File size: 24,132 Bytes
079c32c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
// C++11

#include <iostream>
#include "cnode.h"
#include <algorithm>
#include <map>
#include <cassert>

#ifdef _WIN32
#include "..\..\common_lib\utils.cpp"
#else
#include "../../common_lib/utils.cpp"
#endif


namespace tree
{

    CSearchResults::CSearchResults()
    {
        /*
        Overview:
            Initialization of CSearchResults, the default result number is set to 0.
        */
        this->num = 0;
    }

    CSearchResults::CSearchResults(int num)
    {
        /*
        Overview:
            Initialization of CSearchResults with result number.
        */
        this->num = num;
        for (int i = 0; i < num; ++i)
        {
            this->search_paths.push_back(std::vector<CNode *>());
        }
    }

    CSearchResults::~CSearchResults() {}

    //*********************************************************

    CNode::CNode()
    {
        /*
        Overview:
            Initialization of CNode.
        */
        this->prior = 0;
        this->legal_actions = legal_actions;

        this->visit_count = 0;
        this->value_sum = 0;
        this->best_action = -1;
        this->to_play = 0;
        this->reward = 0.0;
    }

    CNode::CNode(float prior, std::vector<int> &legal_actions)
    {
        /*
        Overview:
            Initialization of CNode with prior value and legal actions.
        Arguments:
            - prior: the prior value of this node.
            - legal_actions: a vector of legal actions of this node.
        */
        this->prior = prior;
        this->legal_actions = legal_actions;

        this->visit_count = 0;
        this->value_sum = 0;
        this->best_action = -1;
        this->to_play = 0;
        this->current_latent_state_index = -1;
        this->batch_index = -1;
    }

    CNode::~CNode() {}

    void CNode::expand(int to_play, int current_latent_state_index, int batch_index, float reward, const std::vector<float> &policy_logits)
    {
        /*
        Overview:
            Expand the child nodes of the current node.
        Arguments:
            - to_play: which player to play the game in the current node.
            - current_latent_state_index: The index of latent state of the leaf node in the search path of the current node.
            - batch_index: The index of latent state of the leaf node in the search path of the current node.
            - reward: the reward of the current node.
            - policy_logits: the logit of the child nodes.
        */
        this->to_play = to_play;
        this->current_latent_state_index = current_latent_state_index;
        this->batch_index = batch_index;
        this->reward = reward;

        int action_num = policy_logits.size();
        if (this->legal_actions.size() == 0)
        {
            for (int i = 0; i < action_num; ++i)
            {
                this->legal_actions.push_back(i);
            }
        }
        float temp_policy;
        float policy_sum = 0.0;

        #ifdef _WIN32
        // 创建动态数组
        float* policy = new float[action_num];
        #else
        float policy[action_num];
        #endif

        float policy_max = FLOAT_MIN;
        for (auto a : this->legal_actions)
        {
            if (policy_max < policy_logits[a])
            {
                policy_max = policy_logits[a];
            }
        }

        for (auto a : this->legal_actions)
        {
            temp_policy = exp(policy_logits[a] - policy_max);
            policy_sum += temp_policy;
            policy[a] = temp_policy;
        }

        float prior;
        for (auto a : this->legal_actions)
        {
            prior = policy[a] / policy_sum;
            std::vector<int> tmp_empty;
            this->children[a] = CNode(prior, tmp_empty); // only for muzero/efficient zero, not support alphazero
        }
        
        #ifdef _WIN32
        // 释放数组内存
        delete[] policy;
        #else
        #endif
    }

    void CNode::add_exploration_noise(float exploration_fraction, const std::vector<float> &noises)
    {
        /*
        Overview:
            Add a noise to the prior of the child nodes.
        Arguments:
            - exploration_fraction: the fraction to add noise.
            - noises: the vector of noises added to each child node.
        */
        float noise, prior;
        for (int i = 0; i < this->legal_actions.size(); ++i)
        {
            noise = noises[i];
            CNode *child = this->get_child(this->legal_actions[i]);

            prior = child->prior;
            child->prior = prior * (1 - exploration_fraction) + noise * exploration_fraction;
        }
    }

    float CNode::compute_mean_q(int isRoot, float parent_q, float discount_factor)
    {
        /*
        Overview:
            Compute the mean q value of the current node.
        Arguments:
            - isRoot: whether the current node is a root node.
            - parent_q: the q value of the parent node.
            - discount_factor: the discount_factor of reward.
        */
        float total_unsigned_q = 0.0;
        int total_visits = 0;
        for (auto a : this->legal_actions)
        {
            CNode *child = this->get_child(a);
            if (child->visit_count > 0)
            {
                float true_reward = child->reward;
                float qsa = true_reward + discount_factor * child->value();
                total_unsigned_q += qsa;
                total_visits += 1;
            }
        }

        float mean_q = 0.0;
        if (isRoot && total_visits > 0)
        {
            mean_q = (total_unsigned_q) / (total_visits);
        }
        else
        {
            mean_q = (parent_q + total_unsigned_q) / (total_visits + 1);
        }
        return mean_q;
    }

    void CNode::print_out()
    {
        return;
    }

    int CNode::expanded()
    {
        /*
        Overview:
            Return whether the current node is expanded.
        */
        return this->children.size() > 0;
    }

    float CNode::value()
    {
        /*
        Overview:
            Return the real value of the current tree.
        */
        float true_value = 0.0;
        if (this->visit_count == 0)
        {
            return true_value;
        }
        else
        {
            true_value = this->value_sum / this->visit_count;
            return true_value;
        }
    }

    std::vector<int> CNode::get_trajectory()
    {
        /*
        Overview:
            Find the current best trajectory starts from the current node.
        Outputs:
            - traj: a vector of node index, which is the current best trajectory from this node.
        */
        std::vector<int> traj;

        CNode *node = this;
        int best_action = node->best_action;
        while (best_action >= 0)
        {
            traj.push_back(best_action);

            node = node->get_child(best_action);
            best_action = node->best_action;
        }
        return traj;
    }

    std::vector<int> CNode::get_children_distribution()
    {
        /*
        Overview:
            Get the distribution of child nodes in the format of visit_count.
        Outputs:
            - distribution: a vector of distribution of child nodes in the format of visit count (i.e. [1,3,0,2,5]).
        */
        std::vector<int> distribution;
        if (this->expanded())
        {
            for (auto a : this->legal_actions)
            {
                CNode *child = this->get_child(a);
                distribution.push_back(child->visit_count);
            }
        }
        return distribution;
    }

    CNode *CNode::get_child(int action)
    {
        /*
        Overview:
            Get the child node corresponding to the input action.
        Arguments:
            - action: the action to get child.
        */
        return &(this->children[action]);
    }

    //*********************************************************

    CRoots::CRoots()
    {
        /*
        Overview:
            The initialization of CRoots.
        */
        this->root_num = 0;
    }

    CRoots::CRoots(int root_num, std::vector<std::vector<int> > &legal_actions_list)
    {
        /*
        Overview:
            The initialization of CRoots with root num and legal action lists.
        Arguments:
            - root_num: the number of the current root.
            - legal_action_list: the vector of the legal action of this root.
        */
        this->root_num = root_num;
        this->legal_actions_list = legal_actions_list;

        for (int i = 0; i < root_num; ++i)
        {
            this->roots.push_back(CNode(0, this->legal_actions_list[i]));
        }
    }

    CRoots::~CRoots() {}

    void CRoots::prepare(float root_noise_weight, const std::vector<std::vector<float> > &noises, const std::vector<float> &rewards, const std::vector<std::vector<float> > &policies, std::vector<int> &to_play_batch)
    {
        /*
        Overview:
            Expand the roots and add noises.
        Arguments:
            - root_noise_weight: the exploration fraction of roots
            - noises: the vector of noise add to the roots.
            - rewards: the vector of rewards of each root.
            - policies: the vector of policy logits of each root.
            - to_play_batch: the vector of the player side of each root.
        */
        for (int i = 0; i < this->root_num; ++i)
        {
            this->roots[i].expand(to_play_batch[i], 0, i, rewards[i], policies[i]);
            this->roots[i].add_exploration_noise(root_noise_weight, noises[i]);

            this->roots[i].visit_count += 1;
        }
    }

    void CRoots::prepare_no_noise(const std::vector<float> &rewards, const std::vector<std::vector<float> > &policies, std::vector<int> &to_play_batch)
    {
        /*
        Overview:
            Expand the roots without noise.
        Arguments:
            - rewards: the vector of rewards of each root.
            - policies: the vector of policy logits of each root.
            - to_play_batch: the vector of the player side of each root.
        */
        for (int i = 0; i < this->root_num; ++i)
        {
            this->roots[i].expand(to_play_batch[i], 0, i, rewards[i], policies[i]);

            this->roots[i].visit_count += 1;
        }
    }

    void CRoots::clear()
    {
        /*
        Overview:
            Clear the roots vector.
        */
        this->roots.clear();
    }

    std::vector<std::vector<int> > CRoots::get_trajectories()
    {
        /*
        Overview:
            Find the current best trajectory starts from each root.
        Outputs:
            - traj: a vector of node index, which is the current best trajectory from each root.
        */
        std::vector<std::vector<int> > trajs;
        trajs.reserve(this->root_num);

        for (int i = 0; i < this->root_num; ++i)
        {
            trajs.push_back(this->roots[i].get_trajectory());
        }
        return trajs;
    }

    std::vector<std::vector<int> > CRoots::get_distributions()
    {
        /*
        Overview:
            Get the children distribution of each root.
        Outputs:
            - distribution: a vector of distribution of child nodes in the format of visit count (i.e. [1,3,0,2,5]).
        */
        std::vector<std::vector<int> > distributions;
        distributions.reserve(this->root_num);

        for (int i = 0; i < this->root_num; ++i)
        {
            distributions.push_back(this->roots[i].get_children_distribution());
        }
        return distributions;
    }

    std::vector<float> CRoots::get_values()
    {
        /*
        Overview:
            Return the real value of each root.
        */
        std::vector<float> values;
        for (int i = 0; i < this->root_num; ++i)
        {
            values.push_back(this->roots[i].value());
        }
        return values;
    }

    //*********************************************************
    //
    void update_tree_q(CNode *root, tools::CMinMaxStats &min_max_stats, float discount_factor, int players)
    {
        /*
        Overview:
            Update the q value of the root and its child nodes.
        Arguments:
            - root: the root that update q value from.
            - min_max_stats: a tool used to min-max normalize the q value.
            - discount_factor: the discount factor of reward.
            - players: the number of players.
        */
        std::stack<CNode *> node_stack;
        node_stack.push(root);
        while (node_stack.size() > 0)
        {
            CNode *node = node_stack.top();
            node_stack.pop();

            if (node != root)
            {
                //                # NOTE: in self-play-mode, value_prefix is not calculated according to the perspective of current player of node,
                //                # but treated as 1 player, just for obtaining the true reward in the perspective of current player of node.
                //                # true_reward = node.value_prefix - (- parent_value_prefix)
                //                float true_reward = node->value_prefix - node->parent_value_prefix;
                float true_reward = node->reward;

                float qsa;
                if (players == 1)
                    qsa = true_reward + discount_factor * node->value();
                else if (players == 2)
                    // TODO(pu):
                    qsa = true_reward + discount_factor * (-1) * node->value();

                min_max_stats.update(qsa);
            }

            for (auto a : node->legal_actions)
            {
                CNode *child = node->get_child(a);
                if (child->expanded())
                {
                    node_stack.push(child);
                }
            }
        }
    }

    void cbackpropagate(std::vector<CNode *> &search_path, tools::CMinMaxStats &min_max_stats, int to_play, float value, float discount_factor)
    {
        /*
        Overview:
            Update the value sum and visit count of nodes along the search path.
        Arguments:
            - search_path: a vector of nodes on the search path.
            - min_max_stats: a tool used to min-max normalize the q value.
            - to_play: which player to play the game in the current node.
            - value: the value to propagate along the search path.
            - discount_factor: the discount factor of reward.
        */
        assert(to_play == -1 || to_play == 1 || to_play == 2);
        if (to_play == -1)
        {           
            // for play-with-bot-mode
            float bootstrap_value = value;
            int path_len = search_path.size();
            for (int i = path_len - 1; i >= 0; --i)
            {
                CNode *node = search_path[i];
                node->value_sum += bootstrap_value;
                node->visit_count += 1;

                float true_reward = node->reward;

                min_max_stats.update(true_reward + discount_factor * node->value());

                bootstrap_value = true_reward + discount_factor * bootstrap_value;
            }
        }
        else
        {
            // for self-play-mode
            float bootstrap_value = value;
            int path_len = search_path.size();
            for (int i = path_len - 1; i >= 0; --i)
            {
                CNode *node = search_path[i];
                if (node->to_play == to_play)
                    node->value_sum += bootstrap_value;
                else
                    node->value_sum += -bootstrap_value;
                node->visit_count += 1;

                // NOTE: in self-play-mode, value_prefix is not calculated according to the perspective of current player of node,
                // but treated as 1 player, just for obtaining the true reward in the perspective of current player of node.
                //                float true_reward = node->value_prefix - parent_value_prefix;
                float true_reward = node->reward;

                // TODO(pu): why in muzero-general is - node.value
                min_max_stats.update(true_reward + discount_factor * -node->value());

                if (node->to_play == to_play)
                    bootstrap_value = -true_reward + discount_factor * bootstrap_value;
                else
                    bootstrap_value = true_reward + discount_factor * bootstrap_value;
            }
        }
    }

    void cbatch_backpropagate(int current_latent_state_index, float discount_factor, const std::vector<float> &value_prefixs, const std::vector<float> &values, const std::vector<std::vector<float> > &policies, tools::CMinMaxStatsList *min_max_stats_lst, CSearchResults &results, std::vector<int> &to_play_batch)
    {
        /*
        Overview:
            Expand the nodes along the search path and update the infos.
        Arguments:
            - current_latent_state_index: The index of latent state of the leaf node in the search path.
            - discount_factor: the discount factor of reward.
            - value_prefixs: the value prefixs of nodes along the search path.
            - values: the values to propagate along the search path.
            - policies: the policy logits of nodes along the search path.
            - min_max_stats: a tool used to min-max normalize the q value.
            - results: the search results.
            - to_play_batch: the batch of which player is playing on this node.
        */
        for (int i = 0; i < results.num; ++i)
        {
            results.nodes[i]->expand(to_play_batch[i], current_latent_state_index, i, value_prefixs[i], policies[i]);
            cbackpropagate(results.search_paths[i], min_max_stats_lst->stats_lst[i], to_play_batch[i], values[i], discount_factor);
        }
    }

    int cselect_child(CNode *root, tools::CMinMaxStats &min_max_stats, int pb_c_base, float pb_c_init, float discount_factor, float mean_q, int players)
    {
        /*
        Overview:
            Select the child node of the roots according to ucb scores.
        Arguments:
            - root: the roots to select the child node.
            - min_max_stats: a tool used to min-max normalize the score.
            - pb_c_base: constants c2 in muzero.
            - pb_c_init: constants c1 in muzero.
            - disount_factor: the discount factor of reward.
            - mean_q: the mean q value of the parent node.
            - players: the number of players.
        Outputs:
            - action: the action to select.
        */
        float max_score = FLOAT_MIN;
        const float epsilon = 0.000001;
        std::vector<int> max_index_lst;
        for (auto a : root->legal_actions)
        {

            CNode *child = root->get_child(a);
            float temp_score = cucb_score(child, min_max_stats, mean_q, root->visit_count - 1, pb_c_base, pb_c_init, discount_factor, players);

            if (max_score < temp_score)
            {
                max_score = temp_score;

                max_index_lst.clear();
                max_index_lst.push_back(a);
            }
            else if (temp_score >= max_score - epsilon)
            {
                max_index_lst.push_back(a);
            }
        }

        int action = 0;
        if (max_index_lst.size() > 0)
        {
            int rand_index = rand() % max_index_lst.size();
            action = max_index_lst[rand_index];
        }
        return action;
    }

    float cucb_score(CNode *child, tools::CMinMaxStats &min_max_stats, float parent_mean_q, float total_children_visit_counts, float pb_c_base, float pb_c_init, float discount_factor, int players)
    {
        /*
        Overview:
            Compute the ucb score of the child.
        Arguments:
            - child: the child node to compute ucb score.
            - min_max_stats: a tool used to min-max normalize the score.
            - mean_q: the mean q value of the parent node.
            - total_children_visit_counts: the total visit counts of the child nodes of the parent node.
            - pb_c_base: constants c2 in muzero.
            - pb_c_init: constants c1 in muzero.
            - disount_factor: the discount factor of reward.
            - players: the number of players.
        Outputs:
            - ucb_value: the ucb score of the child.
        */
        float pb_c = 0.0, prior_score = 0.0, value_score = 0.0;
        pb_c = log((total_children_visit_counts + pb_c_base + 1) / pb_c_base) + pb_c_init;
        pb_c *= (sqrt(total_children_visit_counts) / (child->visit_count + 1));

        prior_score = pb_c * child->prior;
        if (child->visit_count == 0)
        {
            value_score = parent_mean_q;
        }
        else
        {
            float true_reward = child->reward;
            if (players == 1)
                value_score = true_reward + discount_factor * child->value();
            else if (players == 2)
                value_score = true_reward + discount_factor * (-child->value());
        }

        value_score = min_max_stats.normalize(value_score);

        if (value_score < 0)
            value_score = 0;
        if (value_score > 1)
            value_score = 1;

        float ucb_value = prior_score + value_score;
        return ucb_value;
    }

    void cbatch_traverse(CRoots *roots, int pb_c_base, float pb_c_init, float discount_factor, tools::CMinMaxStatsList *min_max_stats_lst, CSearchResults &results, std::vector<int> &virtual_to_play_batch)
    {
        /*
        Overview:
            Search node path from the roots.
        Arguments:
            - roots: the roots that search from.
            - pb_c_base: constants c2 in muzero.
            - pb_c_init: constants c1 in muzero.
            - disount_factor: the discount factor of reward.
            - min_max_stats: a tool used to min-max normalize the score.
            - results: the search results.
            - virtual_to_play_batch: the batch of which player is playing on this node.
        */
        // set seed
        get_time_and_set_rand_seed();

        int last_action = -1;
        float parent_q = 0.0;
        results.search_lens = std::vector<int>();

        int players = 0;
        int largest_element = *max_element(virtual_to_play_batch.begin(), virtual_to_play_batch.end()); // 0 or 2
        if (largest_element == -1)
            players = 1;
        else
            players = 2;

        for (int i = 0; i < results.num; ++i)
        {
            CNode *node = &(roots->roots[i]);
            int is_root = 1;
            int search_len = 0;
            results.search_paths[i].push_back(node);

            while (node->expanded())
            {
                float mean_q = node->compute_mean_q(is_root, parent_q, discount_factor);
                is_root = 0;
                parent_q = mean_q;

                int action = cselect_child(node, min_max_stats_lst->stats_lst[i], pb_c_base, pb_c_init, discount_factor, mean_q, players);
                if (players > 1)
                {
                    assert(virtual_to_play_batch[i] == 1 || virtual_to_play_batch[i] == 2);
                    if (virtual_to_play_batch[i] == 1)
                        virtual_to_play_batch[i] = 2;
                    else
                        virtual_to_play_batch[i] = 1;
                }

                node->best_action = action;
                // next
                node = node->get_child(action);
                last_action = action;
                results.search_paths[i].push_back(node);
                search_len += 1;
            }

            CNode *parent = results.search_paths[i][results.search_paths[i].size() - 2];

            results.latent_state_index_in_search_path.push_back(parent->current_latent_state_index);
            results.latent_state_index_in_batch.push_back(parent->batch_index);

            results.last_actions.push_back(last_action);
            results.search_lens.push_back(search_len);
            results.nodes.push_back(node);
            results.virtual_to_play_batchs.push_back(virtual_to_play_batch[i]);
        }
    }

}