Repository navigation
Expand file tree
/
Copy pathmcts_player_test.cc
More file actions
534 lines (454 loc) · 16.4 KB
/
Copy pathmcts_player_test.cc
File metadata and controls
534 lines (454 loc) · 16.4 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
// 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 <memory>
#include <string>
#include <utility>
#include "absl/memory/memory.h"
#include "cc/algorithm.h"
#include "cc/color.h"
#include "cc/constants.h"
#include "cc/dual_net/fake_dual_net.h"
#include "cc/position.h"
#include "cc/test_utils.h"
#include "gtest/gtest.h"
namespace minigo {
namespace {
static constexpr char kAlmostDoneBoard[] = R"(
.XO.XO.OO
X.XXOOOO.
XXXXXOOOO
XXXXXOOOO
.XXXXOOO.
XXXXXOOOO
.XXXXOOO.
XXXXXOOOO
XXXXOOOOO)";
class TestablePlayer : public MctsPlayer {
public:
explicit TestablePlayer(const Options& options)
: MctsPlayer(absl::make_unique<FakeDualNet>(), options) {}
explicit TestablePlayer(std::unique_ptr<DualNet> network,
const Options& options)
: MctsPlayer(std::move(network), options) {}
TestablePlayer(absl::Span<const float> fake_priors, float fake_value,
const Options& options)
: MctsPlayer(absl::make_unique<FakeDualNet>(fake_priors, fake_value),
options) {}
using MctsPlayer::PickMove;
using MctsPlayer::PlayMove;
using MctsPlayer::ProcessLeaves;
using MctsPlayer::ResetRoot;
using MctsPlayer::rnd;
using MctsPlayer::TreePath;
using MctsPlayer::TreeSearch;
using MctsPlayer::UndoMove;
std::array<float, kNumMoves> Noise() {
std::array<float, kNumMoves> noise;
rnd()->Dirichlet(kDirichletAlpha, &noise);
return noise;
}
DualNet::Output Run(const DualNet::BoardFeatures& features) {
DualNet::Output output;
network()->RunMany({&features}, {&output}, nullptr);
return output;
}
void TreeSearch(int virtual_losses) {
mutable_options()->batch_size = virtual_losses;
TreeSearch();
}
};
std::unique_ptr<TestablePlayer> CreateBasicPlayer(MctsPlayer::Options options) {
// Always use a deterministic random seed.
options.random_seed = 17;
auto player = absl::make_unique<TestablePlayer>(options);
auto* first_node = player->root()->SelectLeaf();
DualNet::BoardFeatures features;
std::vector<const Position::Stones*> positions = {
&player->root()->position.stones()};
DualNet::SetFeatures(positions, Color::kBlack, &features);
auto output = player->Run(features);
first_node->IncorporateResults(output.policy, output.value, player->root());
return player;
}
std::unique_ptr<TestablePlayer> CreateAlmostDonePlayer(int n) {
// Always use a deterministic random seed.
MctsPlayer::Options options;
options.random_seed = 17;
options.komi = 2.5;
// Don't apply random symmetries. If we did, the probabilities we set in
// the FakeDualNet won't be chosen correctly (since the board position will be
// randomly transformed).
options.random_symmetry = false;
std::array<float, kNumMoves> probs;
for (auto&am
F438
p; p : probs) {
p = 0.001;
}
probs[Coord(0, 2)] = 0.2;
probs[Coord(0, 3)] = 0.2;
probs[Coord(0, 4)] = 0.2;
probs[Coord::kPass] = 0.2;
auto player = absl::make_unique<TestablePlayer>(probs, 0, options);
auto board = TestablePosition(kAlmostDoneBoard, Color::kBlack, n);
player->InitializeGame(board);
return player;
}
TEST(MctsPlayerTest, TimeRecommendation) {
// Early in the game with plenty of time left, the time recommendation should
// be the requested number of seconds per move.
EXPECT_EQ(5, TimeRecommendation(0, 5, 1000, 0.98));
EXPECT_EQ(5, TimeRecommendation(1, 5, 1000, 0.98));
EXPECT_EQ(5, TimeRecommendation(10, 5, 1000, 0.98));
EXPECT_EQ(5, TimeRecommendation(50, 5, 1000, 0.98));
// With a small time limit, the time recommendation should immediately be less
// than requested.
EXPECT_GT(1.0f, TimeRecommendation(0, 5, 10, 0.98));
// Time recommendations for even and odd moves should be identical.
EXPECT_EQ(TimeRecommendation(20, 5, 10, 0.98),
TimeRecommendation(21, 5, 10, 0.98));
// If we're later into the game than should really be possible, time
// recommendation should be almost zero.
EXPECT_GT(0.0001, TimeRecommendation(1000, 5, 100, 0.98));
}
TEST(MctsPlayerTest, InjectNoise) {
MctsPlayer::Options options;
auto player = CreateBasicPlayer(options);
auto* root = player->root();
// FakeDualNet should return normalized priors.
float sum_P = 0;
for (int i = 0; i < kNumMoves; ++i) {
sum_P += root->child_P(i);
}
EXPECT_NEAR(1, sum_P, 0.000001);
for (int i = 0; i < kNumMoves; ++i) {
EXPECT_EQ(root->child_U(0), root->child_U(i));
}
root->InjectNoise(player->Noise());
// Priors should still be normalized after injecting noise.
sum_P = 0;
for (int i = 0; i < kNumMoves; ++i) {
sum_P += root->child_P(i);
}
EXPECT_NEAR(1, sum_P, 0.000001);
// With Dirichelet noise, majority of density should be in one node.
int i = ArgMax(root->edges, MctsNode::CmpP);
float max_P = root->child_P(i);
EXPECT_GT(max_P, 3.0 / kNumMoves);
}
// Verify that with soft pick disabled, the player will always choose the best
// move.
TEST(MctsPlayerTest, PickMoveArgMax) {
MctsPlayer::Options options;
options.soft_pick = false;
auto player = CreateBasicPlayer(options);
auto* root = player->root();
std::vector<std::pair<Coord, int>> child_visits = {
{{2, 0}, 10},
{{1, 0}, 5},
{{3, 0}, 1},
};
for (const auto& p : child_visits) {
root->MaybeAddChild(p.first);
root->edges[p.first].N = p.second;
}
for (int i = 0; i < 100; ++i) {
EXPECT_EQ(Coord(2, 0), player->PickMove());
}
}
// Verify that with soft pick enabled, the player will choose moves early in the
// game proportionally to their visit count.
TEST(MctsPlayerTest, PickMoveSoft) {
MctsPlayer::Options options;
options.soft_pick = true;
auto player = CreateBasicPlayer(options);
auto* root = player->root();
root->edges[Coord(2, 0)].N = 10;
root->edges[Coord(1, 0)].N = 5;
root->edges[Coord(3, 0)].N = 1;
int count_1_0 = 0;
int count_2_0 = 0;
int count_3_0 = 0;
for (int i = 0; i < 160; ++i) {
auto move = player->PickMove();
if (move == Coord(1, 0)) {
++count_1_0;
} else if (move == Coord(2, 0)) {
++count_2_0;
} else {
ASSERT_EQ(Coord(3, 0), move);
++count_3_0;
}
}
EXPECT_NEAR(100, count_2_0, 5);
EXPECT_NEAR(50, count_1_0, 5);
EXPECT_NEAR(10, count_3_0, 5);
}
TEST(MctsPlayerTest, DontPassIfLosing) {
auto player = CreateAlmostDonePlayer(0);
auto* root = player->root();
EXPECT_EQ(-0.5, root->position.CalculateScore(player->options().komi));
for (int i = 0; i < 20; ++i) {
player->TreeSearch(1);
}
// Search should converge on D9 as only winning move.
auto best_move = ArgMax(root->edges, MctsNode::CmpN);
ASSERT_EQ(Coord::FromKgs("D9"), best_move);
// D9 should have a positive value.
EXPECT_LT(0, root->child_Q(best_move));
EXPECT_LE(20, root->N());
// Passing should be ineffective.
EXPECT_GT(0, root->child_Q(Coord::kPass));
// No virtual losses should be pending.
EXPECT_EQ(0, CountPendingVirtualLosses(root));
}
TEST(MctsPlayerTest, ParallelTreeSearch) {
auto player = CreateAlmostDonePlayer(0);
auto* root = player->root();
// Initialize the tree so that the root node has populated children.
player->TreeSearch(1);
// Virtual losses should enable multiple searches to happen simultaneously
// without throwing an error...
for (int i = 0; i < 5; ++i) {
player->TreeSearch(5);
}
// Search should converge on D9 as only winning move.
auto best_move = ArgMax(root->edges, MctsNode::CmpN);
EXPECT_EQ(Coord::FromString("D9"), best_move);
// D9 should have a positive value.
EXPECT_LT(0, root->child_Q(best_move));
EXPECT_LE(20, root->N());
// Passing should be ineffective.
EXPECT_GT(0, root->child_Q(Coord::kPass));
// No virtual losses should be pending.
EXPECT_EQ(0, CountPendingVirtualLosses(root));
}
TEST(MctsPlayerTest, RidiculouslyParallelTreeSearch) {
auto player = CreateAlmostDonePlayer(0);
auto* root = player->root();
for (int i = 0; i < 10; ++i) {
// Test that an almost complete game will tree search with
// # parallelism > # legal moves.
player->TreeSearch(50);
}
// No virtual losses should be pending.
EXPECT_EQ(0, CountPendingVirtualLosses(root));
}
TEST(MctsPlayerTest, LongGameTreeSearch) {
auto player = CreateAlmostDonePlayer(kMaxSearchDepth - 2);
// Test that an almost complete game.
for (int i = 0; i < 10; ++i) {
player->TreeSearch(8);
}
EXPECT_EQ(0, CountPendingVirtualLosses(player->root()));
EXPECT_LT(0, player->root()->Q());
}
TEST(MctsPlayerTest, ColdStartParallelTreeSearch) {
MctsPlayer::Options options;
options.random_seed = 17;
auto player = absl::make_unique<TestablePlayer>(absl::Span<const float>(),
0.17, options);
auto* root = player->root();
// Test that parallel tree search doesn't trip on an empty tree.
EXPECT_EQ(0, root->N());
EXPECT_EQ(false, root->HasFlag(MctsNode::Flag::kExpanded));
player->TreeSearch(4);
EXPECT_EQ(0, CountPendingVirtualLosses(root));
// Even though we attempted to run 4 parallel searchs, the root should have
// only been selected once (the subsequent calls to SelectLeaf should have
// returned null).
EXPECT_EQ(1, root->N());
// 0.085 = average(0, 0.17), since 0 is the prior on the root.
EXPECT_NEAR(0.085, root->Q(), 0.01);
}
TEST(MctsPlayerTest, TreeSearchFailsafe) {
// Test that the failsafe works correctly. It can trigger if the MCTS
// repeatedly visits a finished game state.
std::array<float, kNumMoves> probs;
for (auto& p : probs) {
p = 0.001;
}
probs[Coord::kPass] = 1; // Make the dummy net always want to pass.
MctsPlayer::Options options;
options.random_seed = 17;
auto player = absl::make_unique<TestablePlayer>(probs, 0, options);
auto board = TestablePosition("");
board.PlayMove("pass");
player->InitializeGame(board);
player->TreeSearch(1);
EXPECT_EQ(0, CountPendingVirtualLosses(player->root()));
}
// When presented with a situation where the last move was a pass, and we have
// to decide whether to pass, it should be the first thing we check, but not
// more than that.
TEST(MctsPlayerTest, OnlyCheckGameEndOnce) {
BoardVisitor bv;
GroupVisitor gv;
Position position(&bv, &gv, Color::kBlack);
position.PlayMove({3, 3}); // B plays.
position.PlayMove({3, 4}); // W plays.
position.PlayMove({4, 3}); // B plays.
// W passes. If B passes too, B would lose by komi..
position.PlayMove(Coord::kPass);
auto player = absl::make_unique<TestablePlayer>(MctsPlayer::Options());
player->InitializeGame(position);
auto* root = player->root();
// Initialize the root
player->TreeSearch(1);
// Explore a child - should be a pass move.
player->TreeSearch(1);
EXPECT_EQ(1, root->child_N(Coord::kPass));
player->TreeSearch(1);
// Check that we didn't visit the pass node any more times.
EXPECT_EQ(1, root->child_N(Coord::kPass));
}
TEST(MctsPlayerTest, ExtractDataNormalEnd) {
auto player = absl::make_unique<TestablePlayer>(MctsPlayer::Options());
player->TreeSearch(1);
player->PlayMove(Coord::kPass);
player->TreeSearch(1);
player->PlayMove(Coord::kPass);
auto* root = player->root();
EXPECT_TRUE(root->game_over());
EXPECT_EQ(Color::kBlack, root->position.to_play());
ASSERT_EQ(2, player->history().size());
// White wins by komi
EXPECT_EQ(-1, player->result());
EXPECT_EQ("W+7.5", player->result_string());
}
TEST(MctsPlayerTest, ExtractDataResignEnd) {
auto player = absl::make_unique<TestablePlayer>(MctsPlayer::Options());
player->TreeSearch(1);
player->PlayMove({0, 0});
player->TreeSearch(1);
player->PlayMove(Coord::kPass);
player->TreeSearch(1);
player->PlayMove(Coord::kResign);
auto* root = player->root();
// Black is winning on the board.
EXPECT_LT(0, root->position.CalculateScore(player->options().komi));
EXPECT_EQ(-1, player->result());
EXPECT_EQ("W+R", player->result_string());
}
// Fake DualNet implementation used to verify that MctsPlayer symmetries work
// correctly. For each position on the board, MergeFeaturesNet returns a policy
// value depending on the feature planes of that square, if the square or any
// four connected neighbor is set true the policy is set to 0.01.
class MergeFeaturesNet : public DualNet {
public:
void RunMany(std::vector<const BoardFeatures*> features,
std::vector<Output*> outputs, std::string* model) override {
for (size_t i = 0; i < features.size(); ++i) {
Run(*features[i], outputs[i]);
}
if (model != nullptr) {
*model = "MergeFeaturesNet";
}
}
private:
void Run(const BoardFeatures& features, Output* output) {
for (int c = 0; c < kN * kN; ++c) {
bool present = false;
for (const auto n : kNeighborCoords[c]) {
const float* src = features.data() + n * DualNet::kNumStoneFeatures;
for (int f = 0; f < DualNet::kNumStoneFeatures - 1; ++f) {
if (src[f] != 0) {
present = true;
}
}
}
output->policy[c] = 0.01 * present;
}
output->policy[Coord::kPass] = 0.0;
output->value = 0.0;
}
};
TEST(MctsPlayerTest, SymmetriesTest) {
MctsPlayer::Options options;
options.random_seed = 17;
options.random_symmetry = true;
TestablePlayer player(absl::make_unique<MergeFeaturesNet>(), options);
// Without playing a move, all features planes should be zero except the last
// one (it's black's turn to play).
auto* root = player.root();
TestablePlayer::TreePath path(root, root);
root->AddVirtualLoss(root);
player.ProcessLeaves({&path, 1}, true);
for (int i = 0; i < kN * kN; ++i) {
ASSERT_EQ(0.0, root->child_P(i));
}
// Play an odd number of moves.
// Because it's white to play next, the output of the MergeFeaturesNet should
// only be non-zero except near locations where we have played.
std::vector<std::unique_ptr<MctsNode>> nodes;
std::vector<std::string> moves = {"B3", "F1", "C7"};
auto* parent = root;
for (const auto& move : moves) {
nodes.push_back(absl::make_unique<MctsNode>(parent, Coord::FromKgs(move)));
parent = nodes.back().get();
}
// 4 squares near B3 + 3 next to F1 + 4 near C7 should have equal policy.
float policy_fraction = 1.0 / (4 + 3 + 4);
// Run the MergeFeaturesNet many times to have a good chance of exercising all
// the symmetries.
for (int i = 0; i < 100; ++i) {
auto* leaf = nodes.back().get();
path.leaf = leaf;
leaf->AddVirtualLoss(root);
player.ProcessLeaves({&path, 1}, true);
ASSERT_EQ(0.0, leaf->child_P(Coord::FromKgs("pass")));
for (const auto move : moves) {
// Playing where stones exist is illegal and should have been marked as 0.
ASSERT_EQ(0.0, leaf->child_P(Coord::FromKgs(move)));
for (const auto n : kNeighborCoords[Coord::FromKgs(move)]) {
ASSERT_NEAR(policy_fraction, leaf->child_P(n), 1e-7);
}
}
}
}
TEST(MctsPlayerTest, ResetRoot) {
auto player = absl::make_unique<TestablePlayer>(MctsPlayer::Options());
auto* game_root = player->root();
auto c = Coord(12);
player->PlayMove(c);
EXPECT_EQ(c, player->root()->move);
player->ResetRoot();
EXPECT_EQ(game_root, player->root());
}
TEST(MctsPlayerTest, UndoMove) {
auto player = absl::make_unique<TestablePlayer>(MctsPlayer::Options());
// Can't undo without first playing a move.
EXPECT_FALSE(player->UndoMove());
player->PlayMove(Coord::kPass);
player->PlayMove(Coord::kPass);
auto* root = player->root();
EXPECT_TRUE(root->game_over());
EXPECT_EQ(Color::kBlack, root->position.to_play());
ASSERT_EQ(2, player->history().size());
EXPECT_EQ(-1, player->result());
EXPECT_EQ("W+7.5", player->result_string());
// Undo the last pass, the game should no longer be over.
EXPECT_TRUE(player->UndoMove());
root = player->root();
EXPECT_FALSE(root->game_over());
EXPECT_EQ(Coord::kPass, root->move);
EXPECT_EQ(Color::kWhite, root->position.to_play());
ASSERT_EQ(1, player->history().size());
}
} // namespace
} // namespace minigo
int main(int argc, char** argv) {
::testing::InitGoogleTest(&argc, argv);
::minigo::zobrist::Init(614944751);
return RUN_ALL_TESTS();
}