16 #include "playout/elo.h"
17 #include "playout/moggy.h"
18 #include "playout/light.h"
21 #include "distributed/distributed.h"
22 #include "uct/dynkomi.h"
23 #include "uct/internal.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_stop(struct uct
*u
);
34 static void uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
);
36 /* Maximal simulation length. */
37 #define MC_GAMELEN MAX_GAMELEN
41 setup_state(struct uct
*u
, struct board
*b
, enum stone color
)
43 u
->t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0, u
->local_tree_aging
);
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_book
&& 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
->dynkomi
->permove
)
66 u
->t
->extra_komi
= u
->dynkomi
->permove(u
->dynkomi
, b
, u
->t
);
70 uct_prepare_move(struct uct
*u
, struct board
*b
, enum stone color
)
73 /* Verify that we have sane state. */
75 assert(u
->t
&& b
->moves
);
76 if (color
!= stone_other(u
->t
->root_color
)) {
77 fprintf(stderr
, "Fatal: Non-alternating play detected %d %d\n",
78 color
, u
->t
->root_color
);
83 /* We need fresh state. */
85 setup_state(u
, b
, color
);
88 u
->ownermap
.playouts
= 0;
89 memset(u
->ownermap
.map
, 0, board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
90 memset(u
->stats
, 0, board_size2(b
) * sizeof(u
->stats
[0]));
91 u
->played_own
= u
->played_all
= 0;
95 dead_group_list(struct uct
*u
, struct board
*b
, struct move_queue
*mq
)
97 struct group_judgement gj
;
99 gj
.gs
= alloca(board_size2(b
) * sizeof(gj
.gs
[0]));
100 board_ownermap_judge_group(b
, &u
->ownermap
, &gj
);
101 groups_of_status(b
, &gj
, GS_DEAD
, mq
);
105 uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
)
107 if (u
->ownermap
.playouts
< GJ_MINGAMES
)
110 struct move_queue mq
= { .moves
= 0 };
111 dead_group_list(u
, b
, &mq
);
112 if (pass_all_alive
&& mq
.moves
> 0)
113 return false; // We need to remove some dead groups first.
114 return pass_is_safe(b
, color
, &mq
);
118 uct_printhook_ownermap(struct board
*board
, coord_t c
, char *s
, char *end
)
120 struct uct
*u
= board
->es
;
122 const char chr
[] = ":XO,"; // dame, black, white, unclear
123 const char chm
[] = ":xo,";
124 char ch
= chr
[board_ownermap_judge_point(&u
->ownermap
, c
, GJ_THRES
)];
125 if (ch
== ',') { // less precise estimate then?
126 ch
= chm
[board_ownermap_judge_point(&u
->ownermap
, c
, 0.67)];
128 s
+= snprintf(s
, end
- s
, "%c ", ch
);
133 uct_notify_play(struct engine
*e
, struct board
*b
, struct move
*m
)
135 struct uct
*u
= e
->data
;
137 /* No state, create one - this is probably game beginning
138 * and we need to load the opening book right now. */
139 uct_prepare_move(u
, b
, m
->color
);
143 /* Stop pondering, required by tree_promote_at() */
144 uct_pondering_stop(u
);
145 if (UDEBUGL(2) && u
->slave
)
146 tree_dump(u
->t
, u
->dumpthres
);
148 if (is_resign(m
->coord
)) {
154 /* Promote node of the appropriate move to the tree root. */
156 if (!tree_promote_at(u
->t
, b
, m
->coord
)) {
158 fprintf(stderr
, "Warning: Cannot promote move node! Several play commands in row?\n");
163 /* If we are a slave in a distributed engine, start pondering once
164 * we know which move we actually played. See uct_genmove() about
165 * the check for pass. */
166 if (u
->pondering_opt
&& u
->slave
&& m
->color
== u
->my_color
&& !is_pass(m
->coord
))
167 uct_pondering_start(u
, b
, u
->t
, stone_other(m
->color
));
173 uct_chat(struct engine
*e
, struct board
*b
, char *cmd
)
175 struct uct
*u
= e
->data
;
176 static char reply
[1024];
178 cmd
+= strspn(cmd
, " \n\t");
179 if (!strncasecmp(cmd
, "winrate", 7)) {
181 return "no game context (yet?)";
182 enum stone color
= u
->t
->root_color
;
183 struct tree_node
*n
= u
->t
->root
;
184 snprintf(reply
, 1024, "In %d playouts at %d threads, %s %s can win with %.2f%% probability",
185 n
->u
.playouts
, u
->threads
, stone2str(color
), coord2sstr(n
->coord
, b
),
186 tree_node_get_value(u
->t
, -1, n
->u
.value
) * 100);
187 if (u
->t
->use_extra_komi
&& abs(u
->t
->extra_komi
) >= 0.5) {
188 sprintf(reply
+ strlen(reply
), ", while self-imposing extra komi %.1f",
198 uct_dead_group_list(struct engine
*e
, struct board
*b
, struct move_queue
*mq
)
200 struct uct
*u
= e
->data
;
202 /* This means the game is probably over, no use pondering on. */
203 uct_pondering_stop(u
);
205 if (u
->pass_all_alive
)
206 return; // no dead groups
208 bool mock_state
= false;
211 /* No state, but we cannot just back out - we might
212 * have passed earlier, only assuming some stones are
213 * dead, and then re-connected, only to lose counting
214 * when all stones are assumed alive. */
215 /* Mock up some state and seed the ownermap by few
217 uct_prepare_move(u
, b
, S_BLACK
); assert(u
->t
);
218 for (int i
= 0; i
< GJ_MINGAMES
; i
++)
219 uct_playout(u
, b
, S_BLACK
, u
->t
);
223 dead_group_list(u
, b
, mq
);
226 /* Clean up the mock state in case we will receive
227 * a genmove; we could get a non-alternating-move
228 * error from uct_prepare_move() in that case otherwise. */
234 playout_policy_done(struct playout_policy
*p
)
236 if (p
->done
) p
->done(p
);
237 if (p
->data
) free(p
->data
);
242 uct_done(struct engine
*e
)
244 /* This is called on engine reset, especially when clear_board
245 * is received and new game should begin. */
246 struct uct
*u
= e
->data
;
247 uct_pondering_stop(u
);
248 if (u
->t
) reset_state(u
);
249 free(u
->ownermap
.map
);
253 free(u
->random_policy
);
254 playout_policy_done(u
->playout
);
255 uct_prior_done(u
->prior
);
260 /* Run time-limited MCTS search on foreground. */
262 uct_search(struct uct
*u
, struct board
*b
, struct time_info
*ti
, enum stone color
, struct tree
*t
)
264 struct uct_search_state s
;
265 uct_search_start(u
, b
, color
, t
, ti
, &s
);
266 if (UDEBUGL(2) && s
.base_playouts
> 0)
267 fprintf(stderr
, "<pre-simulated %d games>\n", s
.base_playouts
);
269 /* The search tree is ctx->t. This is normally == t, but in case of
270 * TM_ROOT, it is one of the trees belonging to the independent
271 * workers. It is important to reference ctx->t directly since the
272 * thread manager will swap the tree pointer asynchronously. */
273 /* XXX: This means TM_ROOT support is suboptimal since single stalled
274 * thread can stall the others in case of limiting the search by game
275 * count. However, TM_ROOT just does not deserve any more extra code
278 /* Now, just periodically poll the search tree. */
280 time_sleep(TREE_BUSYWAIT_INTERVAL
);
281 /* TREE_BUSYWAIT_INTERVAL should never be less than desired time, or the
282 * time control is broken. But if it happens to be less, we still search
283 * at least 100ms otherwise the move is completely random. */
285 int i
= uct_search_games(&s
);
286 /* Print notifications etc. */
287 uct_search_progress(u
, b
, color
, t
, ti
, &s
, i
);
288 /* Check if we should stop the search. */
289 if (uct_search_check_stop(u
, b
, color
, t
, ti
, &s
, i
))
293 struct uct_thread_ctx
*ctx
= uct_search_stop();
294 if (UDEBUGL(2)) tree_dump(t
, u
->dumpthres
);
296 fprintf(stderr
, "(avg score %f/%d value %f/%d)\n",
297 u
->dynkomi
->score
.value
, u
->dynkomi
->score
.playouts
,
298 u
->dynkomi
->value
.value
, u
->dynkomi
->value
.playouts
);
300 uct_progress_status(u
, t
, color
, ctx
->games
);
302 u
->played_own
+= ctx
->games
;
306 /* Start pondering background with @color to play. */
308 uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
)
311 fprintf(stderr
, "Starting to ponder with color %s\n", stone2str(stone_other(color
)));
314 /* We need a local board copy to ponder upon. */
315 struct board
*b
= malloc2(sizeof(*b
)); board_copy(b
, b0
);
317 /* *b0 did not have the genmove'd move played yet. */
318 struct move m
= { t
->root
->coord
, t
->root_color
};
319 int res
= board_play(b
, &m
);
321 setup_dynkomi(u
, b
, stone_other(m
.color
));
323 /* Start MCTS manager thread "headless". */
324 static struct uct_search_state s
;
325 uct_search_start(u
, b
, color
, t
, NULL
, &s
);
328 /* uct_search_stop() frontend for the pondering (non-genmove) mode, and
329 * to stop the background search for a slave in the distributed engine. */
331 uct_pondering_stop(struct uct
*u
)
333 if (!thread_manager_running
)
336 /* Stop the thread manager. */
337 struct uct_thread_ctx
*ctx
= uct_search_stop();
339 if (u
->pondering
) fprintf(stderr
, "(pondering) ");
340 uct_progress_status(u
, ctx
->t
, ctx
->color
, ctx
->games
);
344 u
->pondering
= false;
350 uct_genmove_setup(struct uct
*u
, struct board
*b
, enum stone color
)
352 if (b
->superko_violation
) {
353 fprintf(stderr
, "!!! WARNING: SUPERKO VIOLATION OCCURED BEFORE THIS MOVE\n");
354 fprintf(stderr
, "Maybe you play with situational instead of positional superko?\n");
355 fprintf(stderr
, "I'm going to ignore the violation, but note that I may miss\n");
356 fprintf(stderr
, "some moves valid under this ruleset because of this.\n");
357 b
->superko_violation
= false;
360 uct_prepare_move(u
, b
, color
);
365 /* How to decide whether to use dynkomi in this game? Since we use
366 * pondering, it's not simple "who-to-play" matter. Decide based on
367 * the last genmove issued. */
368 u
->t
->use_extra_komi
= !!(u
->dynkomi_mask
& color
);
369 setup_dynkomi(u
, b
, color
);
371 if (b
->rules
== RULES_JAPANESE
)
372 u
->territory_scoring
= true;
374 /* Make pessimistic assumption about komi for Japanese rules to
375 * avoid losing by 0.5 when winning by 0.5 with Chinese rules.
376 * The rules usually give the same winner if the integer part of komi
377 * is odd so we adjust the komi only if it is even (for a board of
378 * odd size). We are not trying to get an exact evaluation for rare
379 * cases of seki. For details see http://home.snafu.de/jasiek/parity.html */
380 if (u
->territory_scoring
&& (((int)floor(b
->komi
) + board_size(b
)) & 1)) {
381 b
->komi
+= (color
== S_BLACK
? 1.0 : -1.0);
383 fprintf(stderr
, "Setting komi to %.1f assuming Japanese rules\n",
389 uct_genmove(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
391 double start_time
= time_now();
392 struct uct
*u
= e
->data
;
393 uct_pondering_stop(u
);
394 uct_genmove_setup(u
, b
, color
);
396 /* Start the Monte Carlo Tree Search! */
397 int base_playouts
= u
->t
->root
->u
.playouts
;
398 int played_games
= uct_search(u
, b
, ti
, color
, u
->t
);
401 struct tree_node
*best
;
402 best
= uct_search_result(u
, b
, color
, pass_all_alive
, played_games
, base_playouts
, &best_coord
);
405 double time
= time_now() - start_time
+ 0.000001; /* avoid divide by zero */
406 fprintf(stderr
, "genmove in %0.2fs (%d games/s, %d games/s/thread)\n",
407 time
, (int)(played_games
/time
), (int)(played_games
/time
/u
->threads
));
411 /* Pass or resign. */
413 return coord_copy(best_coord
);
415 tree_promote_node(u
->t
, &best
);
417 /* After a pass, pondering is harmful for two reasons:
418 * (i) We might keep pondering even when the game is over.
419 * Of course this is the case for opponent resign as well.
420 * (ii) More importantly, the ownermap will get skewed since
421 * the UCT will start cutting off any playouts. */
422 if (u
->pondering_opt
&& !is_pass(best
->coord
)) {
423 uct_pondering_start(u
, b
, u
->t
, stone_other(color
));
425 return coord_copy(best_coord
);
430 uct_genbook(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
)
432 struct uct
*u
= e
->data
;
433 if (!u
->t
) uct_prepare_move(u
, b
, color
);
436 if (ti
->dim
== TD_GAMES
) {
437 /* Don't count in games that already went into the book. */
438 ti
->len
.games
+= u
->t
->root
->u
.playouts
;
440 uct_search(u
, b
, ti
, color
, u
->t
);
442 assert(ti
->dim
== TD_GAMES
);
443 tree_save(u
->t
, b
, ti
->len
.games
/ 100);
449 uct_dumpbook(struct engine
*e
, struct board
*b
, enum stone color
)
451 struct uct
*u
= e
->data
;
452 struct tree
*t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0, u
->local_tree_aging
);
460 uct_state_init(char *arg
, struct board
*b
)
462 struct uct
*u
= calloc2(1, sizeof(struct uct
));
463 bool using_elo
= false;
465 u
->debug_level
= debug_level
;
466 u
->gamelen
= MC_GAMELEN
;
470 u
->playout_amaf
= true;
471 u
->playout_amaf_nakade
= false;
472 u
->amaf_prior
= false;
473 u
->max_tree_size
= 3072ULL * 1048576;
475 u
->dynkomi_mask
= S_BLACK
;
478 u
->thread_model
= TM_TREEVL
;
479 u
->parallel_tree
= true;
480 u
->virtual_loss
= true;
482 u
->fuseki_end
= 20; // max time at 361*20% = 72 moves (our 36th move, still 99 to play)
483 u
->yose_start
= 40; // (100-40-25)*361/100/2 = 63 moves still to play by us then
484 u
->bestr_ratio
= 0.02;
485 // 2.5 is clearly too much, but seems to compensate well for overly stern time allocations.
486 // TODO: Further tuning and experiments with better time allocation schemes.
487 u
->best2_ratio
= 2.5;
489 u
->val_scale
= 0.04; u
->val_points
= 40;
492 u
->local_tree_aging
= 2;
495 char *optspec
, *next
= arg
;
498 next
+= strcspn(next
, ",");
499 if (*next
) { *next
++ = 0; } else { *next
= 0; }
501 char *optname
= optspec
;
502 char *optval
= strchr(optspec
, '=');
503 if (optval
) *optval
++ = 0;
505 if (!strcasecmp(optname
, "debug")) {
507 u
->debug_level
= atoi(optval
);
510 } else if (!strcasecmp(optname
, "mercy") && optval
) {
511 /* Minimal difference of black/white captures
512 * to stop playout - "Mercy Rule". Speeds up
513 * hopeless playouts at the expense of some
515 u
->mercymin
= atoi(optval
);
516 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
517 u
->gamelen
= atoi(optval
);
518 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
519 u
->expand_p
= atoi(optval
);
520 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
521 u
->dumpthres
= atoi(optval
);
522 } else if (!strcasecmp(optname
, "best2_ratio") && optval
) {
523 /* If set, prolong simulating while
524 * first_best/second_best playouts ratio
525 * is less than best2_ratio. */
526 u
->best2_ratio
= atof(optval
);
527 } else if (!strcasecmp(optname
, "bestr_ratio") && optval
) {
528 /* If set, prolong simulating while
529 * best,best_best_child values delta
530 * is more than bestr_ratio. */
531 u
->bestr_ratio
= atof(optval
);
532 } else if (!strcasecmp(optname
, "playout_amaf")) {
533 /* Whether to include random playout moves in
534 * AMAF as well. (Otherwise, only tree moves
535 * are included in AMAF. Of course makes sense
536 * only in connection with an AMAF policy.) */
537 /* with-without: 55.5% (+-4.1) */
538 if (optval
&& *optval
== '0')
539 u
->playout_amaf
= false;
541 u
->playout_amaf
= true;
542 } else if (!strcasecmp(optname
, "playout_amaf_nakade")) {
543 /* Whether to include nakade moves from playouts
544 * in the AMAF statistics; this tends to nullify
545 * the playout_amaf effect by adding too much
547 if (optval
&& *optval
== '0')
548 u
->playout_amaf_nakade
= false;
550 u
->playout_amaf_nakade
= true;
551 } else if (!strcasecmp(optname
, "playout_amaf_cutoff") && optval
) {
552 /* Keep only first N% of playout stage AMAF
554 u
->playout_amaf_cutoff
= atoi(optval
);
555 } else if ((!strcasecmp(optname
, "policy") || !strcasecmp(optname
, "random_policy")) && optval
) {
556 char *policyarg
= strchr(optval
, ':');
557 struct uct_policy
**p
= !strcasecmp(optname
, "policy") ? &u
->policy
: &u
->random_policy
;
560 if (!strcasecmp(optval
, "ucb1")) {
561 *p
= policy_ucb1_init(u
, policyarg
);
562 } else if (!strcasecmp(optval
, "ucb1amaf")) {
563 *p
= policy_ucb1amaf_init(u
, policyarg
);
565 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
568 } else if (!strcasecmp(optname
, "playout") && optval
) {
569 char *playoutarg
= strchr(optval
, ':');
572 if (!strcasecmp(optval
, "moggy")) {
573 u
->playout
= playout_moggy_init(playoutarg
, b
);
574 } else if (!strcasecmp(optval
, "light")) {
575 u
->playout
= playout_light_init(playoutarg
, b
);
576 } else if (!strcasecmp(optval
, "elo")) {
577 u
->playout
= playout_elo_init(playoutarg
, b
);
580 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
583 } else if (!strcasecmp(optname
, "prior") && optval
) {
584 u
->prior
= uct_prior_init(optval
, b
);
585 } else if (!strcasecmp(optname
, "amaf_prior") && optval
) {
586 u
->amaf_prior
= atoi(optval
);
587 } else if (!strcasecmp(optname
, "threads") && optval
) {
588 /* By default, Pachi will run with only single
589 * tree search thread! */
590 u
->threads
= atoi(optval
);
591 } else if (!strcasecmp(optname
, "thread_model") && optval
) {
592 if (!strcasecmp(optval
, "root")) {
593 /* Root parallelization - each thread
594 * does independent search, trees are
595 * merged at the end. */
596 u
->thread_model
= TM_ROOT
;
597 u
->parallel_tree
= false;
598 u
->virtual_loss
= false;
599 } else if (!strcasecmp(optval
, "tree")) {
600 /* Tree parallelization - all threads
601 * grind on the same tree. */
602 u
->thread_model
= TM_TREE
;
603 u
->parallel_tree
= true;
604 u
->virtual_loss
= false;
605 } else if (!strcasecmp(optval
, "treevl")) {
606 /* Tree parallelization, but also
607 * with virtual losses - this discou-
608 * rages most threads choosing the
609 * same tree branches to read. */
610 u
->thread_model
= TM_TREEVL
;
611 u
->parallel_tree
= true;
612 u
->virtual_loss
= true;
614 fprintf(stderr
, "UCT: Invalid thread model %s\n", optval
);
617 } else if (!strcasecmp(optname
, "pondering")) {
618 /* Keep searching even during opponent's turn. */
619 u
->pondering_opt
= !optval
|| atoi(optval
);
620 } else if (!strcasecmp(optname
, "fuseki_end") && optval
) {
621 /* At the very beginning it's not worth thinking
622 * too long because the playout evaluations are
623 * very noisy. So gradually increase the thinking
624 * time up to maximum when fuseki_end percent
625 * of the board has been played.
626 * This only applies if we are not in byoyomi. */
627 u
->fuseki_end
= atoi(optval
);
628 } else if (!strcasecmp(optname
, "yose_start") && optval
) {
629 /* When yose_start percent of the board has been
630 * played, or if we are in byoyomi, stop spending
631 * more time and spread the remaining time
633 * Between fuseki_end and yose_start, we spend
634 * a constant proportion of the remaining time
635 * on each move. (yose_start should actually
636 * be much earlier than when real yose start,
637 * but "yose" is a good short name to convey
639 u
->yose_start
= atoi(optval
);
640 } else if (!strcasecmp(optname
, "force_seed") && optval
) {
641 u
->force_seed
= atoi(optval
);
642 } else if (!strcasecmp(optname
, "no_book")) {
644 } else if (!strcasecmp(optname
, "dynkomi") && optval
) {
645 /* Dynamic komi approach; there are multiple
646 * ways to adjust komi dynamically throughout
647 * play. We currently support two: */
648 char *dynkomiarg
= strchr(optval
, ':');
651 if (!strcasecmp(optval
, "none")) {
652 u
->dynkomi
= uct_dynkomi_init_none(u
, dynkomiarg
, b
);
653 } else if (!strcasecmp(optval
, "linear")) {
654 u
->dynkomi
= uct_dynkomi_init_linear(u
, dynkomiarg
, b
);
655 } else if (!strcasecmp(optval
, "adaptive")) {
656 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, dynkomiarg
, b
);
658 fprintf(stderr
, "UCT: Invalid dynkomi mode %s\n", optval
);
661 } else if (!strcasecmp(optname
, "dynkomi_mask") && optval
) {
662 /* Bitmask of colors the player must be
663 * for dynkomi be applied; you may want
664 * to use dynkomi_mask=3 to allow dynkomi
665 * even in games where Pachi is white. */
666 u
->dynkomi_mask
= atoi(optval
);
667 } else if (!strcasecmp(optname
, "dynkomi_interval") && optval
) {
668 /* If non-zero, re-adjust dynamic komi
669 * throughout a single genmove reading,
670 * roughly every N simulations. */
671 /* XXX: Does not work with tree
672 * parallelization. */
673 u
->dynkomi_interval
= atoi(optval
);
674 } else if (!strcasecmp(optname
, "val_scale") && optval
) {
675 /* How much of the game result value should be
676 * influenced by win size. Zero means it isn't. */
677 u
->val_scale
= atof(optval
);
678 } else if (!strcasecmp(optname
, "val_points") && optval
) {
679 /* Maximum size of win to be scaled into game
680 * result value. Zero means boardsize^2. */
681 u
->val_points
= atoi(optval
) * 2; // result values are doubled
682 } else if (!strcasecmp(optname
, "val_extra")) {
683 /* If false, the score coefficient will be simply
684 * added to the value, instead of scaling the result
685 * coefficient because of it. */
686 u
->val_extra
= !optval
|| atoi(optval
);
687 } else if (!strcasecmp(optname
, "local_tree") && optval
) {
688 /* Whether to bias exploration by local tree values
689 * (must be supported by the used policy).
691 * 1: Do, value = result.
692 * Try to temper the result:
693 * 2: Do, value = 0.5+(result-expected)/2.
694 * 3: Do, value = 0.5+bzz((result-expected)^2).
695 * 4: Do, value = 0.5+sqrt(result-expected)/2. */
696 u
->local_tree
= atoi(optval
);
697 } else if (!strcasecmp(optname
, "tenuki_d") && optval
) {
698 /* Tenuki distance at which to break the local tree. */
699 u
->tenuki_d
= atoi(optval
);
700 if (u
->tenuki_d
> TREE_NODE_D_MAX
+ 1) {
701 fprintf(stderr
, "uct: tenuki_d must not be larger than TREE_NODE_D_MAX+1 %d\n", TREE_NODE_D_MAX
+ 1);
704 } else if (!strcasecmp(optname
, "local_tree_aging") && optval
) {
705 /* How much to reduce local tree values between moves. */
706 u
->local_tree_aging
= atof(optval
);
707 } else if (!strcasecmp(optname
, "local_tree_allseq")) {
708 /* By default, only complete sequences are stored
709 * in the local tree. If this is on, also
710 * subsequences starting at each move are stored. */
711 u
->local_tree_allseq
= !optval
|| atoi(optval
);
712 } else if (!strcasecmp(optname
, "local_tree_playout")) {
713 /* Whether to adjust ELO playout probability
714 * distributions according to matched localtree
716 u
->local_tree_playout
= !optval
|| atoi(optval
);
717 } else if (!strcasecmp(optname
, "local_tree_pseqroot")) {
718 /* By default, when we have no sequence move
719 * to suggest in-playout, we give up. If this
720 * is on, we make probability distribution from
721 * sequences first moves instead. */
722 u
->local_tree_pseqroot
= !optval
|| atoi(optval
);
723 } else if (!strcasecmp(optname
, "pass_all_alive")) {
724 /* Whether to consider passing only after all
725 * dead groups were removed from the board;
726 * this is like all genmoves are in fact
727 * kgs-genmove_cleanup. */
728 u
->pass_all_alive
= !optval
|| atoi(optval
);
729 } else if (!strcasecmp(optname
, "territory_scoring")) {
730 /* Use territory scoring (default is area scoring).
731 * An explicit kgs-rules command overrides this. */
732 u
->territory_scoring
= !optval
|| atoi(optval
);
733 } else if (!strcasecmp(optname
, "random_policy_chance") && optval
) {
734 /* If specified (N), with probability 1/N, random_policy policy
735 * descend is used instead of main policy descend; useful
736 * if specified policy (e.g. UCB1AMAF) can make unduly biased
737 * choices sometimes, you can fall back to e.g.
738 * random_policy=UCB1. */
739 u
->random_policy_chance
= atoi(optval
);
740 } else if (!strcasecmp(optname
, "max_tree_size") && optval
) {
741 /* Maximum amount of memory [MiB] consumed by the move tree.
742 * For fast_alloc it includes the temp tree used for pruning.
743 * Default is 3072 (3 GiB). Note that if you use TM_ROOT,
744 * this limits size of only one of the trees, not all of them
746 u
->max_tree_size
= atol(optval
) * 1048576;
747 } else if (!strcasecmp(optname
, "fast_alloc")) {
748 u
->fast_alloc
= !optval
|| atoi(optval
);
749 } else if (!strcasecmp(optname
, "slave")) {
750 /* Act as slave for the distributed engine. */
751 u
->slave
= !optval
|| atoi(optval
);
752 } else if (!strcasecmp(optname
, "banner") && optval
) {
753 /* Additional banner string. This must come as the
754 * last engine parameter. */
755 if (*next
) *--next
= ',';
756 u
->banner
= strdup(optval
);
759 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
765 u
->resign_ratio
= 0.2; /* Resign when most games are lost. */
766 u
->loss_threshold
= 0.85; /* Stop reading if after at least 2000 playouts this is best value. */
768 u
->policy
= policy_ucb1amaf_init(u
, NULL
);
770 if (!!u
->random_policy_chance
^ !!u
->random_policy
) {
771 fprintf(stderr
, "uct: Only one of random_policy and random_policy_chance is set\n");
775 if (!u
->local_tree
) {
776 /* No ltree aging. */
777 u
->local_tree_aging
= 1.0f
;
780 u
->local_tree_playout
= false;
782 if (u
->fast_alloc
&& !u
->parallel_tree
) {
783 fprintf(stderr
, "fast_alloc not supported with root parallelization.\n");
787 u
->max_tree_size
= (100ULL * u
->max_tree_size
) / (100 + MIN_FREE_MEM_PERCENT
);
790 u
->prior
= uct_prior_init(NULL
, b
);
793 u
->playout
= playout_moggy_init(NULL
, b
);
794 u
->playout
->debug_level
= u
->debug_level
;
796 u
->ownermap
.map
= malloc2(board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
797 u
->stats
= malloc2(board_size2(b
) * sizeof(u
->stats
[0]));
800 u
->dynkomi
= uct_dynkomi_init_linear(u
, NULL
, b
);
802 /* Some things remain uninitialized for now - the opening book
803 * is not loaded and the tree not set up. */
804 /* This will be initialized in setup_state() at the first move
805 * received/requested. This is because right now we are not aware
806 * about any komi or handicap setup and such. */
812 engine_uct_init(char *arg
, struct board
*b
)
814 struct uct
*u
= uct_state_init(arg
, b
);
815 struct engine
*e
= calloc2(1, sizeof(struct engine
));
816 e
->name
= "UCT Engine";
817 e
->printhook
= uct_printhook_ownermap
;
818 e
->notify_play
= uct_notify_play
;
820 e
->genmove
= uct_genmove
;
821 e
->genmoves
= uct_genmoves
;
822 e
->dead_group_list
= uct_dead_group_list
;
826 e
->notify
= uct_notify
;
828 const char banner
[] = "I'm playing UCT. When I'm losing, I will resign, "
829 "if I think I win, I play until you pass. "
830 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
831 if (!u
->banner
) u
->banner
= "";
832 e
->comment
= malloc2(sizeof(banner
) + strlen(u
->banner
) + 1);
833 sprintf(e
->comment
, "%s %s", banner
, u
->banner
);