17 #include "playout/elo.h"
18 #include "playout/moggy.h"
19 #include "playout/light.h"
23 #include "uct/internal.h"
24 #include "uct/prior.h"
29 struct uct_policy
*policy_ucb1_init(struct uct
*u
, char *arg
);
30 struct uct_policy
*policy_ucb1amaf_init(struct uct
*u
, char *arg
);
31 static void uct_pondering_stop(struct uct
*u
);
34 /* Default number of simulations to perform per move.
35 * Note that this is now in total over all threads! (Unless TM_ROOT.) */
36 #define MC_GAMES 80000
37 #define MC_GAMELEN MAX_GAMELEN
38 static const struct time_info default_ti
= {
41 .len
= { .games
= MC_GAMES
},
44 /* How big proportion of ownermap counts must be of one color to consider
47 /* How many games to consider at minimum before judging groups. */
48 #define GJ_MINGAMES 500
50 /* How often to inspect the tree from the main thread to check for playout
51 * stop, progress reports, etc. (in seconds) */
52 #define TREE_BUSYWAIT_INTERVAL 0.1 /* 100ms */
54 /* Once per how many simulations (per thread) to show a progress report line. */
55 #define TREE_SIMPROGRESS_INTERVAL 10000
57 /* When terminating uct_search() early, the safety margin to add to the
58 * remaining playout number estimate when deciding whether the result can
60 #define PLAYOUT_DELTA_SAFEMARGIN 1000
64 setup_state(struct uct
*u
, struct board
*b
, enum stone color
)
66 u
->t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0);
68 fast_srandom(u
->force_seed
);
70 fprintf(stderr
, "Fresh board with random seed %lu\n", fast_getseed());
71 //board_print(b, stderr);
72 if (!u
->no_book
&& b
->moves
== 0) {
73 assert(color
== S_BLACK
);
79 reset_state(struct uct
*u
)
82 tree_done(u
->t
); u
->t
= NULL
;
86 setup_dynkomi(struct uct
*u
, struct board
*b
, enum stone to_play
)
88 if (u
->dynkomi
> b
->moves
&& u
->t
->use_extra_komi
)
89 u
->t
->extra_komi
= uct_get_extra_komi(u
, b
);
95 prepare_move(struct engine
*e
, struct board
*b
, enum stone color
)
97 struct uct
*u
= e
->data
;
100 /* Verify that we have sane state. */
102 assert(u
->t
&& b
->moves
);
103 if (color
!= stone_other(u
->t
->root_color
)) {
104 fprintf(stderr
, "Fatal: Non-alternating play detected %d %d\n",
105 color
, u
->t
->root_color
);
110 /* We need fresh state. */
112 setup_state(u
, b
, color
);
115 u
->ownermap
.playouts
= 0;
116 memset(u
->ownermap
.map
, 0, board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
120 dead_group_list(struct uct
*u
, struct board
*b
, struct move_queue
*mq
)
122 struct group_judgement gj
;
124 gj
.gs
= alloca(board_size2(b
) * sizeof(gj
.gs
[0]));
125 board_ownermap_judge_group(b
, &u
->ownermap
, &gj
);
126 groups_of_status(b
, &gj
, GS_DEAD
, mq
);
130 uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
)
132 if (u
->ownermap
.playouts
< GJ_MINGAMES
)
135 struct move_queue mq
= { .moves
= 0 };
137 dead_group_list(u
, b
, &mq
);
138 return pass_is_safe(b
, color
, &mq
);
143 uct_printhook_ownermap(struct board
*board
, coord_t c
, FILE *f
)
145 struct uct
*u
= board
->es
;
147 const char chr
[] = ":XO,"; // dame, black, white, unclear
148 const char chm
[] = ":xo,";
149 char ch
= chr
[board_ownermap_judge_point(&u
->ownermap
, c
, GJ_THRES
)];
150 if (ch
== ',') { // less precise estimate then?
151 ch
= chm
[board_ownermap_judge_point(&u
->ownermap
, c
, 0.67)];
153 fprintf(f
, "%c ", ch
);
157 uct_notify_play(struct engine
*e
, struct board
*b
, struct move
*m
)
159 struct uct
*u
= e
->data
;
161 /* No state, create one - this is probably game beginning
162 * and we need to load the opening book right now. */
163 prepare_move(e
, b
, m
->color
);
167 /* Stop pondering. */
168 /* XXX: If we are about to receive multiple 'play' commands,
169 * e.g. in a rengo, we will not ponder during the rest of them. */
170 uct_pondering_stop(u
);
172 if (is_resign(m
->coord
)) {
178 /* Promote node of the appropriate move to the tree root. */
180 if (!tree_promote_at(u
->t
, b
, m
->coord
)) {
182 fprintf(stderr
, "Warning: Cannot promote move node! Several play commands in row?\n");
186 /* Setting up dynkomi is not necessary here, probably, but we
187 * better do it anyway for consistency reasons. */
188 setup_dynkomi(u
, b
, stone_other(m
->color
));
193 uct_chat(struct engine
*e
, struct board
*b
, char *cmd
)
195 struct uct
*u
= e
->data
;
196 static char reply
[1024];
198 cmd
+= strspn(cmd
, " \n\t");
199 if (!strncasecmp(cmd
, "winrate", 7)) {
201 return "no game context (yet?)";
202 enum stone color
= u
->t
->root_color
;
203 struct tree_node
*n
= u
->t
->root
;
204 snprintf(reply
, 1024, "In %d playouts at %d threads, %s %s can win with %.2f%% probability",
205 n
->u
.playouts
, u
->threads
, stone2str(color
), coord2sstr(n
->coord
, b
),
206 tree_node_get_value(u
->t
, -1, n
->u
.value
) * 100);
207 if (u
->t
->use_extra_komi
&& abs(u
->t
->extra_komi
) >= 0.5) {
208 sprintf(reply
+ strlen(reply
), ", while self-imposing extra komi %.1f",
218 uct_dead_group_list(struct engine
*e
, struct board
*b
, struct move_queue
*mq
)
220 struct uct
*u
= e
->data
;
222 /* This means the game is probably over, no use pondering on. */
223 uct_pondering_stop(u
);
225 if (u
->pass_all_alive
)
226 return; // no dead groups
228 bool mock_state
= false;
231 /* No state, but we cannot just back out - we might
232 * have passed earlier, only assuming some stones are
233 * dead, and then re-connected, only to lose counting
234 * when all stones are assumed alive. */
235 /* Mock up some state and seed the ownermap by few
237 prepare_move(e
, b
, S_BLACK
); assert(u
->t
);
238 for (int i
= 0; i
< GJ_MINGAMES
; i
++)
239 uct_playout(u
, b
, S_BLACK
, u
->t
);
243 dead_group_list(u
, b
, mq
);
246 /* Clean up the mock state in case we will receive
247 * a genmove; we could get a non-alternating-move
248 * error from prepare_move() in that case otherwise. */
254 playout_policy_done(struct playout_policy
*p
)
256 if (p
->done
) p
->done(p
);
257 if (p
->data
) free(p
->data
);
262 uct_done(struct engine
*e
)
264 /* This is called on engine reset, especially when clear_board
265 * is received and new game should begin. */
266 struct uct
*u
= e
->data
;
267 uct_pondering_stop(u
);
268 if (u
->t
) reset_state(u
);
269 free(u
->ownermap
.map
);
272 free(u
->random_policy
);
273 playout_policy_done(u
->playout
);
274 uct_prior_done(u
->prior
);
278 /* Pachi threading structure (if uct_playouts_parallel() is used):
281 * | main(), GTP communication, ...
282 * | starts and stops the search managed by thread_manager
285 * | spawns and collects worker threads
291 * uct_playouts() loop, doing descend-playout until uct_halt
293 * Another way to look at it is by functions (lines denote thread boundaries):
296 * | uct_search() (uct_search_start() .. uct_search_stop())
297 * | -----------------------
298 * | spawn_thread_manager()
299 * | -----------------------
301 * V uct_playouts() */
303 /* Set in thread manager in case the workers should stop. */
304 volatile sig_atomic_t uct_halt
= 0;
305 /* ID of the running worker thread. */
306 __thread
int thread_id
= -1;
307 /* ID of the thread manager. */
308 static pthread_t thread_manager
;
309 static bool thread_manager_running
;
311 static pthread_mutex_t finish_mutex
= PTHREAD_MUTEX_INITIALIZER
;
312 static pthread_cond_t finish_cond
= PTHREAD_COND_INITIALIZER
;
313 static volatile int finish_thread
;
314 static pthread_mutex_t finish_serializer
= PTHREAD_MUTEX_INITIALIZER
;
327 spawn_worker(void *ctx_
)
329 struct spawn_ctx
*ctx
= ctx_
;
331 fast_srandom(ctx
->seed
);
332 thread_id
= ctx
->tid
;
334 ctx
->games
= uct_playouts(ctx
->u
, ctx
->b
, ctx
->color
, ctx
->t
);
336 pthread_mutex_lock(&finish_serializer
);
337 pthread_mutex_lock(&finish_mutex
);
338 finish_thread
= ctx
->tid
;
339 pthread_cond_signal(&finish_cond
);
340 pthread_mutex_unlock(&finish_mutex
);
344 /* Thread manager, controlling worker threads. It must be called with
345 * finish_mutex lock held, but it will unlock it itself before exiting;
346 * this is necessary to be completely deadlock-free. */
347 /* The finish_cond can be signalled for it to stop; in that case,
348 * the caller should set finish_thread = -1. */
349 /* After it is started, it will update mctx->t to point at some tree
350 * used for the actual search (matters only for TM_ROOT), on return
351 * it will set mctx->games to the number of performed simulations. */
353 spawn_thread_manager(void *ctx_
)
355 /* In thread_manager, we use only some of the ctx fields. */
356 struct spawn_ctx
*mctx
= ctx_
;
357 struct uct
*u
= mctx
->u
;
358 struct tree
*t
= mctx
->t
;
359 bool shared_tree
= u
->parallel_tree
;
360 fast_srandom(mctx
->seed
);
362 int played_games
= 0;
363 pthread_t threads
[u
->threads
];
368 /* Spawn threads... */
369 for (int ti
= 0; ti
< u
->threads
; ti
++) {
370 struct spawn_ctx
*ctx
= malloc(sizeof(*ctx
));
371 ctx
->u
= u
; ctx
->b
= mctx
->b
; ctx
->color
= mctx
->color
;
372 mctx
->t
= ctx
->t
= shared_tree
? t
: tree_copy(t
);
373 ctx
->tid
= ti
; ctx
->seed
= fast_random(65536) + ti
;
374 pthread_create(&threads
[ti
], NULL
, spawn_worker
, ctx
);
376 fprintf(stderr
, "Spawned worker %d\n", ti
);
379 /* ...and collect them back: */
380 while (joined
< u
->threads
) {
381 /* Wait for some thread to finish... */
382 pthread_cond_wait(&finish_cond
, &finish_mutex
);
383 if (finish_thread
< 0) {
384 /* Stop-by-caller. Tell the workers to wrap up. */
388 /* ...and gather its remnants. */
389 struct spawn_ctx
*ctx
;
390 pthread_join(threads
[finish_thread
], (void **) &ctx
);
391 played_games
+= ctx
->games
;
394 if (ctx
->t
== mctx
->t
) mctx
->t
= t
;
395 tree_merge(t
, ctx
->t
);
400 fprintf(stderr
, "Joined worker %d\n", finish_thread
);
401 pthread_mutex_unlock(&finish_serializer
);
404 pthread_mutex_unlock(&finish_mutex
);
407 tree_normalize(mctx
->t
, u
->threads
);
409 mctx
->games
= played_games
;
413 static struct spawn_ctx
*
414 uct_search_start(struct uct
*u
, struct board
*b
, enum stone color
, struct tree
*t
)
416 assert(u
->threads
> 0);
417 assert(!thread_manager_running
);
419 struct spawn_ctx ctx
= { .u
= u
, .b
= b
, .color
= color
, .t
= t
, .seed
= fast_random(65536) };
420 static struct spawn_ctx mctx
; mctx
= ctx
;
421 pthread_mutex_lock(&finish_mutex
);
422 pthread_create(&thread_manager
, NULL
, spawn_thread_manager
, &mctx
);
423 thread_manager_running
= true;
427 static struct spawn_ctx
*
428 uct_search_stop(void)
430 assert(thread_manager_running
);
432 /* Signal thread manager to stop the workers. */
433 pthread_mutex_lock(&finish_mutex
);
435 pthread_cond_signal(&finish_cond
);
436 pthread_mutex_unlock(&finish_mutex
);
438 /* Collect the thread manager. */
439 struct spawn_ctx
*pctx
;
440 thread_manager_running
= false;
441 pthread_join(thread_manager
, (void **) &pctx
);
446 /* Run time-limited MCTS search on foreground. */
448 uct_search(struct uct
*u
, struct board
*b
, struct time_info
*ti
, enum stone color
, struct tree
*t
)
450 int base_playouts
= u
->t
->root
->u
.playouts
;
451 if (UDEBUGL(2) && base_playouts
> 0)
452 fprintf(stderr
, "<pre-simulated %d games skipped>\n", base_playouts
);
454 /* Set up time conditions. */
455 if (ti
->period
== TT_NULL
) *ti
= default_ti
;
456 struct time_stop stop
;
457 time_stop_conditions(ti
, b
, u
->fuseki_end
, u
->yose_start
, &stop
);
459 /* Number of last game with progress print. */
460 int last_print
= t
->root
->u
.playouts
;
461 /* Number of simulations to wait before next print. */
462 int print_interval
= TREE_SIMPROGRESS_INTERVAL
* (u
->thread_model
== TM_ROOT
? 1 : u
->threads
);
463 /* Printed notification about full memory? */
464 bool print_fullmem
= false;
466 struct spawn_ctx
*ctx
= uct_search_start(u
, b
, color
, t
);
468 /* The search tree is ctx->t. This is normally == t, but in case of
469 * TM_ROOT, it is one of the trees belonging to the independent
470 * workers. It is important to reference ctx->t directly since the
471 * thread manager will swap the tree pointer asynchronously. */
472 /* XXX: This means TM_ROOT support is suboptimal since single stalled
473 * thread can stall the others in case of limiting the search by game
474 * count. However, TM_ROOT just does not deserve any more extra code
477 struct tree_node
*best
= NULL
, *prev_best
;
478 struct tree_node
*winner
= NULL
, *prev_winner
;
480 double busywait_interval
= TREE_BUSYWAIT_INTERVAL
;
482 /* Now, just periodically poll the search tree. */
484 time_sleep(busywait_interval
);
485 /* busywait_interval should never be less than desired time, or the
486 * time control is broken. But if it happens to be less, we still search
487 * at least 100ms otherwise the move is completely random. */
489 int i
= ctx
->t
->root
->u
.playouts
;
491 /* Print progress? */
492 if (i
- last_print
> print_interval
) {
493 last_print
+= print_interval
; // keep the numbers tidy
494 uct_progress_status(u
, ctx
->t
, color
, last_print
);
496 if (!print_fullmem
&& ctx
->t
->nodes_size
> u
->max_tree_size
) {
498 fprintf(stderr
, "memory limit hit (%ld > %lu)\n", ctx
->t
->nodes_size
, u
->max_tree_size
);
499 print_fullmem
= true;
502 /* Check against time settings. */
503 bool desired_done
= false;
504 if (ti
->dim
== TD_WALLTIME
) {
505 double elapsed
= time_now() - ti
->len
.t
.timer_start
;
506 if (elapsed
> stop
.worst
.time
) break;
507 desired_done
= elapsed
> stop
.desired
.time
;
509 } else { assert(ti
->dim
== TD_GAMES
);
510 if (i
> stop
.worst
.playouts
) break;
511 desired_done
= i
> stop
.desired
.playouts
;
515 best
= u
->policy
->choose(u
->policy
, ctx
->t
->root
, b
, color
, resign
);
516 /* Second-best move. */
517 struct tree_node
*best2
= u
->policy
->choose(u
->policy
, ctx
->t
->root
, b
, color
, best
->coord
);
519 /* Early break in won situation. */
520 if (best
&& ((best
->u
.playouts
>= 2000 && tree_node_get_value(ctx
->t
, 1, best
->u
.value
) >= u
->loss_threshold
)
521 || (best
->u
.playouts
>= 500 && tree_node_get_value(ctx
->t
, 1, best
->u
.value
) >= 0.95)))
523 /* Break early if we estimate the second-best move cannot
524 * catch up in assigned time anymore. We use all our time
525 * if we are in byoyomi with single stone remaining in our
526 * period, however. */
527 if (best
&& best2
&& ti
->dim
== TD_WALLTIME
528 && (ti
->len
.t
.main_time
> 0 || ti
->len
.t
.byoyomi_stones
> 1)) {
529 double elapsed
= time_now() - ti
->len
.t
.timer_start
;
530 double remaining
= stop
.desired
.time
- elapsed
;
531 double pps
= ((double)i
- base_playouts
) / elapsed
;
532 double estplayouts
= remaining
* pps
+ PLAYOUT_DELTA_SAFEMARGIN
;
533 if (best
->u
.playouts
> best2
->u
.playouts
+ estplayouts
) {
535 fprintf(stderr
, "Early stop, result cannot change: "
536 "best %d, best2 %d, estimated %f simulations to go\n",
537 best
->u
.playouts
, best2
->u
.playouts
, estplayouts
);
542 /* We want to stop simulating, but are willing to keep trying
543 * if we aren't completely sure about the winner yet. */
547 fprintf(stderr
, "Did not find best move, still trying...\n");
551 if (u
->best2_ratio
> 0) {
552 /* Check best/best2 simulations ratio. If the
553 * two best moves give very similar results,
554 * keep simulating. */
555 if (best2
&& best2
->u
.playouts
556 && (double)best
->u
.playouts
/ best2
->u
.playouts
< u
->best2_ratio
) {
558 fprintf(stderr
, "Best2 ratio %f < threshold %f\n",
559 (double)best
->u
.playouts
/ best2
->u
.playouts
,
565 if (u
->policy
->winner
&& u
->policy
->evaluate
) {
566 prev_winner
= winner
;
567 winner
= u
->policy
->winner(u
->policy
, ctx
->t
, ctx
->t
->root
);
568 if (winner
&& winner
!= best
) {
569 /* Keep simulating if best explored
570 * does not have also highest value. */
571 if (UDEBUGL(2) && (best
!= prev_best
|| winner
!= prev_winner
)) {
572 fprintf(stderr
, "[%d] best %3s [%d] %f != winner %3s [%d] %f\n", i
,
573 coord2sstr(best
->coord
, ctx
->t
->board
),
574 best
->u
.playouts
, tree_node_get_value(ctx
->t
, 1, best
->u
.value
),
575 coord2sstr(winner
->coord
, ctx
->t
->board
),
576 winner
->u
.playouts
, tree_node_get_value(ctx
->t
, 1, winner
->u
.value
));
582 /* No reason to keep simulating, bye. */
586 /* TODO: Early break if best->variance goes under threshold and we already
587 * have enough playouts (possibly thanks to book or to pondering)? */
590 ctx
= uct_search_stop();
593 tree_dump(t
, u
->dumpthres
);
595 uct_progress_status(u
, t
, color
, ctx
->games
);
601 /* Start pondering background with @color to play. */
603 uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
)
606 fprintf(stderr
, "Starting to ponder with color %s\n", stone2str(stone_other(color
)));
609 /* We need a local board copy to ponder upon. */
610 struct board
*b
= malloc(sizeof(*b
)); board_copy(b
, b0
);
612 /* *b0 did not have the genmove'd move played yet. */
613 struct move m
= { t
->root
->coord
, t
->root_color
};
614 int res
= board_play(b
, &m
);
616 setup_dynkomi(u
, b
, stone_other(m
.color
));
618 /* Start MCTS manager thread "headless". */
619 uct_search_start(u
, b
, color
, t
);
622 /* uct_search_stop() frontend for the pondering (non-genmove) mode. */
624 uct_pondering_stop(struct uct
*u
)
626 u
->pondering
= false;
627 if (!thread_manager_running
)
630 /* Stop the thread manager. */
631 struct spawn_ctx
*ctx
= uct_search_stop();
633 fprintf(stderr
, "(pondering) ");
634 uct_progress_status(u
, ctx
->t
, ctx
->color
, ctx
->games
);
641 uct_genmove(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
643 double start_time
= time_now();
644 struct uct
*u
= e
->data
;
646 if (b
->superko_violation
) {
647 fprintf(stderr
, "!!! WARNING: SUPERKO VIOLATION OCCURED BEFORE THIS MOVE\n");
648 fprintf(stderr
, "Maybe you play with situational instead of positional superko?\n");
649 fprintf(stderr
, "I'm going to ignore the violation, but note that I may miss\n");
650 fprintf(stderr
, "some moves valid under this ruleset because of this.\n");
651 b
->superko_violation
= false;
655 uct_pondering_stop(u
);
656 prepare_move(e
, b
, color
);
659 /* How to decide whether to use dynkomi in this game? Since we use
660 * pondering, it's not simple "who-to-play" matter. Decide based on
661 * the last genmove issued. */
662 u
->t
->use_extra_komi
= !!(u
->dynkomi_mask
& color
);
663 setup_dynkomi(u
, b
, color
);
665 /* Perform the Monte Carlo Tree Search! */
666 int played_games
= uct_search(u
, b
, ti
, color
, u
->t
);
668 /* Choose the best move from the tree. */
669 struct tree_node
*best
= u
->policy
->choose(u
->policy
, u
->t
->root
, b
, color
, resign
);
672 return coord_copy(pass
);
675 fprintf(stderr
, "*** WINNER is %s (%d,%d) with score %1.4f (%d/%d:%d games)\n",
676 coord2sstr(best
->coord
, b
), coord_x(best
->coord
, b
), coord_y(best
->coord
, b
),
677 tree_node_get_value(u
->t
, 1, best
->u
.value
),
678 best
->u
.playouts
, u
->t
->root
->u
.playouts
, played_games
);
680 /* Do not resign if we're so short of time that evaluation of best move is completely
681 * unreliable, we might be winning actually. In this case best is almost random but
682 * still better than resign. */
683 if (tree_node_get_value(u
->t
, 1, best
->u
.value
) < u
->resign_ratio
&& !is_pass(best
->coord
)
684 && best
->u
.playouts
> GJ_MINGAMES
) {
686 return coord_copy(resign
);
689 /* If the opponent just passed and we win counting, always
691 if (b
->moves
> 1 && is_pass(b
->last_move
.coord
)) {
692 /* Make sure enough playouts are simulated. */
693 while (u
->ownermap
.playouts
< GJ_MINGAMES
)
694 uct_playout(u
, b
, color
, u
->t
);
695 if (uct_pass_is_safe(u
, b
, color
, u
->pass_all_alive
|| pass_all_alive
)) {
697 fprintf(stderr
, "<Will rather pass, looks safe enough.>\n");
702 tree_promote_node(u
->t
, &best
);
703 /* After a pass, pondering is harmful for two reasons:
704 * (i) We might keep pondering even when the game is over.
705 * Of course this is the case for opponent resign as well.
706 * (ii) More importantly, the ownermap will get skewed since
707 * the UCT will start cutting off any playouts. */
708 if (u
->pondering_opt
&& !is_pass(best
->coord
)) {
709 uct_pondering_start(u
, b
, u
->t
, stone_other(color
));
712 double time
= time_now() - start_time
+ 0.000001; /* avoid divide by zero */
713 fprintf(stderr
, "genmove in %0.2fs (%d games/s, %d games/s/thread)\n",
714 time
, (int)(played_games
/time
), (int)(played_games
/time
/u
->threads
));
716 return coord_copy(best
->coord
);
721 uct_genbook(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
)
723 struct uct
*u
= e
->data
;
724 if (!u
->t
) prepare_move(e
, b
, color
);
727 if (ti
->dim
== TD_GAMES
) {
728 /* Don't count in games that already went into the book. */
729 ti
->len
.games
+= u
->t
->root
->u
.playouts
;
731 uct_search(u
, b
, ti
, color
, u
->t
);
733 assert(ti
->dim
== TD_GAMES
);
734 tree_save(u
->t
, b
, ti
->len
.games
/ 100);
740 uct_dumpbook(struct engine
*e
, struct board
*b
, enum stone color
)
742 struct uct
*u
= e
->data
;
743 struct tree
*t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0);
751 uct_state_init(char *arg
, struct board
*b
)
753 struct uct
*u
= calloc(1, sizeof(struct uct
));
756 u
->gamelen
= MC_GAMELEN
;
760 u
->playout_amaf
= true;
761 u
->playout_amaf_nakade
= false;
762 u
->amaf_prior
= false;
763 u
->max_tree_size
= 3072ULL * 1048576;
765 if (board_size(b
) - 2 >= 19)
767 u
->dynkomi_mask
= S_BLACK
;
770 u
->thread_model
= TM_TREEVL
;
771 u
->parallel_tree
= true;
772 u
->virtual_loss
= true;
773 u
->fuseki_end
= 20; // max time at 361*20% = 72 moves (our 36th move, still 99 to play)
774 u
->yose_start
= 40; // (100-40-25)*361/100/2 = 63 moves still to play by us then
776 u
->val_scale
= 0.04; u
->val_points
= 40;
779 char *optspec
, *next
= arg
;
782 next
+= strcspn(next
, ",");
783 if (*next
) { *next
++ = 0; } else { *next
= 0; }
785 char *optname
= optspec
;
786 char *optval
= strchr(optspec
, '=');
787 if (optval
) *optval
++ = 0;
789 if (!strcasecmp(optname
, "debug")) {
791 u
->debug_level
= atoi(optval
);
794 } else if (!strcasecmp(optname
, "mercy") && optval
) {
795 /* Minimal difference of black/white captures
796 * to stop playout - "Mercy Rule". Speeds up
797 * hopeless playouts at the expense of some
799 u
->mercymin
= atoi(optval
);
800 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
801 u
->gamelen
= atoi(optval
);
802 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
803 u
->expand_p
= atoi(optval
);
804 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
805 u
->dumpthres
= atoi(optval
);
806 } else if (!strcasecmp(optname
, "best2_ratio") && optval
) {
807 /* If set, prolong simulating while
808 * first_best/second_best playouts ratio
809 * is less than best2_ratio. */
810 u
->best2_ratio
= atof(optval
);
811 } else if (!strcasecmp(optname
, "playout_amaf")) {
812 /* Whether to include random playout moves in
813 * AMAF as well. (Otherwise, only tree moves
814 * are included in AMAF. Of course makes sense
815 * only in connection with an AMAF policy.) */
816 /* with-without: 55.5% (+-4.1) */
817 if (optval
&& *optval
== '0')
818 u
->playout_amaf
= false;
820 u
->playout_amaf
= true;
821 } else if (!strcasecmp(optname
, "playout_amaf_nakade")) {
822 /* Whether to include nakade moves from playouts
823 * in the AMAF statistics; this tends to nullify
824 * the playout_amaf effect by adding too much
826 if (optval
&& *optval
== '0')
827 u
->playout_amaf_nakade
= false;
829 u
->playout_amaf_nakade
= true;
830 } else if (!strcasecmp(optname
, "playout_amaf_cutoff") && optval
) {
831 /* Keep only first N% of playout stage AMAF
833 u
->playout_amaf_cutoff
= atoi(optval
);
834 } else if ((!strcasecmp(optname
, "policy") || !strcasecmp(optname
, "random_policy")) && optval
) {
835 char *policyarg
= strchr(optval
, ':');
836 struct uct_policy
**p
= !strcasecmp(optname
, "policy") ? &u
->policy
: &u
->random_policy
;
839 if (!strcasecmp(optval
, "ucb1")) {
840 *p
= policy_ucb1_init(u
, policyarg
);
841 } else if (!strcasecmp(optval
, "ucb1amaf")) {
842 *p
= policy_ucb1amaf_init(u
, policyarg
);
844 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
847 } else if (!strcasecmp(optname
, "playout") && optval
) {
848 char *playoutarg
= strchr(optval
, ':');
851 if (!strcasecmp(optval
, "moggy")) {
852 u
->playout
= playout_moggy_init(playoutarg
);
853 } else if (!strcasecmp(optval
, "light")) {
854 u
->playout
= playout_light_init(playoutarg
);
855 } else if (!strcasecmp(optval
, "elo")) {
856 u
->playout
= playout_elo_init(playoutarg
);
858 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
861 } else if (!strcasecmp(optname
, "prior") && optval
) {
862 u
->prior
= uct_prior_init(optval
, b
);
863 } else if (!strcasecmp(optname
, "amaf_prior") && optval
) {
864 u
->amaf_prior
= atoi(optval
);
865 } else if (!strcasecmp(optname
, "threads") && optval
) {
866 /* By default, Pachi will run with only single
867 * tree search thread! */
868 u
->threads
= atoi(optval
);
869 } else if (!strcasecmp(optname
, "thread_model") && optval
) {
870 if (!strcasecmp(optval
, "root")) {
871 /* Root parallelization - each thread
872 * does independent search, trees are
873 * merged at the end. */
874 u
->thread_model
= TM_ROOT
;
875 u
->parallel_tree
= false;
876 u
->virtual_loss
= false;
877 } else if (!strcasecmp(optval
, "tree")) {
878 /* Tree parallelization - all threads
879 * grind on the same tree. */
880 u
->thread_model
= TM_TREE
;
881 u
->parallel_tree
= true;
882 u
->virtual_loss
= false;
883 } else if (!strcasecmp(optval
, "treevl")) {
884 /* Tree parallelization, but also
885 * with virtual losses - this discou-
886 * rages most threads choosing the
887 * same tree branches to read. */
888 u
->thread_model
= TM_TREEVL
;
889 u
->parallel_tree
= true;
890 u
->virtual_loss
= true;
892 fprintf(stderr
, "UCT: Invalid thread model %s\n", optval
);
895 } else if (!strcasecmp(optname
, "pondering")) {
896 /* Keep searching even during opponent's turn. */
897 u
->pondering_opt
= !optval
|| atoi(optval
);
898 } else if (!strcasecmp(optname
, "fuseki_end") && optval
) {
899 /* At the very beginning it's not worth thinking
900 * too long because the playout evaluations are
901 * very noisy. So gradually increase the thinking
902 * time up to maximum when fuseki_end percent
903 * of the board has been played.
904 * This only applies if we are not in byoyomi. */
905 u
->fuseki_end
= atoi(optval
);
906 } else if (!strcasecmp(optname
, "yose_start") && optval
) {
907 /* When yose_start percent of the board has been
908 * played, or if we are in byoyomi, stop spending
909 * more time and spread the remaining time
911 * Between fuseki_end and yose_start, we spend
912 * a constant proportion of the remaining time
913 * on each move. (yose_start should actually
914 * be much earlier than when real yose start,
915 * but "yose" is a good short name to convey
917 u
->yose_start
= atoi(optval
);
918 } else if (!strcasecmp(optname
, "force_seed") && optval
) {
919 u
->force_seed
= atoi(optval
);
920 } else if (!strcasecmp(optname
, "no_book")) {
922 } else if (!strcasecmp(optname
, "dynkomi")) {
923 /* Dynamic komi in handicap game; linearly
924 * decreases to basic settings until move
926 u
->dynkomi
= optval
? atoi(optval
) : 150;
927 } else if (!strcasecmp(optname
, "dynkomi_mask") && optval
) {
928 /* Bitmask of colors the player must be
929 * for dynkomi be applied; you may want
930 * to use dynkomi_mask=3 to allow dynkomi
931 * even in games where Pachi is white. */
932 u
->dynkomi_mask
= atoi(optval
);
933 } else if (!strcasecmp(optname
, "val_scale") && optval
) {
934 /* How much of the game result value should be
935 * influenced by win size. Zero means it isn't. */
936 u
->val_scale
= atof(optval
);
937 } else if (!strcasecmp(optname
, "val_points") && optval
) {
938 /* Maximum size of win to be scaled into game
939 * result value. Zero means boardsize^2. */
940 u
->val_points
= atoi(optval
) * 2; // result values are doubled
941 } else if (!strcasecmp(optname
, "val_extra")) {
942 /* If false, the score coefficient will be simply
943 * added to the value, instead of scaling the result
944 * coefficient because of it. */
945 u
->val_extra
= !optval
|| atoi(optval
);
946 } else if (!strcasecmp(optname
, "root_heuristic") && optval
) {
947 /* Whether to bias exploration by root node values
948 * (must be supported by the used policy).
950 * 1: Do, value = result.
951 * Try to temper the result:
952 * 2: Do, value = 0.5+(result-expected)/2.
953 * 3: Do, value = 0.5+bzz((result-expected)^2). */
954 u
->root_heuristic
= atoi(optval
);
955 } else if (!strcasecmp(optname
, "pass_all_alive")) {
956 /* Whether to consider all stones alive at the game
957 * end instead of marking dead groupd. */
958 u
->pass_all_alive
= !optval
|| atoi(optval
);
959 } else if (!strcasecmp(optname
, "random_policy_chance") && optval
) {
960 /* If specified (N), with probability 1/N, random_policy policy
961 * descend is used instead of main policy descend; useful
962 * if specified policy (e.g. UCB1AMAF) can make unduly biased
963 * choices sometimes, you can fall back to e.g.
964 * random_policy=UCB1. */
965 u
->random_policy_chance
= atoi(optval
);
966 } else if (!strcasecmp(optname
, "max_tree_size") && optval
) {
967 /* Maximum amount of memory [MiB] consumed by the move tree.
968 * Default is 3072 (3 GiB). Note that if you use TM_ROOT,
969 * this limits size of only one of the trees, not all of them
971 u
->max_tree_size
= atol(optval
) * 1048576;
972 } else if (!strcasecmp(optname
, "fast_alloc")) {
973 u
->fast_alloc
= !optval
|| atoi(optval
);
974 } else if (!strcasecmp(optname
, "banner") && optval
) {
975 /* Additional banner string. This must come as the
976 * last engine parameter. */
977 if (*next
) *--next
= ',';
978 u
->banner
= strdup(optval
);
981 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
987 u
->resign_ratio
= 0.2; /* Resign when most games are lost. */
988 u
->loss_threshold
= 0.85; /* Stop reading if after at least 5000 playouts this is best value. */
990 u
->policy
= policy_ucb1amaf_init(u
, NULL
);
992 if (!!u
->random_policy_chance
^ !!u
->random_policy
) {
993 fprintf(stderr
, "uct: Only one of random_policy and random_policy_chance is set\n");
997 if (u
->fast_alloc
&& !u
->parallel_tree
) {
998 fprintf(stderr
, "fast_alloc not supported with root parallelization.\n");
1003 u
->prior
= uct_prior_init(NULL
, b
);
1006 u
->playout
= playout_moggy_init(NULL
);
1007 u
->playout
->debug_level
= u
->debug_level
;
1009 u
->ownermap
.map
= malloc(board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
1011 /* Some things remain uninitialized for now - the opening book
1012 * is not loaded and the tree not set up. */
1013 /* This will be initialized in setup_state() at the first move
1014 * received/requested. This is because right now we are not aware
1015 * about any komi or handicap setup and such. */
1021 engine_uct_init(char *arg
, struct board
*b
)
1023 struct uct
*u
= uct_state_init(arg
, b
);
1024 struct engine
*e
= calloc(1, sizeof(struct engine
));
1025 e
->name
= "UCT Engine";
1026 e
->printhook
= uct_printhook_ownermap
;
1027 e
->notify_play
= uct_notify_play
;
1029 e
->genmove
= uct_genmove
;
1030 e
->dead_group_list
= uct_dead_group_list
;
1034 const char banner
[] = "I'm playing UCT. When I'm losing, I will resign, "
1035 "if I think I win, I play until you pass. "
1036 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
1037 if (!u
->banner
) u
->banner
= "";
1038 e
->comment
= malloc(sizeof(banner
) + strlen(u
->banner
) + 1);
1039 sprintf(e
->comment
, "%s %s", banner
, u
->banner
);