15 #include "joseki/base.h"
17 #include "playout/moggy.h"
18 #include "playout/light.h"
19 #include "tactics/util.h"
21 #include "uct/dynkomi.h"
22 #include "uct/internal.h"
23 #include "uct/plugins.h"
24 #include "uct/prior.h"
25 #include "uct/search.h"
26 #include "uct/slave.h"
31 struct uct_policy
*policy_ucb1_init(struct uct
*u
, char *arg
);
32 struct uct_policy
*policy_ucb1amaf_init(struct uct
*u
, char *arg
);
33 static void uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
);
35 /* Maximal simulation length. */
36 #define MC_GAMELEN MAX_GAMELEN
40 setup_state(struct uct
*u
, struct board
*b
, enum stone color
)
42 u
->t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0,
43 u
->max_pruned_size
, u
->pruning_threshold
, u
->local_tree_aging
, u
->stats_hbits
);
45 fast_srandom(u
->force_seed
);
47 fprintf(stderr
, "Fresh board with random seed %lu\n", fast_getseed());
48 //board_print(b, stderr);
49 if (!u
->no_tbook
&& b
->moves
== 0) {
50 assert(color
== S_BLACK
);
56 reset_state(struct uct
*u
)
59 tree_done(u
->t
); u
->t
= NULL
;
63 setup_dynkomi(struct uct
*u
, struct board
*b
, enum stone to_play
)
65 if (u
->t
->use_extra_komi
&& !u
->pondering
&& u
->dynkomi
->permove
)
66 u
->t
->extra_komi
= u
->dynkomi
->permove(u
->dynkomi
, b
, u
->t
);
67 else if (!u
->t
->use_extra_komi
)
72 uct_prepare_move(struct uct
*u
, struct board
*b
, enum stone color
)
75 /* Verify that we have sane state. */
77 assert(u
->t
&& b
->moves
);
78 if (color
!= stone_other(u
->t
->root_color
)) {
79 fprintf(stderr
, "Fatal: Non-alternating play detected %d %d\n",
80 color
, u
->t
->root_color
);
83 uct_htable_reset(u
->t
);
86 /* We need fresh state. */
88 setup_state(u
, b
, color
);
91 u
->ownermap
.playouts
= 0;
92 memset(u
->ownermap
.map
, 0, board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
93 u
->played_own
= u
->played_all
= 0;
97 dead_group_list(struct uct
*u
, struct board
*b
, struct move_queue
*mq
)
99 struct group_judgement gj
;
101 gj
.gs
= alloca(board_size2(b
) * sizeof(gj
.gs
[0]));
102 board_ownermap_judge_group(b
, &u
->ownermap
, &gj
);
103 groups_of_status(b
, &gj
, GS_DEAD
, mq
);
107 uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
)
109 if (u
->ownermap
.playouts
< GJ_MINGAMES
)
112 struct move_queue mq
= { .moves
= 0 };
113 dead_group_list(u
, b
, &mq
);
114 if (pass_all_alive
&& mq
.moves
> 0)
115 return false; // We need to remove some dead groups first.
116 return pass_is_safe(b
, color
, &mq
);
120 uct_printhook_ownermap(struct board
*board
, coord_t c
, char *s
, char *end
)
122 struct uct
*u
= board
->es
;
127 const char chr
[] = ":XO,"; // dame, black, white, unclear
128 const char chm
[] = ":xo,";
129 char ch
= chr
[board_ownermap_judge_point(&u
->ownermap
, c
, GJ_THRES
)];
130 if (ch
== ',') { // less precise estimate then?
131 ch
= chm
[board_ownermap_judge_point(&u
->ownermap
, c
, 0.67)];
133 s
+= snprintf(s
, end
- s
, "%c ", ch
);
138 uct_notify_play(struct engine
*e
, struct board
*b
, struct move
*m
)
140 struct uct
*u
= e
->data
;
142 /* No state, create one - this is probably game beginning
143 * and we need to load the opening tbook right now. */
144 uct_prepare_move(u
, b
, m
->color
);
148 /* Stop pondering, required by tree_promote_at() */
149 uct_pondering_stop(u
);
150 if (UDEBUGL(2) && u
->slave
)
151 tree_dump(u
->t
, u
->dumpthres
);
153 if (is_resign(m
->coord
)) {
159 /* Promote node of the appropriate move to the tree root. */
161 if (!tree_promote_at(u
->t
, b
, m
->coord
)) {
163 fprintf(stderr
, "Warning: Cannot promote move node! Several play commands in row?\n");
168 /* If we are a slave in a distributed engine, start pondering once
169 * we know which move we actually played. See uct_genmove() about
170 * the check for pass. */
171 if (u
->pondering_opt
&& u
->slave
&& m
->color
== u
->my_color
&& !is_pass(m
->coord
))
172 uct_pondering_start(u
, b
, u
->t
, stone_other(m
->color
));
178 uct_undo(struct engine
*e
, struct board
*b
)
180 struct uct
*u
= e
->data
;
182 if (!u
->t
) return NULL
;
183 uct_pondering_stop(u
);
189 uct_result(struct engine
*e
, struct board
*b
)
191 struct uct
*u
= e
->data
;
192 static char reply
[1024];
196 enum stone color
= u
->t
->root_color
;
197 struct tree_node
*n
= u
->t
->root
;
198 snprintf(reply
, 1024, "%s %s %d %.2f %.1f",
199 stone2str(color
), coord2sstr(n
->coord
, b
),
200 n
->u
.playouts
, tree_node_get_value(u
->t
, -1, n
->u
.value
),
201 u
->t
->use_extra_komi
? u
->t
->extra_komi
: 0);
206 uct_chat(struct engine
*e
, struct board
*b
, char *cmd
)
208 struct uct
*u
= e
->data
;
209 static char reply
[1024];
211 cmd
+= strspn(cmd
, " \n\t");
212 if (!strncasecmp(cmd
, "winrate", 7)) {
214 return "no game context (yet?)";
215 enum stone color
= u
->t
->root_color
;
216 struct tree_node
*n
= u
->t
->root
;
217 snprintf(reply
, 1024, "In %d playouts at %d threads, %s %s can win with %.2f%% probability",
218 n
->u
.playouts
, u
->threads
, stone2str(color
), coord2sstr(n
->coord
, b
),
219 tree_node_get_value(u
->t
, -1, n
->u
.value
) * 100);
220 if (u
->t
->use_extra_komi
&& abs(u
->t
->extra_komi
) >= 0.5) {
221 sprintf(reply
+ strlen(reply
), ", while self-imposing extra komi %.1f",
231 uct_dead_group_list(struct engine
*e
, struct board
*b
, struct move_queue
*mq
)
233 struct uct
*u
= e
->data
;
235 /* This means the game is probably over, no use pondering on. */
236 uct_pondering_stop(u
);
238 if (u
->pass_all_alive
)
239 return; // no dead groups
241 bool mock_state
= false;
244 /* No state, but we cannot just back out - we might
245 * have passed earlier, only assuming some stones are
246 * dead, and then re-connected, only to lose counting
247 * when all stones are assumed alive. */
248 uct_prepare_move(u
, b
, S_BLACK
); assert(u
->t
);
251 /* Make sure the ownermap is well-seeded. */
252 while (u
->ownermap
.playouts
< GJ_MINGAMES
)
253 uct_playout(u
, b
, S_BLACK
, u
->t
);
254 /* Show the ownermap: */
256 board_print_custom(b
, stderr
, uct_printhook_ownermap
);
258 dead_group_list(u
, b
, mq
);
261 /* Clean up the mock state in case we will receive
262 * a genmove; we could get a non-alternating-move
263 * error from uct_prepare_move() in that case otherwise. */
269 playout_policy_done(struct playout_policy
*p
)
271 if (p
->done
) p
->done(p
);
272 if (p
->data
) free(p
->data
);
277 uct_done(struct engine
*e
)
279 /* This is called on engine reset, especially when clear_board
280 * is received and new game should begin. */
281 struct uct
*u
= e
->data
;
282 uct_pondering_stop(u
);
283 if (u
->t
) reset_state(u
);
284 free(u
->ownermap
.map
);
287 free(u
->random_policy
);
288 playout_policy_done(u
->playout
);
289 uct_prior_done(u
->prior
);
290 joseki_done(u
->jdict
);
291 pluginset_done(u
->plugins
);
296 /* Run time-limited MCTS search on foreground. */
298 uct_search(struct uct
*u
, struct board
*b
, struct time_info
*ti
, enum stone color
, struct tree
*t
)
300 struct uct_search_state s
;
301 uct_search_start(u
, b
, color
, t
, ti
, &s
);
302 if (UDEBUGL(2) && s
.base_playouts
> 0)
303 fprintf(stderr
, "<pre-simulated %d games>\n", s
.base_playouts
);
305 /* The search tree is ctx->t. This is currently == . It is important
306 * to reference ctx->t directly since the
307 * thread manager will swap the tree pointer asynchronously. */
309 /* Now, just periodically poll the search tree. */
310 /* Note that in case of TD_GAMES, threads will terminate independently
311 * of the uct_search_check_stop() signalization. */
313 time_sleep(TREE_BUSYWAIT_INTERVAL
);
314 /* TREE_BUSYWAIT_INTERVAL should never be less than desired time, or the
315 * time control is broken. But if it happens to be less, we still search
316 * at least 100ms otherwise the move is completely random. */
318 int i
= uct_search_games(&s
);
319 /* Print notifications etc. */
320 uct_search_progress(u
, b
, color
, t
, ti
, &s
, i
);
321 /* Check if we should stop the search. */
322 if (uct_search_check_stop(u
, b
, color
, t
, ti
, &s
, i
))
326 struct uct_thread_ctx
*ctx
= uct_search_stop();
327 if (UDEBUGL(2)) tree_dump(t
, u
->dumpthres
);
329 fprintf(stderr
, "(avg score %f/%d value %f/%d)\n",
330 u
->dynkomi
->score
.value
, u
->dynkomi
->score
.playouts
,
331 u
->dynkomi
->value
.value
, u
->dynkomi
->value
.playouts
);
333 uct_progress_status(u
, t
, color
, ctx
->games
);
335 u
->played_own
+= ctx
->games
;
339 /* Start pondering background with @color to play. */
341 uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
)
344 fprintf(stderr
, "Starting to ponder with color %s\n", stone2str(stone_other(color
)));
347 /* We need a local board copy to ponder upon. */
348 struct board
*b
= malloc2(sizeof(*b
)); board_copy(b
, b0
);
350 /* *b0 did not have the genmove'd move played yet. */
351 struct move m
= { t
->root
->coord
, t
->root_color
};
352 int res
= board_play(b
, &m
);
354 setup_dynkomi(u
, b
, stone_other(m
.color
));
356 /* Start MCTS manager thread "headless". */
357 static struct uct_search_state s
;
358 uct_search_start(u
, b
, color
, t
, NULL
, &s
);
361 /* uct_search_stop() frontend for the pondering (non-genmove) mode, and
362 * to stop the background search for a slave in the distributed engine. */
364 uct_pondering_stop(struct uct
*u
)
366 if (!thread_manager_running
)
369 /* Stop the thread manager. */
370 struct uct_thread_ctx
*ctx
= uct_search_stop();
372 if (u
->pondering
) fprintf(stderr
, "(pondering) ");
373 uct_progress_status(u
, ctx
->t
, ctx
->color
, ctx
->games
);
377 u
->pondering
= false;
383 uct_genmove_setup(struct uct
*u
, struct board
*b
, enum stone color
)
385 if (b
->superko_violation
) {
386 fprintf(stderr
, "!!! WARNING: SUPERKO VIOLATION OCCURED BEFORE THIS MOVE\n");
387 fprintf(stderr
, "Maybe you play with situational instead of positional superko?\n");
388 fprintf(stderr
, "I'm going to ignore the violation, but note that I may miss\n");
389 fprintf(stderr
, "some moves valid under this ruleset because of this.\n");
390 b
->superko_violation
= false;
393 uct_prepare_move(u
, b
, color
);
398 /* How to decide whether to use dynkomi in this game? Since we use
399 * pondering, it's not simple "who-to-play" matter. Decide based on
400 * the last genmove issued. */
401 u
->t
->use_extra_komi
= !!(u
->dynkomi_mask
& color
);
402 /* Moreover, we do not use extra komi at the game end - we are not
403 * to fool ourselves at this point. */
404 if (board_estimated_moves_left(b
) <= MIN_MOVES_LEFT
)
405 u
->t
->use_extra_komi
= false;
406 setup_dynkomi(u
, b
, color
);
408 if (b
->rules
== RULES_JAPANESE
)
409 u
->territory_scoring
= true;
411 /* Make pessimistic assumption about komi for Japanese rules to
412 * avoid losing by 0.5 when winning by 0.5 with Chinese rules.
413 * The rules usually give the same winner if the integer part of komi
414 * is odd so we adjust the komi only if it is even (for a board of
415 * odd size). We are not trying to get an exact evaluation for rare
416 * cases of seki. For details see http://home.snafu.de/jasiek/parity.html */
417 if (u
->territory_scoring
&& (((int)floor(b
->komi
) + board_size(b
)) & 1)) {
418 b
->komi
+= (color
== S_BLACK
? 1.0 : -1.0);
420 fprintf(stderr
, "Setting komi to %.1f assuming Japanese rules\n",
426 uct_genmove(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
428 double start_time
= time_now();
429 struct uct
*u
= e
->data
;
430 uct_pondering_stop(u
);
431 uct_genmove_setup(u
, b
, color
);
433 /* Start the Monte Carlo Tree Search! */
434 int base_playouts
= u
->t
->root
->u
.playouts
;
435 int played_games
= uct_search(u
, b
, ti
, color
, u
->t
);
438 struct tree_node
*best
;
439 best
= uct_search_result(u
, b
, color
, pass_all_alive
, played_games
, base_playouts
, &best_coord
);
442 double time
= time_now() - start_time
+ 0.000001; /* avoid divide by zero */
443 fprintf(stderr
, "genmove in %0.2fs (%d games/s, %d games/s/thread)\n",
444 time
, (int)(played_games
/time
), (int)(played_games
/time
/u
->threads
));
448 /* Pass or resign. */
450 return coord_copy(best_coord
);
452 tree_promote_node(u
->t
, &best
);
454 /* After a pass, pondering is harmful for two reasons:
455 * (i) We might keep pondering even when the game is over.
456 * Of course this is the case for opponent resign as well.
457 * (ii) More importantly, the ownermap will get skewed since
458 * the UCT will start cutting off any playouts. */
459 if (u
->pondering_opt
&& !is_pass(best
->coord
)) {
460 uct_pondering_start(u
, b
, u
->t
, stone_other(color
));
462 return coord_copy(best_coord
);
467 uct_gentbook(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
)
469 struct uct
*u
= e
->data
;
470 if (!u
->t
) uct_prepare_move(u
, b
, color
);
473 if (ti
->dim
== TD_GAMES
) {
474 /* Don't count in games that already went into the tbook. */
475 ti
->len
.games
+= u
->t
->root
->u
.playouts
;
477 uct_search(u
, b
, ti
, color
, u
->t
);
479 assert(ti
->dim
== TD_GAMES
);
480 tree_save(u
->t
, b
, ti
->len
.games
/ 100);
486 uct_dumptbook(struct engine
*e
, struct board
*b
, enum stone color
)
488 struct uct
*u
= e
->data
;
489 struct tree
*t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0,
490 u
->max_pruned_size
, u
->pruning_threshold
, u
->local_tree_aging
, 0);
498 uct_evaluate(struct engine
*e
, struct board
*b
, struct time_info
*ti
, coord_t c
, enum stone color
)
500 struct uct
*u
= e
->data
;
504 struct move m
= { c
, color
};
505 int res
= board_play(&b2
, &m
);
508 color
= stone_other(color
);
510 if (u
->t
) reset_state(u
);
511 uct_prepare_move(u
, &b2
, color
);
515 uct_search(u
, &b2
, ti
, color
, u
->t
);
516 struct tree_node
*best
= u
->policy
->choose(u
->policy
, u
->t
->root
, &b2
, color
, resign
);
518 bestval
= NAN
; // the opponent has no reply!
520 bestval
= tree_node_get_value(u
->t
, 1, best
->u
.value
);
523 reset_state(u
); // clean our junk
525 return isnan(bestval
) ? NAN
: 1.0f
- bestval
;
530 uct_state_init(char *arg
, struct board
*b
)
532 struct uct
*u
= calloc2(1, sizeof(struct uct
));
534 u
->debug_level
= debug_level
;
535 u
->gamelen
= MC_GAMELEN
;
536 u
->resign_threshold
= 0.2;
537 u
->sure_win_threshold
= 0.85;
539 u
->significant_threshold
= 50;
542 u
->playout_amaf
= true;
543 u
->playout_amaf_nakade
= false;
544 u
->amaf_prior
= false;
545 u
->max_tree_size
= 1408ULL * 1048576;
546 u
->fast_alloc
= true;
547 u
->pruning_threshold
= 0;
550 u
->thread_model
= TM_TREEVL
;
553 u
->fuseki_end
= 20; // max time at 361*20% = 72 moves (our 36th move, still 99 to play)
554 u
->yose_start
= 40; // (100-40-25)*361/100/2 = 63 moves still to play by us then
555 u
->bestr_ratio
= 0.02;
556 // 2.5 is clearly too much, but seems to compensate well for overly stern time allocations.
557 // TODO: Further tuning and experiments with better time allocation schemes.
558 u
->best2_ratio
= 2.5;
559 u
->max_maintime_ratio
= 8.0;
561 u
->val_scale
= 0.04; u
->val_points
= 40;
562 u
->dynkomi_interval
= 1000;
563 u
->dynkomi_mask
= S_BLACK
| S_WHITE
;
566 u
->local_tree_aging
= 80;
567 u
->local_tree_allseq
= 1;
568 u
->local_tree_rootseqval
= 1;
569 u
->local_tree_depth_decay
= 1.5;
571 u
->stats_delay
= 0.001; // 1 ms
573 u
->plugins
= pluginset_init(b
);
575 u
->jdict
= joseki_load(b
->size
);
578 char *optspec
, *next
= arg
;
581 next
+= strcspn(next
, ",");
582 if (*next
) { *next
++ = 0; } else { *next
= 0; }
584 char *optname
= optspec
;
585 char *optval
= strchr(optspec
, '=');
586 if (optval
) *optval
++ = 0;
590 if (!strcasecmp(optname
, "debug")) {
592 u
->debug_level
= atoi(optval
);
595 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
596 /* When dumping the UCT tree on output, include
597 * nodes with at least this many playouts.
598 * (This value is re-scaled "intelligently"
599 * in case of very large trees.) */
600 u
->dumpthres
= atoi(optval
);
601 } else if (!strcasecmp(optname
, "resign_threshold") && optval
) {
602 /* Resign when this ratio of games is lost
603 * after GJ_MINGAMES sample is taken. */
604 u
->resign_threshold
= atof(optval
);
605 } else if (!strcasecmp(optname
, "sure_win_threshold") && optval
) {
606 /* Stop reading when this ratio of games is won
607 * after PLAYOUT_EARLY_BREAK_MIN sample is
608 * taken. (Prevents stupid time losses,
609 * friendly to human opponents.) */
610 u
->sure_win_threshold
= atof(optval
);
611 } else if (!strcasecmp(optname
, "force_seed") && optval
) {
612 /* Set RNG seed at the tree setup. */
613 u
->force_seed
= atoi(optval
);
614 } else if (!strcasecmp(optname
, "no_tbook")) {
615 /* Disable UCT opening tbook. */
617 } else if (!strcasecmp(optname
, "pass_all_alive")) {
618 /* Whether to consider passing only after all
619 * dead groups were removed from the board;
620 * this is like all genmoves are in fact
621 * kgs-genmove_cleanup. */
622 u
->pass_all_alive
= !optval
|| atoi(optval
);
623 } else if (!strcasecmp(optname
, "territory_scoring")) {
624 /* Use territory scoring (default is area scoring).
625 * An explicit kgs-rules command overrides this. */
626 u
->territory_scoring
= !optval
|| atoi(optval
);
627 } else if (!strcasecmp(optname
, "banner") && optval
) {
628 /* Additional banner string. This must come as the
629 * last engine parameter. */
630 if (*next
) *--next
= ',';
631 u
->banner
= strdup(optval
);
633 } else if (!strcasecmp(optname
, "plugin") && optval
) {
634 /* Load an external plugin; filename goes before the colon,
635 * extra arguments after the colon. */
636 char *pluginarg
= strchr(optval
, ':');
639 plugin_load(u
->plugins
, optval
, pluginarg
);
641 /** UCT behavior and policies */
643 } else if ((!strcasecmp(optname
, "policy")
644 /* Node selection policy. ucb1amaf is the
645 * default policy implementing RAVE, while
646 * ucb1 is the simple exploration/exploitation
647 * policy. Policies can take further extra
649 || !strcasecmp(optname
, "random_policy")) && optval
) {
650 /* A policy to be used randomly with small
651 * chance instead of the default policy. */
652 char *policyarg
= strchr(optval
, ':');
653 struct uct_policy
**p
= !strcasecmp(optname
, "policy") ? &u
->policy
: &u
->random_policy
;
656 if (!strcasecmp(optval
, "ucb1")) {
657 *p
= policy_ucb1_init(u
, policyarg
);
658 } else if (!strcasecmp(optval
, "ucb1amaf")) {
659 *p
= policy_ucb1amaf_init(u
, policyarg
);
661 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
664 } else if (!strcasecmp(optname
, "playout") && optval
) {
665 /* Random simulation (playout) policy.
666 * moggy is the default policy with large
667 * amount of domain-specific knowledge and
668 * heuristics. light is a simple uniformly
669 * random move selection policy. */
670 char *playoutarg
= strchr(optval
, ':');
673 if (!strcasecmp(optval
, "moggy")) {
674 u
->playout
= playout_moggy_init(playoutarg
, b
, u
->jdict
);
675 } else if (!strcasecmp(optval
, "light")) {
676 u
->playout
= playout_light_init(playoutarg
, b
);
678 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
681 } else if (!strcasecmp(optname
, "prior") && optval
) {
682 /* Node priors policy. When expanding a node,
683 * it will seed node values heuristically
684 * (most importantly, based on playout policy
685 * opinion, but also with regard to other
686 * things). See uct/prior.c for details.
687 * Use prior=eqex=0 to disable priors. */
688 u
->prior
= uct_prior_init(optval
, b
);
689 } else if (!strcasecmp(optname
, "mercy") && optval
) {
690 /* Minimal difference of black/white captures
691 * to stop playout - "Mercy Rule". Speeds up
692 * hopeless playouts at the expense of some
694 u
->mercymin
= atoi(optval
);
695 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
696 /* Maximum length of single simulation
698 u
->gamelen
= atoi(optval
);
699 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
700 /* Expand UCT nodes after it has been
701 * visited this many times. */
702 u
->expand_p
= atoi(optval
);
703 } else if (!strcasecmp(optname
, "random_policy_chance") && optval
) {
704 /* If specified (N), with probability 1/N, random_policy policy
705 * descend is used instead of main policy descend; useful
706 * if specified policy (e.g. UCB1AMAF) can make unduly biased
707 * choices sometimes, you can fall back to e.g.
708 * random_policy=UCB1. */
709 u
->random_policy_chance
= atoi(optval
);
711 /** General AMAF behavior */
712 /* (Only relevant if the policy supports AMAF.
713 * More variables can be tuned as policy
716 } else if (!strcasecmp(optname
, "playout_amaf")) {
717 /* Whether to include random playout moves in
718 * AMAF as well. (Otherwise, only tree moves
719 * are included in AMAF. Of course makes sense
720 * only in connection with an AMAF policy.) */
721 /* with-without: 55.5% (+-4.1) */
722 if (optval
&& *optval
== '0')
723 u
->playout_amaf
= false;
725 u
->playout_amaf
= true;
726 } else if (!strcasecmp(optname
, "playout_amaf_nakade")) {
727 /* Whether to include nakade moves from playouts
728 * in the AMAF statistics; this tends to nullify
729 * the playout_amaf effect by adding too much
731 if (optval
&& *optval
== '0')
732 u
->playout_amaf_nakade
= false;
734 u
->playout_amaf_nakade
= true;
735 } else if (!strcasecmp(optname
, "playout_amaf_cutoff") && optval
) {
736 /* Keep only first N% of playout stage AMAF
738 u
->playout_amaf_cutoff
= atoi(optval
);
739 } else if (!strcasecmp(optname
, "amaf_prior") && optval
) {
740 /* In node policy, consider prior values
741 * part of the real result term or part
742 * of the AMAF term? */
743 u
->amaf_prior
= atoi(optval
);
745 /** Performance and memory management */
747 } else if (!strcasecmp(optname
, "threads") && optval
) {
748 /* By default, Pachi will run with only single
749 * tree search thread! */
750 u
->threads
= atoi(optval
);
751 } else if (!strcasecmp(optname
, "thread_model") && optval
) {
752 if (!strcasecmp(optval
, "tree")) {
753 /* Tree parallelization - all threads
754 * grind on the same tree. */
755 u
->thread_model
= TM_TREE
;
757 } else if (!strcasecmp(optval
, "treevl")) {
758 /* Tree parallelization, but also
759 * with virtual losses - this discou-
760 * rages most threads choosing the
761 * same tree branches to read. */
762 u
->thread_model
= TM_TREEVL
;
764 fprintf(stderr
, "UCT: Invalid thread model %s\n", optval
);
767 } else if (!strcasecmp(optname
, "virtual_loss")) {
768 /* Number of virtual losses added before evaluating a node. */
769 u
->virtual_loss
= !optval
|| atoi(optval
);
770 } else if (!strcasecmp(optname
, "pondering")) {
771 /* Keep searching even during opponent's turn. */
772 u
->pondering_opt
= !optval
|| atoi(optval
);
773 } else if (!strcasecmp(optname
, "max_tree_size") && optval
) {
774 /* Maximum amount of memory [MiB] consumed by the move tree.
775 * For fast_alloc it includes the temp tree used for pruning.
776 * Default is 3072 (3 GiB). */
777 u
->max_tree_size
= atol(optval
) * 1048576;
778 } else if (!strcasecmp(optname
, "fast_alloc")) {
779 u
->fast_alloc
= !optval
|| atoi(optval
);
780 } else if (!strcasecmp(optname
, "pruning_threshold") && optval
) {
781 /* Force pruning at beginning of a move if the tree consumes
782 * more than this [MiB]. Default is 10% of max_tree_size.
783 * Increase to reduce pruning time overhead if memory is plentiful.
784 * This option is meaningful only for fast_alloc. */
785 u
->pruning_threshold
= atol(optval
) * 1048576;
789 } else if (!strcasecmp(optname
, "best2_ratio") && optval
) {
790 /* If set, prolong simulating while
791 * first_best/second_best playouts ratio
792 * is less than best2_ratio. */
793 u
->best2_ratio
= atof(optval
);
794 } else if (!strcasecmp(optname
, "bestr_ratio") && optval
) {
795 /* If set, prolong simulating while
796 * best,best_best_child values delta
797 * is more than bestr_ratio. */
798 u
->bestr_ratio
= atof(optval
);
799 } else if (!strcasecmp(optname
, "max_maintime_ratio") && optval
) {
800 /* If set and while not in byoyomi, prolong simulating no more than
801 * max_maintime_ratio times the normal desired thinking time. */
802 u
->max_maintime_ratio
= atof(optval
);
803 } else if (!strcasecmp(optname
, "fuseki_end") && optval
) {
804 /* At the very beginning it's not worth thinking
805 * too long because the playout evaluations are
806 * very noisy. So gradually increase the thinking
807 * time up to maximum when fuseki_end percent
808 * of the board has been played.
809 * This only applies if we are not in byoyomi. */
810 u
->fuseki_end
= atoi(optval
);
811 } else if (!strcasecmp(optname
, "yose_start") && optval
) {
812 /* When yose_start percent of the board has been
813 * played, or if we are in byoyomi, stop spending
814 * more time and spread the remaining time
816 * Between fuseki_end and yose_start, we spend
817 * a constant proportion of the remaining time
818 * on each move. (yose_start should actually
819 * be much earlier than when real yose start,
820 * but "yose" is a good short name to convey
822 u
->yose_start
= atoi(optval
);
826 } else if (!strcasecmp(optname
, "dynkomi") && optval
) {
827 /* Dynamic komi approach; there are multiple
828 * ways to adjust komi dynamically throughout
829 * play. We currently support two: */
830 char *dynkomiarg
= strchr(optval
, ':');
833 if (!strcasecmp(optval
, "none")) {
834 u
->dynkomi
= uct_dynkomi_init_none(u
, dynkomiarg
, b
);
835 } else if (!strcasecmp(optval
, "linear")) {
836 /* You should set dynkomi_mask=1
837 * since this doesn't work well
838 * for white handicaps! */
839 u
->dynkomi
= uct_dynkomi_init_linear(u
, dynkomiarg
, b
);
840 } else if (!strcasecmp(optval
, "adaptive")) {
841 /* There are many more knobs to
842 * crank - see uct/dynkomi.c. */
843 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, dynkomiarg
, b
);
845 fprintf(stderr
, "UCT: Invalid dynkomi mode %s\n", optval
);
848 } else if (!strcasecmp(optname
, "dynkomi_mask") && optval
) {
849 /* Bitmask of colors the player must be
850 * for dynkomi be applied; you may want
851 * to use dynkomi_mask=3 to allow dynkomi
852 * even in games where Pachi is white. */
853 u
->dynkomi_mask
= atoi(optval
);
854 } else if (!strcasecmp(optname
, "dynkomi_interval") && optval
) {
855 /* If non-zero, re-adjust dynamic komi
856 * throughout a single genmove reading,
857 * roughly every N simulations. */
858 /* XXX: Does not work with tree
859 * parallelization. */
860 u
->dynkomi_interval
= atoi(optval
);
862 /** Node value result scaling */
864 } else if (!strcasecmp(optname
, "val_scale") && optval
) {
865 /* How much of the game result value should be
866 * influenced by win size. Zero means it isn't. */
867 u
->val_scale
= atof(optval
);
868 } else if (!strcasecmp(optname
, "val_points") && optval
) {
869 /* Maximum size of win to be scaled into game
870 * result value. Zero means boardsize^2. */
871 u
->val_points
= atoi(optval
) * 2; // result values are doubled
872 } else if (!strcasecmp(optname
, "val_extra")) {
873 /* If false, the score coefficient will be simply
874 * added to the value, instead of scaling the result
875 * coefficient because of it. */
876 u
->val_extra
= !optval
|| atoi(optval
);
879 /* (Purely experimental. Does not work - yet!) */
881 } else if (!strcasecmp(optname
, "local_tree") && optval
) {
882 /* Whether to bias exploration by local tree values
883 * (must be supported by the used policy).
885 * 1: Do, value = result.
886 * Try to temper the result:
887 * 2: Do, value = 0.5+(result-expected)/2.
888 * 3: Do, value = 0.5+bzz((result-expected)^2).
889 * 4: Do, value = 0.5+sqrt(result-expected)/2. */
890 u
->local_tree
= atoi(optval
);
891 } else if (!strcasecmp(optname
, "tenuki_d") && optval
) {
892 /* Tenuki distance at which to break the local tree. */
893 u
->tenuki_d
= atoi(optval
);
894 if (u
->tenuki_d
> TREE_NODE_D_MAX
+ 1) {
895 fprintf(stderr
, "uct: tenuki_d must not be larger than TREE_NODE_D_MAX+1 %d\n", TREE_NODE_D_MAX
+ 1);
898 } else if (!strcasecmp(optname
, "local_tree_aging") && optval
) {
899 /* How much to reduce local tree values between moves. */
900 u
->local_tree_aging
= atof(optval
);
901 } else if (!strcasecmp(optname
, "local_tree_depth_decay") && optval
) {
902 /* With value x>0, during the descent the node
903 * contributes 1/x^depth playouts in
904 * the local tree. I.e., with x>1, nodes more
905 * distant from local situation contribute more
906 * than nodes near the root. */
907 u
->local_tree_depth_decay
= atof(optval
);
908 } else if (!strcasecmp(optname
, "local_tree_allseq")) {
909 /* If disabled, only complete sequences are stored
910 * in the local tree. If this is on, also
911 * subsequences starting at each move are stored. */
912 u
->local_tree_allseq
= !optval
|| atoi(optval
);
913 } else if (!strcasecmp(optname
, "local_tree_rootseqval")) {
914 /* If disabled, expected node value is computed by
915 * summing up values through the whole descent.
916 * If enabled, expected node value for
917 * each sequence is the value at the root of the
919 u
->local_tree_rootseqval
= !optval
|| atoi(optval
);
921 /** Other heuristics */
922 } else if (!strcasecmp(optname
, "significant_threshold") && optval
) {
923 /* Some heuristics (XXX: none in mainline) rely
924 * on the knowledge of the last "significant"
925 * node in the descent. Such a node is
926 * considered reasonably trustworthy to carry
927 * some meaningful information in the values
928 * of the node and its children. */
929 u
->significant_threshold
= atoi(optval
);
931 /** Distributed engine slaves setup */
933 } else if (!strcasecmp(optname
, "slave")) {
934 /* Act as slave for the distributed engine. */
935 u
->slave
= !optval
|| atoi(optval
);
936 } else if (!strcasecmp(optname
, "shared_nodes") && optval
) {
937 /* Share at most shared_nodes between master and slave at each genmoves.
938 * Must use the same value in master and slaves. */
939 u
->shared_nodes
= atoi(optval
);
940 } else if (!strcasecmp(optname
, "shared_levels") && optval
) {
941 /* Share only nodes of level <= shared_levels. */
942 u
->shared_levels
= atoi(optval
);
943 } else if (!strcasecmp(optname
, "stats_hbits") && optval
) {
944 /* Set hash table size to 2^stats_hbits for the shared stats. */
945 u
->stats_hbits
= atoi(optval
);
946 } else if (!strcasecmp(optname
, "stats_delay") && optval
) {
947 /* How long to wait in slave for initial stats to build up before
948 * replying to the genmoves command (in ms) */
949 u
->stats_delay
= 0.001 * atof(optval
);
952 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
959 u
->policy
= policy_ucb1amaf_init(u
, NULL
);
961 if (!!u
->random_policy_chance
^ !!u
->random_policy
) {
962 fprintf(stderr
, "uct: Only one of random_policy and random_policy_chance is set\n");
966 if (!u
->local_tree
) {
967 /* No ltree aging. */
968 u
->local_tree_aging
= 1.0f
;
972 if (u
->pruning_threshold
< u
->max_tree_size
/ 10)
973 u
->pruning_threshold
= u
->max_tree_size
/ 10;
974 if (u
->pruning_threshold
> u
->max_tree_size
/ 2)
975 u
->pruning_threshold
= u
->max_tree_size
/ 2;
977 /* Limit pruning temp space to 20% of memory. Beyond this we discard
978 * the nodes and recompute them at the next move if necessary. */
979 u
->max_pruned_size
= u
->max_tree_size
/ 5;
980 u
->max_tree_size
-= u
->max_pruned_size
;
982 /* Reserve 5% memory in case the background free() are slower
983 * than the concurrent allocations. */
984 u
->max_tree_size
-= u
->max_tree_size
/ 20;
988 u
->prior
= uct_prior_init(NULL
, b
);
991 u
->playout
= playout_moggy_init(NULL
, b
, u
->jdict
);
992 if (!u
->playout
->debug_level
)
993 u
->playout
->debug_level
= u
->debug_level
;
995 u
->ownermap
.map
= malloc2(board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
998 if (!u
->stats_hbits
) u
->stats_hbits
= DEFAULT_STATS_HBITS
;
999 if (!u
->shared_nodes
) u
->shared_nodes
= DEFAULT_SHARED_NODES
;
1000 assert(u
->shared_levels
* board_bits2(b
) <= 8 * (int)sizeof(path_t
));
1004 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, NULL
, b
);
1006 /* Some things remain uninitialized for now - the opening tbook
1007 * is not loaded and the tree not set up. */
1008 /* This will be initialized in setup_state() at the first move
1009 * received/requested. This is because right now we are not aware
1010 * about any komi or handicap setup and such. */
1016 engine_uct_init(char *arg
, struct board
*b
)
1018 struct uct
*u
= uct_state_init(arg
, b
);
1019 struct engine
*e
= calloc2(1, sizeof(struct engine
));
1020 e
->name
= "UCT Engine";
1021 e
->printhook
= uct_printhook_ownermap
;
1022 e
->notify_play
= uct_notify_play
;
1025 e
->result
= uct_result
;
1026 e
->genmove
= uct_genmove
;
1027 e
->genmoves
= uct_genmoves
;
1028 e
->dead_group_list
= uct_dead_group_list
;
1032 e
->notify
= uct_notify
;
1034 const char banner
[] = "I'm playing UCT. When I'm losing, I will resign, "
1035 "if I think I win, I play until you pass. "
1036 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
1037 if (!u
->banner
) u
->banner
= "";
1038 e
->comment
= malloc2(sizeof(banner
) + strlen(u
->banner
) + 1);
1039 sprintf(e
->comment
, "%s %s", banner
, u
->banner
);