),
m_nodes, m_rootstate);
m_run = true;
ThreadGroup tg(thread_pool);
for (auto i = size_t{1}; i < cfg_num_threads; i++) {
tg.add_task(UCTWorker(m_rootstate, this, m_root.get()));
}
Time start;
auto keeprunning = true;
auto last_output = 0;
do {
auto currstate = std::make_unique<GameState>(m_rootstate);
auto result = play_simulation(*currstate, m_root.get());
if (result.valid()) {
increment_playouts();
}
if (cfg_analyze_tags.interval_centis()) {
Time elapsed;
int elapsed_centis = Time::timediff_centis(start, elapsed);
if (elapsed_centis - last_output > cfg_analyze_tags.interval_centis()) {
last_output = elapsed_centis;
output_analysis(m_rootstate, *m_root);
}
}
keeprunning = is_running();
keeprunning &= !stop_thinking(0, 1);
} while (!Utils::input_pending() && keeprunning);
// Make sure to post at least once.
if (cfg_analyze_tags.interval_centis() && last_output == 0) {
output_analysis(m_rootstate, *m_root);
}
// Stop the search.
m_run = false;
tg.wait_all();
// Display search info.
myprintf("\n");
dump_stats(m_rootstate, *m_root);
myprintf("\n%d visits, %d nodes\n\n", m_root->get_visits(), m_nodes.load());
// Copy the root state. Use to check for tree re-use in future calls.
if (!disable_reuse) {
m_last_rootstate = std::make_unique<GameState>(m_rootstate);
}
}
void UCTSearch::set_playout_limit(int playouts) {
static_assert(std::is_convertible<decltype(playouts),
decltype(m_maxplayouts)>::value,
"Inconsistent types for playout amount.");
m_maxplayouts = std::min(playouts, UNLIMITED_PLAYOUTS);
}
void UCTSearch::set_visit_limit(int visits) {
static_assert(std::is_convertible<decltype(visits),
decltype(m_maxvisits)>::value,
"Inconsistent types for visits amount.");
// Limit to type max / 2 to prevent overflow when multithreading.
m_maxvisits = std::min(visits, UNLIMITED_PLAYOUTS);
}