15 #include "joseki/base.h"
17 #include "playout/elo.h"
18 #include "playout/moggy.h"
19 #include "playout/light.h"
20 #include "tactics/util.h"
22 #include "uct/dynkomi.h"
23 #include "uct/internal.h"
24 #include "uct/plugins.h"
25 #include "uct/prior.h"
26 #include "uct/search.h"
27 #include "uct/slave.h"
32 struct uct_policy
*policy_ucb1_init(struct uct
*u
, char *arg
);
33 struct uct_policy
*policy_ucb1amaf_init(struct uct
*u
, char *arg
);
34 static void uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
);
36 /* Maximal simulation length. */
37 #define MC_GAMELEN MAX_GAMELEN
41 setup_state(struct uct
*u
, struct board
*b
, enum stone color
)
43 u
->t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0,
44 u
->max_pruned_size
, u
->pruning_threshold
, u
->local_tree_aging
, u
->stats_hbits
);
46 fast_srandom(u
->force_seed
);
48 fprintf(stderr
, "Fresh board with random seed %lu\n", fast_getseed());
49 //board_print(b, stderr);
50 if (!u
->no_tbook
&& b
->moves
== 0) {
51 assert(color
== S_BLACK
);
57 reset_state(struct uct
*u
)
60 tree_done(u
->t
); u
->t
= NULL
;
64 setup_dynkomi(struct uct
*u
, struct board
*b
, enum stone to_play
)
66 if (u
->t
->use_extra_komi
&& !u
->pondering
&& u
->dynkomi
->permove
)
67 u
->t
->extra_komi
= u
->dynkomi
->permove(u
->dynkomi
, b
, u
->t
);
68 else if (!u
->t
->use_extra_komi
)
73 uct_prepare_move(struct uct
*u
, struct board
*b
, enum stone color
)
76 /* Verify that we have sane state. */
78 assert(u
->t
&& b
->moves
);
79 if (color
!= stone_other(u
->t
->root_color
)) {
80 fprintf(stderr
, "Fatal: Non-alternating play detected %d %d\n",
81 color
, u
->t
->root_color
);
84 uct_htable_reset(u
->t
);
87 /* We need fresh state. */
89 setup_state(u
, b
, color
);
92 u
->ownermap
.playouts
= 0;
93 memset(u
->ownermap
.map
, 0, board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
94 u
->played_own
= u
->played_all
= 0;
98 dead_group_list(struct uct
*u
, struct board
*b
, struct move_queue
*mq
)
100 struct group_judgement gj
;
102 gj
.gs
= alloca(board_size2(b
) * sizeof(gj
.gs
[0]));
103 board_ownermap_judge_group(b
, &u
->ownermap
, &gj
);
104 groups_of_status(b
, &gj
, GS_DEAD
, mq
);
108 uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
)
110 if (u
->ownermap
.playouts
< GJ_MINGAMES
)
113 struct move_queue mq
= { .moves
= 0 };
114 dead_group_list(u
, b
, &mq
);
115 if (pass_all_alive
&& mq
.moves
> 0)
116 return false; // We need to remove some dead groups first.
117 return pass_is_safe(b
, color
, &mq
);
121 uct_printhook_ownermap(struct board
*board
, coord_t c
, char *s
, char *end
)
123 struct uct
*u
= board
->es
;
128 const char chr
[] = ":XO,"; // dame, black, white, unclear
129 const char chm
[] = ":xo,";
130 char ch
= chr
[board_ownermap_judge_point(&u
->ownermap
, c
, GJ_THRES
)];
131 if (ch
== ',') { // less precise estimate then?
132 ch
= chm
[board_ownermap_judge_point(&u
->ownermap
, c
, 0.67)];
134 s
+= snprintf(s
, end
- s
, "%c ", ch
);
139 uct_notify_play(struct engine
*e
, struct board
*b
, struct move
*m
)
141 struct uct
*u
= e
->data
;
143 /* No state, create one - this is probably game beginning
144 * and we need to load the opening tbook right now. */
145 uct_prepare_move(u
, b
, m
->color
);
149 /* Stop pondering, required by tree_promote_at() */
150 uct_pondering_stop(u
);
151 if (UDEBUGL(2) && u
->slave
)
152 tree_dump(u
->t
, u
->dumpthres
);
154 if (is_resign(m
->coord
)) {
160 /* Promote node of the appropriate move to the tree root. */
162 if (!tree_promote_at(u
->t
, b
, m
->coord
)) {
164 fprintf(stderr
, "Warning: Cannot promote move node! Several play commands in row?\n");
169 /* If we are a slave in a distributed engine, start pondering once
170 * we know which move we actually played. See uct_genmove() about
171 * the check for pass. */
172 if (u
->pondering_opt
&& u
->slave
&& m
->color
== u
->my_color
&& !is_pass(m
->coord
))
173 uct_pondering_start(u
, b
, u
->t
, stone_other(m
->color
));
179 uct_result(struct engine
*e
, struct board
*b
)
181 struct uct
*u
= e
->data
;
182 static char reply
[1024];
186 enum stone color
= u
->t
->root_color
;
187 struct tree_node
*n
= u
->t
->root
;
188 snprintf(reply
, 1024, "%s %s %d %.2f %.1f",
189 stone2str(color
), coord2sstr(n
->coord
, b
),
190 n
->u
.playouts
, tree_node_get_value(u
->t
, -1, n
->u
.value
),
191 u
->t
->use_extra_komi
? u
->t
->extra_komi
: 0);
196 uct_chat(struct engine
*e
, struct board
*b
, char *cmd
)
198 struct uct
*u
= e
->data
;
199 static char reply
[1024];
201 cmd
+= strspn(cmd
, " \n\t");
202 if (!strncasecmp(cmd
, "winrate", 7)) {
204 return "no game context (yet?)";
205 enum stone color
= u
->t
->root_color
;
206 struct tree_node
*n
= u
->t
->root
;
207 snprintf(reply
, 1024, "In %d playouts at %d threads, %s %s can win with %.2f%% probability",
208 n
->u
.playouts
, u
->threads
, stone2str(color
), coord2sstr(n
->coord
, b
),
209 tree_node_get_value(u
->t
, -1, n
->u
.value
) * 100);
210 if (u
->t
->use_extra_komi
&& abs(u
->t
->extra_komi
) >= 0.5) {
211 sprintf(reply
+ strlen(reply
), ", while self-imposing extra komi %.1f",
221 uct_dead_group_list(struct engine
*e
, struct board
*b
, struct move_queue
*mq
)
223 struct uct
*u
= e
->data
;
225 /* This means the game is probably over, no use pondering on. */
226 uct_pondering_stop(u
);
228 if (u
->pass_all_alive
)
229 return; // no dead groups
231 bool mock_state
= false;
234 /* No state, but we cannot just back out - we might
235 * have passed earlier, only assuming some stones are
236 * dead, and then re-connected, only to lose counting
237 * when all stones are assumed alive. */
238 uct_prepare_move(u
, b
, S_BLACK
); assert(u
->t
);
241 /* Make sure the ownermap is well-seeded. */
242 while (u
->ownermap
.playouts
< GJ_MINGAMES
)
243 uct_playout(u
, b
, S_BLACK
, u
->t
);
244 /* Show the ownermap: */
246 board_print_custom(b
, stderr
, uct_printhook_ownermap
);
248 dead_group_list(u
, b
, mq
);
251 /* Clean up the mock state in case we will receive
252 * a genmove; we could get a non-alternating-move
253 * error from uct_prepare_move() in that case otherwise. */
259 playout_policy_done(struct playout_policy
*p
)
261 if (p
->done
) p
->done(p
);
262 if (p
->data
) free(p
->data
);
267 uct_done(struct engine
*e
)
269 /* This is called on engine reset, especially when clear_board
270 * is received and new game should begin. */
271 struct uct
*u
= e
->data
;
272 uct_pondering_stop(u
);
273 if (u
->t
) reset_state(u
);
274 free(u
->ownermap
.map
);
277 free(u
->random_policy
);
278 playout_policy_done(u
->playout
);
279 uct_prior_done(u
->prior
);
280 joseki_done(u
->jdict
);
281 pluginset_done(u
->plugins
);
286 /* Run time-limited MCTS search on foreground. */
288 uct_search(struct uct
*u
, struct board
*b
, struct time_info
*ti
, enum stone color
, struct tree
*t
)
290 struct uct_search_state s
;
291 uct_search_start(u
, b
, color
, t
, ti
, &s
);
292 if (UDEBUGL(2) && s
.base_playouts
> 0)
293 fprintf(stderr
, "<pre-simulated %d games>\n", s
.base_playouts
);
295 /* The search tree is ctx->t. This is currently == . It is important
296 * to reference ctx->t directly since the
297 * thread manager will swap the tree pointer asynchronously. */
299 /* Now, just periodically poll the search tree. */
300 /* Note that in case of TD_GAMES, threads will terminate independently
301 * of the uct_search_check_stop() signalization. */
303 time_sleep(TREE_BUSYWAIT_INTERVAL
);
304 /* TREE_BUSYWAIT_INTERVAL should never be less than desired time, or the
305 * time control is broken. But if it happens to be less, we still search
306 * at least 100ms otherwise the move is completely random. */
308 int i
= uct_search_games(&s
);
309 /* Print notifications etc. */
310 uct_search_progress(u
, b
, color
, t
, ti
, &s
, i
);
311 /* Check if we should stop the search. */
312 if (uct_search_check_stop(u
, b
, color
, t
, ti
, &s
, i
))
316 struct uct_thread_ctx
*ctx
= uct_search_stop();
317 if (UDEBUGL(2)) tree_dump(t
, u
->dumpthres
);
319 fprintf(stderr
, "(avg score %f/%d value %f/%d)\n",
320 u
->dynkomi
->score
.value
, u
->dynkomi
->score
.playouts
,
321 u
->dynkomi
->value
.value
, u
->dynkomi
->value
.playouts
);
323 uct_progress_status(u
, t
, color
, ctx
->games
);
325 u
->played_own
+= ctx
->games
;
329 /* Start pondering background with @color to play. */
331 uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
)
334 fprintf(stderr
, "Starting to ponder with color %s\n", stone2str(stone_other(color
)));
337 /* We need a local board copy to ponder upon. */
338 struct board
*b
= malloc2(sizeof(*b
)); board_copy(b
, b0
);
340 /* *b0 did not have the genmove'd move played yet. */
341 struct move m
= { t
->root
->coord
, t
->root_color
};
342 int res
= board_play(b
, &m
);
344 setup_dynkomi(u
, b
, stone_other(m
.color
));
346 /* Start MCTS manager thread "headless". */
347 static struct uct_search_state s
;
348 uct_search_start(u
, b
, color
, t
, NULL
, &s
);
351 /* uct_search_stop() frontend for the pondering (non-genmove) mode, and
352 * to stop the background search for a slave in the distributed engine. */
354 uct_pondering_stop(struct uct
*u
)
356 if (!thread_manager_running
)
359 /* Stop the thread manager. */
360 struct uct_thread_ctx
*ctx
= uct_search_stop();
362 if (u
->pondering
) fprintf(stderr
, "(pondering) ");
363 uct_progress_status(u
, ctx
->t
, ctx
->color
, ctx
->games
);
367 u
->pondering
= false;
373 uct_genmove_setup(struct uct
*u
, struct board
*b
, enum stone color
)
375 if (b
->superko_violation
) {
376 fprintf(stderr
, "!!! WARNING: SUPERKO VIOLATION OCCURED BEFORE THIS MOVE\n");
377 fprintf(stderr
, "Maybe you play with situational instead of positional superko?\n");
378 fprintf(stderr
, "I'm going to ignore the violation, but note that I may miss\n");
379 fprintf(stderr
, "some moves valid under this ruleset because of this.\n");
380 b
->superko_violation
= false;
383 uct_prepare_move(u
, b
, color
);
388 /* How to decide whether to use dynkomi in this game? Since we use
389 * pondering, it's not simple "who-to-play" matter. Decide based on
390 * the last genmove issued. */
391 u
->t
->use_extra_komi
= !!(u
->dynkomi_mask
& color
);
392 /* Moreover, we do not use extra komi at the game end - we are not
393 * to fool ourselves at this point. */
394 if (board_estimated_moves_left(b
) <= MIN_MOVES_LEFT
)
395 u
->t
->use_extra_komi
= false;
396 setup_dynkomi(u
, b
, color
);
398 if (b
->rules
== RULES_JAPANESE
)
399 u
->territory_scoring
= true;
401 /* Make pessimistic assumption about komi for Japanese rules to
402 * avoid losing by 0.5 when winning by 0.5 with Chinese rules.
403 * The rules usually give the same winner if the integer part of komi
404 * is odd so we adjust the komi only if it is even (for a board of
405 * odd size). We are not trying to get an exact evaluation for rare
406 * cases of seki. For details see http://home.snafu.de/jasiek/parity.html */
407 if (u
->territory_scoring
&& (((int)floor(b
->komi
) + board_size(b
)) & 1)) {
408 b
->komi
+= (color
== S_BLACK
? 1.0 : -1.0);
410 fprintf(stderr
, "Setting komi to %.1f assuming Japanese rules\n",
416 uct_genmove(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
418 double start_time
= time_now();
419 struct uct
*u
= e
->data
;
420 uct_pondering_stop(u
);
421 uct_genmove_setup(u
, b
, color
);
423 /* Start the Monte Carlo Tree Search! */
424 int base_playouts
= u
->t
->root
->u
.playouts
;
425 int played_games
= uct_search(u
, b
, ti
, color
, u
->t
);
428 struct tree_node
*best
;
429 best
= uct_search_result(u
, b
, color
, pass_all_alive
, played_games
, base_playouts
, &best_coord
);
432 double time
= time_now() - start_time
+ 0.000001; /* avoid divide by zero */
433 fprintf(stderr
, "genmove in %0.2fs (%d games/s, %d games/s/thread)\n",
434 time
, (int)(played_games
/time
), (int)(played_games
/time
/u
->threads
));
438 /* Pass or resign. */
440 return coord_copy(best_coord
);
442 tree_promote_node(u
->t
, &best
);
444 /* After a pass, pondering is harmful for two reasons:
445 * (i) We might keep pondering even when the game is over.
446 * Of course this is the case for opponent resign as well.
447 * (ii) More importantly, the ownermap will get skewed since
448 * the UCT will start cutting off any playouts. */
449 if (u
->pondering_opt
&& !is_pass(best
->coord
)) {
450 uct_pondering_start(u
, b
, u
->t
, stone_other(color
));
452 return coord_copy(best_coord
);
457 uct_gentbook(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
)
459 struct uct
*u
= e
->data
;
460 if (!u
->t
) uct_prepare_move(u
, b
, color
);
463 if (ti
->dim
== TD_GAMES
) {
464 /* Don't count in games that already went into the tbook. */
465 ti
->len
.games
+= u
->t
->root
->u
.playouts
;
467 uct_search(u
, b
, ti
, color
, u
->t
);
469 assert(ti
->dim
== TD_GAMES
);
470 tree_save(u
->t
, b
, ti
->len
.games
/ 100);
476 uct_dumptbook(struct engine
*e
, struct board
*b
, enum stone color
)
478 struct uct
*u
= e
->data
;
479 struct tree
*t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0,
480 u
->max_pruned_size
, u
->pruning_threshold
, u
->local_tree_aging
, 0);
488 uct_evaluate(struct engine
*e
, struct board
*b
, struct time_info
*ti
, coord_t c
, enum stone color
)
490 struct uct
*u
= e
->data
;
494 struct move m
= { c
, color
};
495 int res
= board_play(&b2
, &m
);
498 color
= stone_other(color
);
500 if (u
->t
) reset_state(u
);
501 uct_prepare_move(u
, &b2
, color
);
505 uct_search(u
, &b2
, ti
, color
, u
->t
);
506 struct tree_node
*best
= u
->policy
->choose(u
->policy
, u
->t
->root
, &b2
, color
, resign
);
508 bestval
= NAN
; // the opponent has no reply!
510 bestval
= tree_node_get_value(u
->t
, 1, best
->u
.value
);
513 reset_state(u
); // clean our junk
515 return isnan(bestval
) ? NAN
: 1.0f
- bestval
;
520 uct_state_init(char *arg
, struct board
*b
)
522 struct uct
*u
= calloc2(1, sizeof(struct uct
));
523 bool using_elo
= false;
525 u
->debug_level
= debug_level
;
526 u
->gamelen
= MC_GAMELEN
;
527 u
->resign_threshold
= 0.2;
528 u
->sure_win_threshold
= 0.85;
530 u
->significant_threshold
= 50;
533 u
->playout_amaf
= true;
534 u
->playout_amaf_nakade
= false;
535 u
->amaf_prior
= false;
536 u
->max_tree_size
= 3072ULL * 1048576;
537 u
->pruning_threshold
= 0;
540 u
->thread_model
= TM_TREEVL
;
541 u
->virtual_loss
= true;
543 u
->fuseki_end
= 20; // max time at 361*20% = 72 moves (our 36th move, still 99 to play)
544 u
->yose_start
= 40; // (100-40-25)*361/100/2 = 63 moves still to play by us then
545 u
->bestr_ratio
= 0.02;
546 // 2.5 is clearly too much, but seems to compensate well for overly stern time allocations.
547 // TODO: Further tuning and experiments with better time allocation schemes.
548 u
->best2_ratio
= 2.5;
550 u
->val_scale
= 0.04; u
->val_points
= 40;
551 u
->dynkomi_interval
= 1000;
552 u
->dynkomi_mask
= S_BLACK
| S_WHITE
;
555 u
->local_tree_aging
= 2;
557 u
->plugins
= pluginset_init(b
);
559 u
->jdict
= joseki_load(b
->size
);
562 char *optspec
, *next
= arg
;
565 next
+= strcspn(next
, ",");
566 if (*next
) { *next
++ = 0; } else { *next
= 0; }
568 char *optname
= optspec
;
569 char *optval
= strchr(optspec
, '=');
570 if (optval
) *optval
++ = 0;
574 if (!strcasecmp(optname
, "debug")) {
576 u
->debug_level
= atoi(optval
);
579 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
580 /* When dumping the UCT tree on output, include
581 * nodes with at least this many playouts.
582 * (This value is re-scaled "intelligently"
583 * in case of very large trees.) */
584 u
->dumpthres
= atoi(optval
);
585 } else if (!strcasecmp(optname
, "resign_threshold") && optval
) {
586 /* Resign when this ratio of games is lost
587 * after GJ_MINGAMES sample is taken. */
588 u
->resign_threshold
= atof(optval
);
589 } else if (!strcasecmp(optname
, "sure_win_threshold") && optval
) {
590 /* Stop reading when this ratio of games is won
591 * after PLAYOUT_EARLY_BREAK_MIN sample is
592 * taken. (Prevents stupid time losses,
593 * friendly to human opponents.) */
594 u
->sure_win_threshold
= atof(optval
);
595 } else if (!strcasecmp(optname
, "force_seed") && optval
) {
596 /* Set RNG seed at the tree setup. */
597 u
->force_seed
= atoi(optval
);
598 } else if (!strcasecmp(optname
, "no_tbook")) {
599 /* Disable UCT opening tbook. */
601 } else if (!strcasecmp(optname
, "pass_all_alive")) {
602 /* Whether to consider passing only after all
603 * dead groups were removed from the board;
604 * this is like all genmoves are in fact
605 * kgs-genmove_cleanup. */
606 u
->pass_all_alive
= !optval
|| atoi(optval
);
607 } else if (!strcasecmp(optname
, "territory_scoring")) {
608 /* Use territory scoring (default is area scoring).
609 * An explicit kgs-rules command overrides this. */
610 u
->territory_scoring
= !optval
|| atoi(optval
);
611 } else if (!strcasecmp(optname
, "banner") && optval
) {
612 /* Additional banner string. This must come as the
613 * last engine parameter. */
614 if (*next
) *--next
= ',';
615 u
->banner
= strdup(optval
);
617 } else if (!strcasecmp(optname
, "plugin") && optval
) {
618 /* Load an external plugin; filename goes before the colon,
619 * extra arguments after the colon. */
620 char *pluginarg
= strchr(optval
, ':');
623 plugin_load(u
->plugins
, optval
, pluginarg
);
625 /** UCT behavior and policies */
627 } else if ((!strcasecmp(optname
, "policy")
628 /* Node selection policy. ucb1amaf is the
629 * default policy implementing RAVE, while
630 * ucb1 is the simple exploration/exploitation
631 * policy. Policies can take further extra
633 || !strcasecmp(optname
, "random_policy")) && optval
) {
634 /* A policy to be used randomly with small
635 * chance instead of the default policy. */
636 char *policyarg
= strchr(optval
, ':');
637 struct uct_policy
**p
= !strcasecmp(optname
, "policy") ? &u
->policy
: &u
->random_policy
;
640 if (!strcasecmp(optval
, "ucb1")) {
641 *p
= policy_ucb1_init(u
, policyarg
);
642 } else if (!strcasecmp(optval
, "ucb1amaf")) {
643 *p
= policy_ucb1amaf_init(u
, policyarg
);
645 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
648 } else if (!strcasecmp(optname
, "playout") && optval
) {
649 /* Random simulation (playout) policy.
650 * moggy is the default policy with large
651 * amount of domain-specific knowledge and
652 * heuristics. light is a simple uniformly
653 * random move selection policy. */
654 char *playoutarg
= strchr(optval
, ':');
657 if (!strcasecmp(optval
, "moggy")) {
658 u
->playout
= playout_moggy_init(playoutarg
, b
, u
->jdict
);
659 } else if (!strcasecmp(optval
, "light")) {
660 u
->playout
= playout_light_init(playoutarg
, b
);
661 } else if (!strcasecmp(optval
, "elo")) {
662 u
->playout
= playout_elo_init(playoutarg
, b
);
665 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
668 } else if (!strcasecmp(optname
, "prior") && optval
) {
669 /* Node priors policy. When expanding a node,
670 * it will seed node values heuristically
671 * (most importantly, based on playout policy
672 * opinion, but also with regard to other
673 * things). See uct/prior.c for details.
674 * Use prior=eqex=0 to disable priors. */
675 u
->prior
= uct_prior_init(optval
, b
);
676 } else if (!strcasecmp(optname
, "mercy") && optval
) {
677 /* Minimal difference of black/white captures
678 * to stop playout - "Mercy Rule". Speeds up
679 * hopeless playouts at the expense of some
681 u
->mercymin
= atoi(optval
);
682 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
683 /* Maximum length of single simulation
685 u
->gamelen
= atoi(optval
);
686 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
687 /* Expand UCT nodes after it has been
688 * visited this many times. */
689 u
->expand_p
= atoi(optval
);
690 } else if (!strcasecmp(optname
, "random_policy_chance") && optval
) {
691 /* If specified (N), with probability 1/N, random_policy policy
692 * descend is used instead of main policy descend; useful
693 * if specified policy (e.g. UCB1AMAF) can make unduly biased
694 * choices sometimes, you can fall back to e.g.
695 * random_policy=UCB1. */
696 u
->random_policy_chance
= atoi(optval
);
698 /** General AMAF behavior */
699 /* (Only relevant if the policy supports AMAF.
700 * More variables can be tuned as policy
703 } else if (!strcasecmp(optname
, "playout_amaf")) {
704 /* Whether to include random playout moves in
705 * AMAF as well. (Otherwise, only tree moves
706 * are included in AMAF. Of course makes sense
707 * only in connection with an AMAF policy.) */
708 /* with-without: 55.5% (+-4.1) */
709 if (optval
&& *optval
== '0')
710 u
->playout_amaf
= false;
712 u
->playout_amaf
= true;
713 } else if (!strcasecmp(optname
, "playout_amaf_nakade")) {
714 /* Whether to include nakade moves from playouts
715 * in the AMAF statistics; this tends to nullify
716 * the playout_amaf effect by adding too much
718 if (optval
&& *optval
== '0')
719 u
->playout_amaf_nakade
= false;
721 u
->playout_amaf_nakade
= true;
722 } else if (!strcasecmp(optname
, "playout_amaf_cutoff") && optval
) {
723 /* Keep only first N% of playout stage AMAF
725 u
->playout_amaf_cutoff
= atoi(optval
);
726 } else if (!strcasecmp(optname
, "amaf_prior") && optval
) {
727 /* In node policy, consider prior values
728 * part of the real result term or part
729 * of the AMAF term? */
730 u
->amaf_prior
= atoi(optval
);
732 /** Performance and memory management */
734 } else if (!strcasecmp(optname
, "threads") && optval
) {
735 /* By default, Pachi will run with only single
736 * tree search thread! */
737 u
->threads
= atoi(optval
);
738 } else if (!strcasecmp(optname
, "thread_model") && optval
) {
739 if (!strcasecmp(optval
, "tree")) {
740 /* Tree parallelization - all threads
741 * grind on the same tree. */
742 u
->thread_model
= TM_TREE
;
743 u
->virtual_loss
= false;
744 } else if (!strcasecmp(optval
, "treevl")) {
745 /* Tree parallelization, but also
746 * with virtual losses - this discou-
747 * rages most threads choosing the
748 * same tree branches to read. */
749 u
->thread_model
= TM_TREEVL
;
750 u
->virtual_loss
= true;
752 fprintf(stderr
, "UCT: Invalid thread model %s\n", optval
);
755 } else if (!strcasecmp(optname
, "pondering")) {
756 /* Keep searching even during opponent's turn. */
757 u
->pondering_opt
= !optval
|| atoi(optval
);
758 } else if (!strcasecmp(optname
, "max_tree_size") && optval
) {
759 /* Maximum amount of memory [MiB] consumed by the move tree.
760 * For fast_alloc it includes the temp tree used for pruning.
761 * Default is 3072 (3 GiB). */
762 u
->max_tree_size
= atol(optval
) * 1048576;
763 } else if (!strcasecmp(optname
, "fast_alloc")) {
764 u
->fast_alloc
= !optval
|| atoi(optval
);
765 } else if (!strcasecmp(optname
, "pruning_threshold") && optval
) {
766 /* Force pruning at beginning of a move if the tree consumes
767 * more than this [MiB]. Default is 10% of max_tree_size.
768 * Increase to reduce pruning time overhead if memory is plentiful.
769 * This option is meaningful only for fast_alloc. */
770 u
->pruning_threshold
= atol(optval
) * 1048576;
774 } else if (!strcasecmp(optname
, "best2_ratio") && optval
) {
775 /* If set, prolong simulating while
776 * first_best/second_best playouts ratio
777 * is less than best2_ratio. */
778 u
->best2_ratio
= atof(optval
);
779 } else if (!strcasecmp(optname
, "bestr_ratio") && optval
) {
780 /* If set, prolong simulating while
781 * best,best_best_child values delta
782 * is more than bestr_ratio. */
783 u
->bestr_ratio
= atof(optval
);
784 } else if (!strcasecmp(optname
, "fuseki_end") && optval
) {
785 /* At the very beginning it's not worth thinking
786 * too long because the playout evaluations are
787 * very noisy. So gradually increase the thinking
788 * time up to maximum when fuseki_end percent
789 * of the board has been played.
790 * This only applies if we are not in byoyomi. */
791 u
->fuseki_end
= atoi(optval
);
792 } else if (!strcasecmp(optname
, "yose_start") && optval
) {
793 /* When yose_start percent of the board has been
794 * played, or if we are in byoyomi, stop spending
795 * more time and spread the remaining time
797 * Between fuseki_end and yose_start, we spend
798 * a constant proportion of the remaining time
799 * on each move. (yose_start should actually
800 * be much earlier than when real yose start,
801 * but "yose" is a good short name to convey
803 u
->yose_start
= atoi(optval
);
807 } else if (!strcasecmp(optname
, "dynkomi") && optval
) {
808 /* Dynamic komi approach; there are multiple
809 * ways to adjust komi dynamically throughout
810 * play. We currently support two: */
811 char *dynkomiarg
= strchr(optval
, ':');
814 if (!strcasecmp(optval
, "none")) {
815 u
->dynkomi
= uct_dynkomi_init_none(u
, dynkomiarg
, b
);
816 } else if (!strcasecmp(optval
, "linear")) {
817 /* You should set dynkomi_mask=1
818 * since this doesn't work well
819 * for white handicaps! */
820 u
->dynkomi
= uct_dynkomi_init_linear(u
, dynkomiarg
, b
);
821 } else if (!strcasecmp(optval
, "adaptive")) {
822 /* There are many more knobs to
823 * crank - see uct/dynkomi.c. */
824 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, dynkomiarg
, b
);
826 fprintf(stderr
, "UCT: Invalid dynkomi mode %s\n", optval
);
829 } else if (!strcasecmp(optname
, "dynkomi_mask") && optval
) {
830 /* Bitmask of colors the player must be
831 * for dynkomi be applied; you may want
832 * to use dynkomi_mask=3 to allow dynkomi
833 * even in games where Pachi is white. */
834 u
->dynkomi_mask
= atoi(optval
);
835 } else if (!strcasecmp(optname
, "dynkomi_interval") && optval
) {
836 /* If non-zero, re-adjust dynamic komi
837 * throughout a single genmove reading,
838 * roughly every N simulations. */
839 /* XXX: Does not work with tree
840 * parallelization. */
841 u
->dynkomi_interval
= atoi(optval
);
843 /** Node value result scaling */
845 } else if (!strcasecmp(optname
, "val_scale") && optval
) {
846 /* How much of the game result value should be
847 * influenced by win size. Zero means it isn't. */
848 u
->val_scale
= atof(optval
);
849 } else if (!strcasecmp(optname
, "val_points") && optval
) {
850 /* Maximum size of win to be scaled into game
851 * result value. Zero means boardsize^2. */
852 u
->val_points
= atoi(optval
) * 2; // result values are doubled
853 } else if (!strcasecmp(optname
, "val_extra")) {
854 /* If false, the score coefficient will be simply
855 * added to the value, instead of scaling the result
856 * coefficient because of it. */
857 u
->val_extra
= !optval
|| atoi(optval
);
860 /* (Purely experimental. Does not work - yet!) */
862 } else if (!strcasecmp(optname
, "local_tree") && optval
) {
863 /* Whether to bias exploration by local tree values
864 * (must be supported by the used policy).
866 * 1: Do, value = result.
867 * Try to temper the result:
868 * 2: Do, value = 0.5+(result-expected)/2.
869 * 3: Do, value = 0.5+bzz((result-expected)^2).
870 * 4: Do, value = 0.5+sqrt(result-expected)/2. */
871 u
->local_tree
= atoi(optval
);
872 } else if (!strcasecmp(optname
, "tenuki_d") && optval
) {
873 /* Tenuki distance at which to break the local tree. */
874 u
->tenuki_d
= atoi(optval
);
875 if (u
->tenuki_d
> TREE_NODE_D_MAX
+ 1) {
876 fprintf(stderr
, "uct: tenuki_d must not be larger than TREE_NODE_D_MAX+1 %d\n", TREE_NODE_D_MAX
+ 1);
879 } else if (!strcasecmp(optname
, "local_tree_aging") && optval
) {
880 /* How much to reduce local tree values between moves. */
881 u
->local_tree_aging
= atof(optval
);
882 } else if (!strcasecmp(optname
, "local_tree_allseq")) {
883 /* By default, only complete sequences are stored
884 * in the local tree. If this is on, also
885 * subsequences starting at each move are stored. */
886 u
->local_tree_allseq
= !optval
|| atoi(optval
);
887 } else if (!strcasecmp(optname
, "local_tree_playout")) {
888 /* Whether to adjust ELO playout probability
889 * distributions according to matched localtree
891 u
->local_tree_playout
= !optval
|| atoi(optval
);
892 } else if (!strcasecmp(optname
, "local_tree_pseqroot")) {
893 /* By default, when we have no sequence move
894 * to suggest in-playout, we give up. If this
895 * is on, we make probability distribution from
896 * sequences first moves instead. */
897 u
->local_tree_pseqroot
= !optval
|| atoi(optval
);
899 /** Other heuristics */
900 } else if (!strcasecmp(optname
, "significant_threshold") && optval
) {
901 /* Some heuristics (treepool) rely
902 * on the knowledge of the last "significant"
903 * node in the descent. Such a node is
904 * considered reasonably trustworthy to carry
905 * some meaningful information in the values
906 * of the node and its children. */
907 u
->significant_threshold
= atoi(optval
);
908 } else if (!strcasecmp(optname
, "treepool_chance") && optval
) {
909 /* Chance of applying the treepool heuristic:
910 * one of the best N children of the last
911 * significant node is tried on each turn
912 * of the simulation. */
913 /* This is in form of two numbers:
914 * PREMOVE:POSTMOVE. Each is percentage
915 * value, one is the chance of the move
916 * tried before playout policy, one is the
917 * chance of it being applied if the policy
918 * has not picked anymove. */
919 char *optval2
= strchr(optval
, ':');
921 fprintf(stderr
, "uct: treepool_chance takes two comma-separated numbers\n");
924 u
->treepool_chance
[0] = atoi(optval
);
926 u
->treepool_chance
[1] = atoi(optval
);
927 } else if (!strcasecmp(optname
, "treepool_size") && optval
) {
928 /* Number of top significant children
929 * to pick from. Too low means low effect,
930 * too high means even lousy moves are
932 u
->treepool_size
= atoi(optval
);
933 } else if (!strcasecmp(optname
, "treepool_type") && optval
) {
934 /* How to sort the children. */
935 if (!strcasecmp(optval
, "rave_playouts"))
936 u
->treepool_type
= UTT_RAVE_PLAYOUTS
;
937 else if (!strcasecmp(optval
, "rave_value"))
938 u
->treepool_type
= UTT_RAVE_VALUE
;
939 else if (!strcasecmp(optval
, "uct_playouts"))
940 u
->treepool_type
= UTT_UCT_PLAYOUTS
;
941 else if (!strcasecmp(optval
, "uct_value"))
942 u
->treepool_type
= UTT_UCT_VALUE
;
943 else if (!strcasecmp(optval
, "evaluate"))
944 /* The proper combination of RAVE,
945 * UCT and prior, as used through
947 u
->treepool_type
= UTT_EVALUATE
;
949 fprintf(stderr
, "uct: unknown treepool_type %s\n", optval
);
952 } else if (!strcasecmp(optname
, "treepool_pickfactor") && optval
) {
953 /* Pick factor influencing children choice.
954 * By default (if this is 0 or 10), coords
955 * have uniform probability to be chosen;
956 * otherwise, children are tried from best
957 * to worst, each picked with probability
958 * (1/n * pickfactor/10).
959 * I.e., better children may be preferred
960 * if pickfactor > 10. */
961 u
->treepool_pickfactor
= atoi(optval
);
963 /** Distributed engine slaves setup */
965 } else if (!strcasecmp(optname
, "slave")) {
966 /* Act as slave for the distributed engine. */
967 u
->slave
= !optval
|| atoi(optval
);
968 } else if (!strcasecmp(optname
, "shared_nodes") && optval
) {
969 /* Share at most shared_nodes between master and slave at each genmoves.
970 * Must use the same value in master and slaves. */
971 u
->shared_nodes
= atoi(optval
);
972 } else if (!strcasecmp(optname
, "shared_levels") && optval
) {
973 /* Share only nodes of level <= shared_levels. */
974 u
->shared_levels
= atoi(optval
);
975 } else if (!strcasecmp(optname
, "stats_hbits") && optval
) {
976 /* Set hash table size to 2^stats_hbits for the shared stats. */
977 u
->stats_hbits
= atoi(optval
);
980 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
987 u
->policy
= policy_ucb1amaf_init(u
, NULL
);
989 if (!!u
->random_policy_chance
^ !!u
->random_policy
) {
990 fprintf(stderr
, "uct: Only one of random_policy and random_policy_chance is set\n");
994 if (!u
->local_tree
) {
995 /* No ltree aging. */
996 u
->local_tree_aging
= 1.0f
;
999 u
->local_tree_playout
= false;
1001 if (u
->fast_alloc
) {
1002 if (u
->pruning_threshold
< u
->max_tree_size
/ 10)
1003 u
->pruning_threshold
= u
->max_tree_size
/ 10;
1004 if (u
->pruning_threshold
> u
->max_tree_size
/ 2)
1005 u
->pruning_threshold
= u
->max_tree_size
/ 2;
1007 /* Limit pruning temp space to 20% of memory. Beyond this we discard
1008 * the nodes and recompute them at the next move if necessary. */
1009 u
->max_pruned_size
= u
->max_tree_size
/ 5;
1010 u
->max_tree_size
-= u
->max_pruned_size
;
1012 /* Reserve 5% memory in case the background free() are slower
1013 * than the concurrent allocations. */
1014 u
->max_tree_size
-= u
->max_tree_size
/ 20;
1018 u
->prior
= uct_prior_init(NULL
, b
);
1021 u
->playout
= playout_moggy_init(NULL
, b
, u
->jdict
);
1022 if (!u
->playout
->debug_level
)
1023 u
->playout
->debug_level
= u
->debug_level
;
1025 u
->ownermap
.map
= malloc2(board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
1028 if (!u
->stats_hbits
) u
->stats_hbits
= DEFAULT_STATS_HBITS
;
1029 if (!u
->shared_nodes
) u
->shared_nodes
= DEFAULT_SHARED_NODES
;
1030 assert(u
->shared_levels
* board_bits2(b
) <= 8 * (int)sizeof(path_t
));
1034 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, NULL
, b
);
1036 /* Some things remain uninitialized for now - the opening tbook
1037 * is not loaded and the tree not set up. */
1038 /* This will be initialized in setup_state() at the first move
1039 * received/requested. This is because right now we are not aware
1040 * about any komi or handicap setup and such. */
1046 engine_uct_init(char *arg
, struct board
*b
)
1048 struct uct
*u
= uct_state_init(arg
, b
);
1049 struct engine
*e
= calloc2(1, sizeof(struct engine
));
1050 e
->name
= "UCT Engine";
1051 e
->printhook
= uct_printhook_ownermap
;
1052 e
->notify_play
= uct_notify_play
;
1054 e
->result
= uct_result
;
1055 e
->genmove
= uct_genmove
;
1056 e
->genmoves
= uct_genmoves
;
1057 e
->dead_group_list
= uct_dead_group_list
;
1061 e
->notify
= uct_notify
;
1063 const char banner
[] = "I'm playing UCT. When I'm losing, I will resign, "
1064 "if I think I win, I play until you pass. "
1065 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
1066 if (!u
->banner
) u
->banner
= "";
1067 e
->comment
= malloc2(sizeof(banner
) + strlen(u
->banner
) + 1);
1068 sprintf(e
->comment
, "%s %s", banner
, u
->banner
);