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
, struct board
*board
);
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 if (!u
->no_tbook
&& b
->moves
== 0) {
49 assert(color
== S_BLACK
);
55 reset_state(struct uct
*u
)
58 tree_done(u
->t
); u
->t
= NULL
;
62 setup_dynkomi(struct uct
*u
, struct board
*b
, enum stone to_play
)
64 if (u
->t
->use_extra_komi
&& !u
->pondering
&& u
->dynkomi
->permove
)
65 u
->t
->extra_komi
= u
->dynkomi
->permove(u
->dynkomi
, b
, u
->t
);
66 else if (!u
->t
->use_extra_komi
)
71 uct_prepare_move(struct uct
*u
, struct board
*b
, enum stone color
)
74 /* Verify that we have sane state. */
76 assert(u
->t
&& b
->moves
);
77 if (color
!= stone_other(u
->t
->root_color
)) {
78 fprintf(stderr
, "Fatal: Non-alternating play detected %d %d\n",
79 color
, u
->t
->root_color
);
82 uct_htable_reset(u
->t
);
85 /* We need fresh state. */
87 setup_state(u
, b
, color
);
90 u
->ownermap
.playouts
= 0;
91 memset(u
->ownermap
.map
, 0, board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
92 u
->played_own
= u
->played_all
= 0;
96 dead_group_list(struct uct
*u
, struct board
*b
, struct move_queue
*mq
)
98 struct group_judgement gj
;
100 gj
.gs
= alloca(board_size2(b
) * sizeof(gj
.gs
[0]));
101 board_ownermap_judge_group(b
, &u
->ownermap
, &gj
);
102 groups_of_status(b
, &gj
, GS_DEAD
, mq
);
106 uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
)
108 /* Make sure enough playouts are simulated to get a reasonable dead group list. */
109 while (u
->ownermap
.playouts
< GJ_MINGAMES
)
110 uct_playout(u
, b
, color
, u
->t
);
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(node_coord(n
), 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(node_coord(n
), 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 not wait for
311 * 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
= { node_coord(t
->root
), 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 setup_dynkomi(u
, b
, color
);
404 if (b
->rules
== RULES_JAPANESE
)
405 u
->territory_scoring
= true;
407 /* Make pessimistic assumption about komi for Japanese rules to
408 * avoid losing by 0.5 when winning by 0.5 with Chinese rules.
409 * The rules usually give the same winner if the integer part of komi
410 * is odd so we adjust the komi only if it is even (for a board of
411 * odd size). We are not trying to get an exact evaluation for rare
412 * cases of seki. For details see http://home.snafu.de/jasiek/parity.html */
413 if (u
->territory_scoring
&& (((int)floor(b
->komi
) + board_size(b
)) & 1)) {
414 b
->komi
+= (color
== S_BLACK
? 1.0 : -1.0);
416 fprintf(stderr
, "Setting komi to %.1f assuming Japanese rules\n",
422 uct_genmove(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
424 double start_time
= time_now();
425 struct uct
*u
= e
->data
;
426 u
->pass_all_alive
|= pass_all_alive
;
427 uct_pondering_stop(u
);
428 uct_genmove_setup(u
, b
, color
);
430 /* Start the Monte Carlo Tree Search! */
431 int base_playouts
= u
->t
->root
->u
.playouts
;
432 int played_games
= uct_search(u
, b
, ti
, color
, u
->t
);
435 struct tree_node
*best
;
436 best
= uct_search_result(u
, b
, color
, u
->pass_all_alive
, played_games
, base_playouts
, &best_coord
);
439 double time
= time_now() - start_time
+ 0.000001; /* avoid divide by zero */
440 fprintf(stderr
, "genmove in %0.2fs (%d games/s, %d games/s/thread)\n",
441 time
, (int)(played_games
/time
), (int)(played_games
/time
/u
->threads
));
445 /* Pass or resign. */
447 return coord_copy(best_coord
);
449 tree_promote_node(u
->t
, &best
);
451 /* After a pass, pondering is harmful for two reasons:
452 * (i) We might keep pondering even when the game is over.
453 * Of course this is the case for opponent resign as well.
454 * (ii) More importantly, the ownermap will get skewed since
455 * the UCT will start cutting off any playouts. */
456 if (u
->pondering_opt
&& !is_pass(node_coord(best
))) {
457 uct_pondering_start(u
, b
, u
->t
, stone_other(color
));
459 return coord_copy(best_coord
);
464 uct_gentbook(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
)
466 struct uct
*u
= e
->data
;
467 if (!u
->t
) uct_prepare_move(u
, b
, color
);
470 if (ti
->dim
== TD_GAMES
) {
471 /* Don't count in games that already went into the tbook. */
472 ti
->len
.games
+= u
->t
->root
->u
.playouts
;
474 uct_search(u
, b
, ti
, color
, u
->t
);
476 assert(ti
->dim
== TD_GAMES
);
477 tree_save(u
->t
, b
, ti
->len
.games
/ 100);
483 uct_dumptbook(struct engine
*e
, struct board
*b
, enum stone color
)
485 struct uct
*u
= e
->data
;
486 struct tree
*t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0,
487 u
->max_pruned_size
, u
->pruning_threshold
, u
->local_tree_aging
, 0);
495 uct_evaluate_one(struct engine
*e
, struct board
*b
, struct time_info
*ti
, coord_t c
, enum stone color
)
497 struct uct
*u
= e
->data
;
501 struct move m
= { c
, color
};
502 int res
= board_play(&b2
, &m
);
505 color
= stone_other(color
);
507 if (u
->t
) reset_state(u
);
508 uct_prepare_move(u
, &b2
, color
);
512 uct_search(u
, &b2
, ti
, color
, u
->t
);
513 struct tree_node
*best
= u
->policy
->choose(u
->policy
, u
->t
->root
, &b2
, color
, resign
);
515 bestval
= NAN
; // the opponent has no reply!
517 bestval
= tree_node_get_value(u
->t
, 1, best
->u
.value
);
520 reset_state(u
); // clean our junk
522 return isnan(bestval
) ? NAN
: 1.0f
- bestval
;
526 uct_evaluate(struct engine
*e
, struct board
*b
, struct time_info
*ti
, floating_t
*vals
, enum stone color
)
528 for (int i
= 0; i
< b
->flen
; i
++) {
529 if (is_pass(b
->f
[i
]))
532 vals
[i
] = uct_evaluate_one(e
, b
, ti
, b
->f
[i
], color
);
538 uct_state_init(char *arg
, struct board
*b
)
540 struct uct
*u
= calloc2(1, sizeof(struct uct
));
541 bool pat_setup
= false;
543 u
->debug_level
= debug_level
;
544 u
->gamelen
= MC_GAMELEN
;
545 u
->resign_threshold
= 0.2;
546 u
->sure_win_threshold
= 0.9;
548 u
->significant_threshold
= 50;
551 u
->playout_amaf
= true;
552 u
->amaf_prior
= false;
553 u
->max_tree_size
= 1408ULL * 1048576;
554 u
->fast_alloc
= true;
555 u
->pruning_threshold
= 0;
558 u
->thread_model
= TM_TREEVL
;
561 u
->fuseki_end
= 20; // max time at 361*20% = 72 moves (our 36th move, still 99 to play)
562 u
->yose_start
= 40; // (100-40-25)*361/100/2 = 63 moves still to play by us then
563 u
->bestr_ratio
= 0.02;
564 // 2.5 is clearly too much, but seems to compensate well for overly stern time allocations.
565 // TODO: Further tuning and experiments with better time allocation schemes.
566 u
->best2_ratio
= 2.5;
567 u
->max_maintime_ratio
= 3.0;
569 u
->val_scale
= 0; u
->val_points
= 40;
570 u
->dynkomi_interval
= 1000;
571 u
->dynkomi_mask
= S_BLACK
| S_WHITE
;
574 u
->local_tree_aging
= 80;
575 u
->local_tree_depth_decay
= 1.5;
576 u
->local_tree_eval
= LTE_ROOT
;
577 u
->local_tree_neival
= true;
581 u
->stats_delay
= 0.01; // 10 ms
583 u
->plugins
= pluginset_init(b
);
585 u
->jdict
= joseki_load(b
->size
);
588 char *optspec
, *next
= arg
;
591 next
+= strcspn(next
, ",");
592 if (*next
) { *next
++ = 0; } else { *next
= 0; }
594 char *optname
= optspec
;
595 char *optval
= strchr(optspec
, '=');
596 if (optval
) *optval
++ = 0;
600 if (!strcasecmp(optname
, "debug")) {
602 u
->debug_level
= atoi(optval
);
605 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
606 /* When dumping the UCT tree on output, include
607 * nodes with at least this many playouts.
608 * (This value is re-scaled "intelligently"
609 * in case of very large trees.) */
610 u
->dumpthres
= atoi(optval
);
611 } else if (!strcasecmp(optname
, "resign_threshold") && optval
) {
612 /* Resign when this ratio of games is lost
613 * after GJ_MINGAMES sample is taken. */
614 u
->resign_threshold
= atof(optval
);
615 } else if (!strcasecmp(optname
, "sure_win_threshold") && optval
) {
616 /* Stop reading when this ratio of games is won
617 * after PLAYOUT_EARLY_BREAK_MIN sample is
618 * taken. (Prevents stupid time losses,
619 * friendly to human opponents.) */
620 u
->sure_win_threshold
= atof(optval
);
621 } else if (!strcasecmp(optname
, "force_seed") && optval
) {
622 /* Set RNG seed at the tree setup. */
623 u
->force_seed
= atoi(optval
);
624 } else if (!strcasecmp(optname
, "no_tbook")) {
625 /* Disable UCT opening tbook. */
627 } else if (!strcasecmp(optname
, "pass_all_alive")) {
628 /* Whether to consider passing only after all
629 * dead groups were removed from the board;
630 * this is like all genmoves are in fact
631 * kgs-genmove_cleanup. */
632 u
->pass_all_alive
= !optval
|| atoi(optval
);
633 } else if (!strcasecmp(optname
, "territory_scoring")) {
634 /* Use territory scoring (default is area scoring).
635 * An explicit kgs-rules command overrides this. */
636 u
->territory_scoring
= !optval
|| atoi(optval
);
637 } else if (!strcasecmp(optname
, "stones_only")) {
638 /* Do not count eyes. Nice to teach go to kids.
639 * http://strasbourg.jeudego.org/regle_strasbourgeoise.htm */
640 b
->rules
= RULES_STONES_ONLY
;
641 u
->pass_all_alive
= true;
642 } else if (!strcasecmp(optname
, "banner") && optval
) {
643 /* Additional banner string. This must come as the
644 * last engine parameter. */
645 if (*next
) *--next
= ',';
646 u
->banner
= strdup(optval
);
648 } else if (!strcasecmp(optname
, "plugin") && optval
) {
649 /* Load an external plugin; filename goes before the colon,
650 * extra arguments after the colon. */
651 char *pluginarg
= strchr(optval
, ':');
654 plugin_load(u
->plugins
, optval
, pluginarg
);
656 /** UCT behavior and policies */
658 } else if ((!strcasecmp(optname
, "policy")
659 /* Node selection policy. ucb1amaf is the
660 * default policy implementing RAVE, while
661 * ucb1 is the simple exploration/exploitation
662 * policy. Policies can take further extra
664 || !strcasecmp(optname
, "random_policy")) && optval
) {
665 /* A policy to be used randomly with small
666 * chance instead of the default policy. */
667 char *policyarg
= strchr(optval
, ':');
668 struct uct_policy
**p
= !strcasecmp(optname
, "policy") ? &u
->policy
: &u
->random_policy
;
671 if (!strcasecmp(optval
, "ucb1")) {
672 *p
= policy_ucb1_init(u
, policyarg
);
673 } else if (!strcasecmp(optval
, "ucb1amaf")) {
674 *p
= policy_ucb1amaf_init(u
, policyarg
, b
);
676 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
679 } else if (!strcasecmp(optname
, "playout") && optval
) {
680 /* Random simulation (playout) policy.
681 * moggy is the default policy with large
682 * amount of domain-specific knowledge and
683 * heuristics. light is a simple uniformly
684 * random move selection policy. */
685 char *playoutarg
= strchr(optval
, ':');
688 if (!strcasecmp(optval
, "moggy")) {
689 u
->playout
= playout_moggy_init(playoutarg
, b
, u
->jdict
);
690 } else if (!strcasecmp(optval
, "light")) {
691 u
->playout
= playout_light_init(playoutarg
, b
);
693 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
696 } else if (!strcasecmp(optname
, "prior") && optval
) {
697 /* Node priors policy. When expanding a node,
698 * it will seed node values heuristically
699 * (most importantly, based on playout policy
700 * opinion, but also with regard to other
701 * things). See uct/prior.c for details.
702 * Use prior=eqex=0 to disable priors. */
703 u
->prior
= uct_prior_init(optval
, b
, u
);
704 } else if (!strcasecmp(optname
, "mercy") && optval
) {
705 /* Minimal difference of black/white captures
706 * to stop playout - "Mercy Rule". Speeds up
707 * hopeless playouts at the expense of some
709 u
->mercymin
= atoi(optval
);
710 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
711 /* Maximum length of single simulation
713 u
->gamelen
= atoi(optval
);
714 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
715 /* Expand UCT nodes after it has been
716 * visited this many times. */
717 u
->expand_p
= atoi(optval
);
718 } else if (!strcasecmp(optname
, "random_policy_chance") && optval
) {
719 /* If specified (N), with probability 1/N, random_policy policy
720 * descend is used instead of main policy descend; useful
721 * if specified policy (e.g. UCB1AMAF) can make unduly biased
722 * choices sometimes, you can fall back to e.g.
723 * random_policy=UCB1. */
724 u
->random_policy_chance
= atoi(optval
);
726 /** General AMAF behavior */
727 /* (Only relevant if the policy supports AMAF.
728 * More variables can be tuned as policy
731 } else if (!strcasecmp(optname
, "playout_amaf")) {
732 /* Whether to include random playout moves in
733 * AMAF as well. (Otherwise, only tree moves
734 * are included in AMAF. Of course makes sense
735 * only in connection with an AMAF policy.) */
736 /* with-without: 55.5% (+-4.1) */
737 if (optval
&& *optval
== '0')
738 u
->playout_amaf
= false;
740 u
->playout_amaf
= true;
741 } else if (!strcasecmp(optname
, "playout_amaf_cutoff") && optval
) {
742 /* Keep only first N% of playout stage AMAF
744 u
->playout_amaf_cutoff
= atoi(optval
);
745 } else if (!strcasecmp(optname
, "amaf_prior") && optval
) {
746 /* In node policy, consider prior values
747 * part of the real result term or part
748 * of the AMAF term? */
749 u
->amaf_prior
= atoi(optval
);
751 /** Performance and memory management */
753 } else if (!strcasecmp(optname
, "threads") && optval
) {
754 /* By default, Pachi will run with only single
755 * tree search thread! */
756 u
->threads
= atoi(optval
);
757 } else if (!strcasecmp(optname
, "thread_model") && optval
) {
758 if (!strcasecmp(optval
, "tree")) {
759 /* Tree parallelization - all threads
760 * grind on the same tree. */
761 u
->thread_model
= TM_TREE
;
763 } else if (!strcasecmp(optval
, "treevl")) {
764 /* Tree parallelization, but also
765 * with virtual losses - this discou-
766 * rages most threads choosing the
767 * same tree branches to read. */
768 u
->thread_model
= TM_TREEVL
;
770 fprintf(stderr
, "UCT: Invalid thread model %s\n", optval
);
773 } else if (!strcasecmp(optname
, "virtual_loss")) {
774 /* Number of virtual losses added before evaluating a node. */
775 u
->virtual_loss
= !optval
|| atoi(optval
);
776 } else if (!strcasecmp(optname
, "pondering")) {
777 /* Keep searching even during opponent's turn. */
778 u
->pondering_opt
= !optval
|| atoi(optval
);
779 } else if (!strcasecmp(optname
, "max_tree_size") && optval
) {
780 /* Maximum amount of memory [MiB] consumed by the move tree.
781 * For fast_alloc it includes the temp tree used for pruning.
782 * Default is 3072 (3 GiB). */
783 u
->max_tree_size
= atol(optval
) * 1048576;
784 } else if (!strcasecmp(optname
, "fast_alloc")) {
785 u
->fast_alloc
= !optval
|| atoi(optval
);
786 } else if (!strcasecmp(optname
, "pruning_threshold") && optval
) {
787 /* Force pruning at beginning of a move if the tree consumes
788 * more than this [MiB]. Default is 10% of max_tree_size.
789 * Increase to reduce pruning time overhead if memory is plentiful.
790 * This option is meaningful only for fast_alloc. */
791 u
->pruning_threshold
= atol(optval
) * 1048576;
795 } else if (!strcasecmp(optname
, "best2_ratio") && optval
) {
796 /* If set, prolong simulating while
797 * first_best/second_best playouts ratio
798 * is less than best2_ratio. */
799 u
->best2_ratio
= atof(optval
);
800 } else if (!strcasecmp(optname
, "bestr_ratio") && optval
) {
801 /* If set, prolong simulating while
802 * best,best_best_child values delta
803 * is more than bestr_ratio. */
804 u
->bestr_ratio
= atof(optval
);
805 } else if (!strcasecmp(optname
, "max_maintime_ratio") && optval
) {
806 /* If set and while not in byoyomi, prolong simulating no more than
807 * max_maintime_ratio times the normal desired thinking time. */
808 u
->max_maintime_ratio
= atof(optval
);
809 } else if (!strcasecmp(optname
, "fuseki_end") && optval
) {
810 /* At the very beginning it's not worth thinking
811 * too long because the playout evaluations are
812 * very noisy. So gradually increase the thinking
813 * time up to maximum when fuseki_end percent
814 * of the board has been played.
815 * This only applies if we are not in byoyomi. */
816 u
->fuseki_end
= atoi(optval
);
817 } else if (!strcasecmp(optname
, "yose_start") && optval
) {
818 /* When yose_start percent of the board has been
819 * played, or if we are in byoyomi, stop spending
820 * more time and spread the remaining time
822 * Between fuseki_end and yose_start, we spend
823 * a constant proportion of the remaining time
824 * on each move. (yose_start should actually
825 * be much earlier than when real yose start,
826 * but "yose" is a good short name to convey
828 u
->yose_start
= atoi(optval
);
832 } else if (!strcasecmp(optname
, "dynkomi") && optval
) {
833 /* Dynamic komi approach; there are multiple
834 * ways to adjust komi dynamically throughout
835 * play. We currently support two: */
836 char *dynkomiarg
= strchr(optval
, ':');
839 if (!strcasecmp(optval
, "none")) {
840 u
->dynkomi
= uct_dynkomi_init_none(u
, dynkomiarg
, b
);
841 } else if (!strcasecmp(optval
, "linear")) {
842 /* You should set dynkomi_mask=1 or a very low
843 * handicap_value for white. */
844 u
->dynkomi
= uct_dynkomi_init_linear(u
, dynkomiarg
, b
);
845 } else if (!strcasecmp(optval
, "adaptive")) {
846 /* There are many more knobs to
847 * crank - see uct/dynkomi.c. */
848 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, dynkomiarg
, b
);
850 fprintf(stderr
, "UCT: Invalid dynkomi mode %s\n", optval
);
853 } else if (!strcasecmp(optname
, "dynkomi_mask") && optval
) {
854 /* Bitmask of colors the player must be
855 * for dynkomi be applied; the default dynkomi_mask=3 allows
856 * dynkomi even in games where Pachi is white. */
857 u
->dynkomi_mask
= atoi(optval
);
858 } else if (!strcasecmp(optname
, "dynkomi_interval") && optval
) {
859 /* If non-zero, re-adjust dynamic komi
860 * throughout a single genmove reading,
861 * roughly every N simulations. */
862 /* XXX: Does not work with tree
863 * parallelization. */
864 u
->dynkomi_interval
= atoi(optval
);
866 /** Node value result scaling */
868 } else if (!strcasecmp(optname
, "val_scale") && optval
) {
869 /* How much of the game result value should be
870 * influenced by win size. Zero means it isn't. */
871 u
->val_scale
= atof(optval
);
872 } else if (!strcasecmp(optname
, "val_points") && optval
) {
873 /* Maximum size of win to be scaled into game
874 * result value. Zero means boardsize^2. */
875 u
->val_points
= atoi(optval
) * 2; // result values are doubled
876 } else if (!strcasecmp(optname
, "val_extra")) {
877 /* If false, the score coefficient will be simply
878 * added to the value, instead of scaling the result
879 * coefficient because of it. */
880 u
->val_extra
= !optval
|| atoi(optval
);
883 /* (Purely experimental. Does not work - yet!) */
885 } else if (!strcasecmp(optname
, "local_tree")) {
886 /* Whether to bias exploration by local tree values. */
887 u
->local_tree
= !optval
|| atoi(optval
);
888 } else if (!strcasecmp(optname
, "tenuki_d") && optval
) {
889 /* Tenuki distance at which to break the local tree. */
890 u
->tenuki_d
= atoi(optval
);
891 if (u
->tenuki_d
> TREE_NODE_D_MAX
+ 1) {
892 fprintf(stderr
, "uct: tenuki_d must not be larger than TREE_NODE_D_MAX+1 %d\n", TREE_NODE_D_MAX
+ 1);
895 } else if (!strcasecmp(optname
, "local_tree_aging") && optval
) {
896 /* How much to reduce local tree values between moves. */
897 u
->local_tree_aging
= atof(optval
);
898 } else if (!strcasecmp(optname
, "local_tree_depth_decay") && optval
) {
899 /* With value x>0, during the descent the node
900 * contributes 1/x^depth playouts in
901 * the local tree. I.e., with x>1, nodes more
902 * distant from local situation contribute more
903 * than nodes near the root. */
904 u
->local_tree_depth_decay
= atof(optval
);
905 } else if (!strcasecmp(optname
, "local_tree_allseq")) {
906 /* If disabled, only complete sequences are stored
907 * in the local tree. If this is on, also
908 * subsequences starting at each move are stored. */
909 u
->local_tree_allseq
= !optval
|| atoi(optval
);
910 } else if (!strcasecmp(optname
, "local_tree_neival")) {
911 /* If disabled, local node value is not
912 * computed just based on terminal status
913 * of the coordinate, but also its neighbors. */
914 u
->local_tree_neival
= !optval
|| atoi(optval
);
915 } else if (!strcasecmp(optname
, "local_tree_eval")) {
916 /* How is the value inserted in the local tree
918 if (!strcasecmp(optval
, "root"))
919 /* All moves within a tree branch are
920 * considered wrt. their merit
921 * reaching tachtical goal of making
922 * the first move in the branch
924 u
->local_tree_eval
= LTE_ROOT
;
925 else if (!strcasecmp(optval
, "each"))
926 /* Each move is considered wrt.
927 * its own survival. */
928 u
->local_tree_eval
= LTE_EACH
;
929 else if (!strcasecmp(optval
, "total"))
930 /* The tactical goal is the survival
931 * of all the moves of my color and
932 * non-survival of all the opponent
933 * moves. Local values (and their
934 * inverses) are averaged. */
935 u
->local_tree_eval
= LTE_TOTAL
;
937 fprintf(stderr
, "uct: unknown local_tree_eval %s\n", optval
);
940 } else if (!strcasecmp(optname
, "local_tree_rootchoose")) {
941 /* If disabled, only moves within the local
942 * tree branch are considered; the values
943 * of the branch roots (i.e. root children)
944 * are ignored. This may make sense together
945 * with eval!=each, we consider only moves
946 * that influence the goal, not the "rating"
947 * of the goal itself. (The real solution
948 * will be probably using criticality to pick
949 * local tree branches.) */
950 u
->local_tree_rootchoose
= !optval
|| atoi(optval
);
952 /** Other heuristics */
953 } else if (!strcasecmp(optname
, "patterns")) {
954 /* Load pattern database. Various modules
955 * (priors, policies etc.) may make use
956 * of this database. They will request
957 * it automatically in that case, but you
958 * can use this option to tweak the pattern
960 patterns_init(&u
->pat
, optval
, false, true);
961 u
->want_pat
= pat_setup
= true;
962 } else if (!strcasecmp(optname
, "significant_threshold") && optval
) {
963 /* Some heuristics (XXX: none in mainline) rely
964 * on the knowledge of the last "significant"
965 * node in the descent. Such a node is
966 * considered reasonably trustworthy to carry
967 * some meaningful information in the values
968 * of the node and its children. */
969 u
->significant_threshold
= atoi(optval
);
971 /** Distributed engine slaves setup */
973 } else if (!strcasecmp(optname
, "slave")) {
974 /* Act as slave for the distributed engine. */
975 u
->slave
= !optval
|| atoi(optval
);
976 } else if (!strcasecmp(optname
, "slave_index") && optval
) {
977 /* Optional index if per-slave behavior is desired.
978 * Must be given as index/max */
979 u
->slave_index
= atoi(optval
);
980 char *p
= strchr(optval
, '/');
981 if (p
) u
->max_slaves
= atoi(++p
);
982 } else if (!strcasecmp(optname
, "shared_nodes") && optval
) {
983 /* Share at most shared_nodes between master and slave at each genmoves.
984 * Must use the same value in master and slaves. */
985 u
->shared_nodes
= atoi(optval
);
986 } else if (!strcasecmp(optname
, "shared_levels") && optval
) {
987 /* Share only nodes of level <= shared_levels. */
988 u
->shared_levels
= atoi(optval
);
989 } else if (!strcasecmp(optname
, "stats_hbits") && optval
) {
990 /* Set hash table size to 2^stats_hbits for the shared stats. */
991 u
->stats_hbits
= atoi(optval
);
992 } else if (!strcasecmp(optname
, "stats_delay") && optval
) {
993 /* How long to wait in slave for initial stats to build up before
994 * replying to the genmoves command (in ms) */
995 u
->stats_delay
= 0.001 * atof(optval
);
998 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
1005 u
->policy
= policy_ucb1amaf_init(u
, NULL
, b
);
1007 if (!!u
->random_policy_chance
^ !!u
->random_policy
) {
1008 fprintf(stderr
, "uct: Only one of random_policy and random_policy_chance is set\n");
1012 if (!u
->local_tree
) {
1013 /* No ltree aging. */
1014 u
->local_tree_aging
= 1.0f
;
1017 if (u
->fast_alloc
) {
1018 if (u
->pruning_threshold
< u
->max_tree_size
/ 10)
1019 u
->pruning_threshold
= u
->max_tree_size
/ 10;
1020 if (u
->pruning_threshold
> u
->max_tree_size
/ 2)
1021 u
->pruning_threshold
= u
->max_tree_size
/ 2;
1023 /* Limit pruning temp space to 20% of memory. Beyond this we discard
1024 * the nodes and recompute them at the next move if necessary. */
1025 u
->max_pruned_size
= u
->max_tree_size
/ 5;
1026 u
->max_tree_size
-= u
->max_pruned_size
;
1028 /* Reserve 5% memory in case the background free() are slower
1029 * than the concurrent allocations. */
1030 u
->max_tree_size
-= u
->max_tree_size
/ 20;
1034 u
->prior
= uct_prior_init(NULL
, b
, u
);
1037 u
->playout
= playout_moggy_init(NULL
, b
, u
->jdict
);
1038 if (!u
->playout
->debug_level
)
1039 u
->playout
->debug_level
= u
->debug_level
;
1041 if (u
->want_pat
&& !pat_setup
)
1042 patterns_init(&u
->pat
, NULL
, false, true);
1044 u
->ownermap
.map
= malloc2(board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
1047 if (!u
->stats_hbits
) u
->stats_hbits
= DEFAULT_STATS_HBITS
;
1048 if (!u
->shared_nodes
) u
->shared_nodes
= DEFAULT_SHARED_NODES
;
1049 assert(u
->shared_levels
* board_bits2(b
) <= 8 * (int)sizeof(path_t
));
1053 u
->dynkomi
= uct_dynkomi_init_linear(u
, NULL
, b
);
1055 /* Some things remain uninitialized for now - the opening tbook
1056 * is not loaded and the tree not set up. */
1057 /* This will be initialized in setup_state() at the first move
1058 * received/requested. This is because right now we are not aware
1059 * about any komi or handicap setup and such. */
1065 engine_uct_init(char *arg
, struct board
*b
)
1067 struct uct
*u
= uct_state_init(arg
, b
);
1068 struct engine
*e
= calloc2(1, sizeof(struct engine
));
1070 e
->printhook
= uct_printhook_ownermap
;
1071 e
->notify_play
= uct_notify_play
;
1074 e
->result
= uct_result
;
1075 e
->genmove
= uct_genmove
;
1076 e
->genmoves
= uct_genmoves
;
1077 e
->evaluate
= uct_evaluate
;
1078 e
->dead_group_list
= uct_dead_group_list
;
1082 e
->notify
= uct_notify
;
1084 const char banner
[] = "If you believe you have won but I am still playing, "
1085 "please help me understand by capturing all dead stones. "
1086 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
1087 if (!u
->banner
) u
->banner
= "";
1088 e
->comment
= malloc2(sizeof(banner
) + strlen(u
->banner
) + 1);
1089 sprintf(e
->comment
, "%s %s", banner
, u
->banner
);