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
->plugins
= pluginset_init(b
);
573 u
->jdict
= joseki_load(b
->size
);
576 char *optspec
, *next
= arg
;
579 next
+= strcspn(next
, ",");
580 if (*next
) { *next
++ = 0; } else { *next
= 0; }
582 char *optname
= optspec
;
583 char *optval
= strchr(optspec
, '=');
584 if (optval
) *optval
++ = 0;
588 if (!strcasecmp(optname
, "debug")) {
590 u
->debug_level
= atoi(optval
);
593 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
594 /* When dumping the UCT tree on output, include
595 * nodes with at least this many playouts.
596 * (This value is re-scaled "intelligently"
597 * in case of very large trees.) */
598 u
->dumpthres
= atoi(optval
);
599 } else if (!strcasecmp(optname
, "resign_threshold") && optval
) {
600 /* Resign when this ratio of games is lost
601 * after GJ_MINGAMES sample is taken. */
602 u
->resign_threshold
= atof(optval
);
603 } else if (!strcasecmp(optname
, "sure_win_threshold") && optval
) {
604 /* Stop reading when this ratio of games is won
605 * after PLAYOUT_EARLY_BREAK_MIN sample is
606 * taken. (Prevents stupid time losses,
607 * friendly to human opponents.) */
608 u
->sure_win_threshold
= atof(optval
);
609 } else if (!strcasecmp(optname
, "force_seed") && optval
) {
610 /* Set RNG seed at the tree setup. */
611 u
->force_seed
= atoi(optval
);
612 } else if (!strcasecmp(optname
, "no_tbook")) {
613 /* Disable UCT opening tbook. */
615 } else if (!strcasecmp(optname
, "pass_all_alive")) {
616 /* Whether to consider passing only after all
617 * dead groups were removed from the board;
618 * this is like all genmoves are in fact
619 * kgs-genmove_cleanup. */
620 u
->pass_all_alive
= !optval
|| atoi(optval
);
621 } else if (!strcasecmp(optname
, "territory_scoring")) {
622 /* Use territory scoring (default is area scoring).
623 * An explicit kgs-rules command overrides this. */
624 u
->territory_scoring
= !optval
|| atoi(optval
);
625 } else if (!strcasecmp(optname
, "banner") && optval
) {
626 /* Additional banner string. This must come as the
627 * last engine parameter. */
628 if (*next
) *--next
= ',';
629 u
->banner
= strdup(optval
);
631 } else if (!strcasecmp(optname
, "plugin") && optval
) {
632 /* Load an external plugin; filename goes before the colon,
633 * extra arguments after the colon. */
634 char *pluginarg
= strchr(optval
, ':');
637 plugin_load(u
->plugins
, optval
, pluginarg
);
639 /** UCT behavior and policies */
641 } else if ((!strcasecmp(optname
, "policy")
642 /* Node selection policy. ucb1amaf is the
643 * default policy implementing RAVE, while
644 * ucb1 is the simple exploration/exploitation
645 * policy. Policies can take further extra
647 || !strcasecmp(optname
, "random_policy")) && optval
) {
648 /* A policy to be used randomly with small
649 * chance instead of the default policy. */
650 char *policyarg
= strchr(optval
, ':');
651 struct uct_policy
**p
= !strcasecmp(optname
, "policy") ? &u
->policy
: &u
->random_policy
;
654 if (!strcasecmp(optval
, "ucb1")) {
655 *p
= policy_ucb1_init(u
, policyarg
);
656 } else if (!strcasecmp(optval
, "ucb1amaf")) {
657 *p
= policy_ucb1amaf_init(u
, policyarg
);
659 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
662 } else if (!strcasecmp(optname
, "playout") && optval
) {
663 /* Random simulation (playout) policy.
664 * moggy is the default policy with large
665 * amount of domain-specific knowledge and
666 * heuristics. light is a simple uniformly
667 * random move selection policy. */
668 char *playoutarg
= strchr(optval
, ':');
671 if (!strcasecmp(optval
, "moggy")) {
672 u
->playout
= playout_moggy_init(playoutarg
, b
, u
->jdict
);
673 } else if (!strcasecmp(optval
, "light")) {
674 u
->playout
= playout_light_init(playoutarg
, b
);
676 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
679 } else if (!strcasecmp(optname
, "prior") && optval
) {
680 /* Node priors policy. When expanding a node,
681 * it will seed node values heuristically
682 * (most importantly, based on playout policy
683 * opinion, but also with regard to other
684 * things). See uct/prior.c for details.
685 * Use prior=eqex=0 to disable priors. */
686 u
->prior
= uct_prior_init(optval
, b
);
687 } else if (!strcasecmp(optname
, "mercy") && optval
) {
688 /* Minimal difference of black/white captures
689 * to stop playout - "Mercy Rule". Speeds up
690 * hopeless playouts at the expense of some
692 u
->mercymin
= atoi(optval
);
693 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
694 /* Maximum length of single simulation
696 u
->gamelen
= atoi(optval
);
697 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
698 /* Expand UCT nodes after it has been
699 * visited this many times. */
700 u
->expand_p
= atoi(optval
);
701 } else if (!strcasecmp(optname
, "random_policy_chance") && optval
) {
702 /* If specified (N), with probability 1/N, random_policy policy
703 * descend is used instead of main policy descend; useful
704 * if specified policy (e.g. UCB1AMAF) can make unduly biased
705 * choices sometimes, you can fall back to e.g.
706 * random_policy=UCB1. */
707 u
->random_policy_chance
= atoi(optval
);
709 /** General AMAF behavior */
710 /* (Only relevant if the policy supports AMAF.
711 * More variables can be tuned as policy
714 } else if (!strcasecmp(optname
, "playout_amaf")) {
715 /* Whether to include random playout moves in
716 * AMAF as well. (Otherwise, only tree moves
717 * are included in AMAF. Of course makes sense
718 * only in connection with an AMAF policy.) */
719 /* with-without: 55.5% (+-4.1) */
720 if (optval
&& *optval
== '0')
721 u
->playout_amaf
= false;
723 u
->playout_amaf
= true;
724 } else if (!strcasecmp(optname
, "playout_amaf_nakade")) {
725 /* Whether to include nakade moves from playouts
726 * in the AMAF statistics; this tends to nullify
727 * the playout_amaf effect by adding too much
729 if (optval
&& *optval
== '0')
730 u
->playout_amaf_nakade
= false;
732 u
->playout_amaf_nakade
= true;
733 } else if (!strcasecmp(optname
, "playout_amaf_cutoff") && optval
) {
734 /* Keep only first N% of playout stage AMAF
736 u
->playout_amaf_cutoff
= atoi(optval
);
737 } else if (!strcasecmp(optname
, "amaf_prior") && optval
) {
738 /* In node policy, consider prior values
739 * part of the real result term or part
740 * of the AMAF term? */
741 u
->amaf_prior
= atoi(optval
);
743 /** Performance and memory management */
745 } else if (!strcasecmp(optname
, "threads") && optval
) {
746 /* By default, Pachi will run with only single
747 * tree search thread! */
748 u
->threads
= atoi(optval
);
749 } else if (!strcasecmp(optname
, "thread_model") && optval
) {
750 if (!strcasecmp(optval
, "tree")) {
751 /* Tree parallelization - all threads
752 * grind on the same tree. */
753 u
->thread_model
= TM_TREE
;
755 } else if (!strcasecmp(optval
, "treevl")) {
756 /* Tree parallelization, but also
757 * with virtual losses - this discou-
758 * rages most threads choosing the
759 * same tree branches to read. */
760 u
->thread_model
= TM_TREEVL
;
762 fprintf(stderr
, "UCT: Invalid thread model %s\n", optval
);
765 } else if (!strcasecmp(optname
, "virtual_loss")) {
766 /* Number of virtual losses added before evaluating a node. */
767 u
->virtual_loss
= !optval
|| atoi(optval
);
768 } else if (!strcasecmp(optname
, "pondering")) {
769 /* Keep searching even during opponent's turn. */
770 u
->pondering_opt
= !optval
|| atoi(optval
);
771 } else if (!strcasecmp(optname
, "max_tree_size") && optval
) {
772 /* Maximum amount of memory [MiB] consumed by the move tree.
773 * For fast_alloc it includes the temp tree used for pruning.
774 * Default is 3072 (3 GiB). */
775 u
->max_tree_size
= atol(optval
) * 1048576;
776 } else if (!strcasecmp(optname
, "fast_alloc")) {
777 u
->fast_alloc
= !optval
|| atoi(optval
);
778 } else if (!strcasecmp(optname
, "pruning_threshold") && optval
) {
779 /* Force pruning at beginning of a move if the tree consumes
780 * more than this [MiB]. Default is 10% of max_tree_size.
781 * Increase to reduce pruning time overhead if memory is plentiful.
782 * This option is meaningful only for fast_alloc. */
783 u
->pruning_threshold
= atol(optval
) * 1048576;
787 } else if (!strcasecmp(optname
, "best2_ratio") && optval
) {
788 /* If set, prolong simulating while
789 * first_best/second_best playouts ratio
790 * is less than best2_ratio. */
791 u
->best2_ratio
= atof(optval
);
792 } else if (!strcasecmp(optname
, "bestr_ratio") && optval
) {
793 /* If set, prolong simulating while
794 * best,best_best_child values delta
795 * is more than bestr_ratio. */
796 u
->bestr_ratio
= atof(optval
);
797 } else if (!strcasecmp(optname
, "max_maintime_ratio") && optval
) {
798 /* If set and while not in byoyomi, prolong simulating no more than
799 * max_maintime_ratio times the normal desired thinking time. */
800 u
->max_maintime_ratio
= atof(optval
);
801 } else if (!strcasecmp(optname
, "fuseki_end") && optval
) {
802 /* At the very beginning it's not worth thinking
803 * too long because the playout evaluations are
804 * very noisy. So gradually increase the thinking
805 * time up to maximum when fuseki_end percent
806 * of the board has been played.
807 * This only applies if we are not in byoyomi. */
808 u
->fuseki_end
= atoi(optval
);
809 } else if (!strcasecmp(optname
, "yose_start") && optval
) {
810 /* When yose_start percent of the board has been
811 * played, or if we are in byoyomi, stop spending
812 * more time and spread the remaining time
814 * Between fuseki_end and yose_start, we spend
815 * a constant proportion of the remaining time
816 * on each move. (yose_start should actually
817 * be much earlier than when real yose start,
818 * but "yose" is a good short name to convey
820 u
->yose_start
= atoi(optval
);
824 } else if (!strcasecmp(optname
, "dynkomi") && optval
) {
825 /* Dynamic komi approach; there are multiple
826 * ways to adjust komi dynamically throughout
827 * play. We currently support two: */
828 char *dynkomiarg
= strchr(optval
, ':');
831 if (!strcasecmp(optval
, "none")) {
832 u
->dynkomi
= uct_dynkomi_init_none(u
, dynkomiarg
, b
);
833 } else if (!strcasecmp(optval
, "linear")) {
834 /* You should set dynkomi_mask=1
835 * since this doesn't work well
836 * for white handicaps! */
837 u
->dynkomi
= uct_dynkomi_init_linear(u
, dynkomiarg
, b
);
838 } else if (!strcasecmp(optval
, "adaptive")) {
839 /* There are many more knobs to
840 * crank - see uct/dynkomi.c. */
841 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, dynkomiarg
, b
);
843 fprintf(stderr
, "UCT: Invalid dynkomi mode %s\n", optval
);
846 } else if (!strcasecmp(optname
, "dynkomi_mask") && optval
) {
847 /* Bitmask of colors the player must be
848 * for dynkomi be applied; you may want
849 * to use dynkomi_mask=3 to allow dynkomi
850 * even in games where Pachi is white. */
851 u
->dynkomi_mask
= atoi(optval
);
852 } else if (!strcasecmp(optname
, "dynkomi_interval") && optval
) {
853 /* If non-zero, re-adjust dynamic komi
854 * throughout a single genmove reading,
855 * roughly every N simulations. */
856 /* XXX: Does not work with tree
857 * parallelization. */
858 u
->dynkomi_interval
= atoi(optval
);
860 /** Node value result scaling */
862 } else if (!strcasecmp(optname
, "val_scale") && optval
) {
863 /* How much of the game result value should be
864 * influenced by win size. Zero means it isn't. */
865 u
->val_scale
= atof(optval
);
866 } else if (!strcasecmp(optname
, "val_points") && optval
) {
867 /* Maximum size of win to be scaled into game
868 * result value. Zero means boardsize^2. */
869 u
->val_points
= atoi(optval
) * 2; // result values are doubled
870 } else if (!strcasecmp(optname
, "val_extra")) {
871 /* If false, the score coefficient will be simply
872 * added to the value, instead of scaling the result
873 * coefficient because of it. */
874 u
->val_extra
= !optval
|| atoi(optval
);
877 /* (Purely experimental. Does not work - yet!) */
879 } else if (!strcasecmp(optname
, "local_tree") && optval
) {
880 /* Whether to bias exploration by local tree values
881 * (must be supported by the used policy).
883 * 1: Do, value = result.
884 * Try to temper the result:
885 * 2: Do, value = 0.5+(result-expected)/2.
886 * 3: Do, value = 0.5+bzz((result-expected)^2).
887 * 4: Do, value = 0.5+sqrt(result-expected)/2. */
888 u
->local_tree
= atoi(optval
);
889 } else if (!strcasecmp(optname
, "tenuki_d") && optval
) {
890 /* Tenuki distance at which to break the local tree. */
891 u
->tenuki_d
= atoi(optval
);
892 if (u
->tenuki_d
> TREE_NODE_D_MAX
+ 1) {
893 fprintf(stderr
, "uct: tenuki_d must not be larger than TREE_NODE_D_MAX+1 %d\n", TREE_NODE_D_MAX
+ 1);
896 } else if (!strcasecmp(optname
, "local_tree_aging") && optval
) {
897 /* How much to reduce local tree values between moves. */
898 u
->local_tree_aging
= atof(optval
);
899 } else if (!strcasecmp(optname
, "local_tree_depth_decay") && optval
) {
900 /* With value x>0, during the descent the node
901 * contributes 1/x^depth playouts in
902 * the local tree. I.e., with x>1, nodes more
903 * distant from local situation contribute more
904 * than nodes near the root. */
905 u
->local_tree_depth_decay
= atof(optval
);
906 } else if (!strcasecmp(optname
, "local_tree_allseq")) {
907 /* If disabled, only complete sequences are stored
908 * in the local tree. If this is on, also
909 * subsequences starting at each move are stored. */
910 u
->local_tree_allseq
= !optval
|| atoi(optval
);
911 } else if (!strcasecmp(optname
, "local_tree_rootseqval")) {
912 /* If disabled, expected node value is computed by
913 * summing up values through the whole descent.
914 * If enabled, expected node value for
915 * each sequence is the value at the root of the
917 u
->local_tree_rootseqval
= !optval
|| atoi(optval
);
919 /** Other heuristics */
920 } else if (!strcasecmp(optname
, "significant_threshold") && optval
) {
921 /* Some heuristics (XXX: none in mainline) rely
922 * on the knowledge of the last "significant"
923 * node in the descent. Such a node is
924 * considered reasonably trustworthy to carry
925 * some meaningful information in the values
926 * of the node and its children. */
927 u
->significant_threshold
= atoi(optval
);
929 /** Distributed engine slaves setup */
931 } else if (!strcasecmp(optname
, "slave")) {
932 /* Act as slave for the distributed engine. */
933 u
->slave
= !optval
|| atoi(optval
);
934 } else if (!strcasecmp(optname
, "shared_nodes") && optval
) {
935 /* Share at most shared_nodes between master and slave at each genmoves.
936 * Must use the same value in master and slaves. */
937 u
->shared_nodes
= atoi(optval
);
938 } else if (!strcasecmp(optname
, "shared_levels") && optval
) {
939 /* Share only nodes of level <= shared_levels. */
940 u
->shared_levels
= atoi(optval
);
941 } else if (!strcasecmp(optname
, "stats_hbits") && optval
) {
942 /* Set hash table size to 2^stats_hbits for the shared stats. */
943 u
->stats_hbits
= atoi(optval
);
946 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
953 u
->policy
= policy_ucb1amaf_init(u
, NULL
);
955 if (!!u
->random_policy_chance
^ !!u
->random_policy
) {
956 fprintf(stderr
, "uct: Only one of random_policy and random_policy_chance is set\n");
960 if (!u
->local_tree
) {
961 /* No ltree aging. */
962 u
->local_tree_aging
= 1.0f
;
966 if (u
->pruning_threshold
< u
->max_tree_size
/ 10)
967 u
->pruning_threshold
= u
->max_tree_size
/ 10;
968 if (u
->pruning_threshold
> u
->max_tree_size
/ 2)
969 u
->pruning_threshold
= u
->max_tree_size
/ 2;
971 /* Limit pruning temp space to 20% of memory. Beyond this we discard
972 * the nodes and recompute them at the next move if necessary. */
973 u
->max_pruned_size
= u
->max_tree_size
/ 5;
974 u
->max_tree_size
-= u
->max_pruned_size
;
976 /* Reserve 5% memory in case the background free() are slower
977 * than the concurrent allocations. */
978 u
->max_tree_size
-= u
->max_tree_size
/ 20;
982 u
->prior
= uct_prior_init(NULL
, b
);
985 u
->playout
= playout_moggy_init(NULL
, b
, u
->jdict
);
986 if (!u
->playout
->debug_level
)
987 u
->playout
->debug_level
= u
->debug_level
;
989 u
->ownermap
.map
= malloc2(board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
992 if (!u
->stats_hbits
) u
->stats_hbits
= DEFAULT_STATS_HBITS
;
993 if (!u
->shared_nodes
) u
->shared_nodes
= DEFAULT_SHARED_NODES
;
994 assert(u
->shared_levels
* board_bits2(b
) <= 8 * (int)sizeof(path_t
));
998 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, NULL
, b
);
1000 /* Some things remain uninitialized for now - the opening tbook
1001 * is not loaded and the tree not set up. */
1002 /* This will be initialized in setup_state() at the first move
1003 * received/requested. This is because right now we are not aware
1004 * about any komi or handicap setup and such. */
1010 engine_uct_init(char *arg
, struct board
*b
)
1012 struct uct
*u
= uct_state_init(arg
, b
);
1013 struct engine
*e
= calloc2(1, sizeof(struct engine
));
1014 e
->name
= "UCT Engine";
1015 e
->printhook
= uct_printhook_ownermap
;
1016 e
->notify_play
= uct_notify_play
;
1019 e
->result
= uct_result
;
1020 e
->genmove
= uct_genmove
;
1021 e
->genmoves
= uct_genmoves
;
1022 e
->dead_group_list
= uct_dead_group_list
;
1026 e
->notify
= uct_notify
;
1028 const char banner
[] = "I'm playing UCT. When I'm losing, I will resign, "
1029 "if I think I win, I play until you pass. "
1030 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
1031 if (!u
->banner
) u
->banner
= "";
1032 e
->comment
= malloc2(sizeof(banner
) + strlen(u
->banner
) + 1);
1033 sprintf(e
->comment
, "%s %s", banner
, u
->banner
);