18 #include "playout/elo.h"
19 #include "playout/moggy.h"
20 #include "playout/light.h"
24 #include "distributed/distributed.h"
25 #include "uct/dynkomi.h"
26 #include "uct/internal.h"
27 #include "uct/prior.h"
32 struct uct_policy
*policy_ucb1_init(struct uct
*u
, char *arg
);
33 struct uct_policy
*policy_ucb1amaf_init(struct uct
*u
, char *arg
);
34 static void uct_pondering_stop(struct uct
*u
);
35 static void uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
);
38 /* Default number of simulations to perform per move.
39 * Note that this is now in total over all threads! (Unless TM_ROOT.) */
40 #define MC_GAMES 80000
41 #define MC_GAMELEN MAX_GAMELEN
42 static const struct time_info default_ti
= {
45 .len
= { .games
= MC_GAMES
},
48 /* How big proportion of ownermap counts must be of one color to consider
51 /* How many games to consider at minimum before judging groups. */
52 #define GJ_MINGAMES 500
54 /* How often to inspect the tree from the main thread to check for playout
55 * stop, progress reports, etc. (in seconds) */
56 #define TREE_BUSYWAIT_INTERVAL 0.1 /* 100ms */
58 /* Once per how many simulations (per thread) to show a progress report line. */
59 #define TREE_SIMPROGRESS_INTERVAL 10000
61 /* When terminating uct_search() early, the safety margin to add to the
62 * remaining playout number estimate when deciding whether the result can
64 #define PLAYOUT_DELTA_SAFEMARGIN 1000
68 setup_state(struct uct
*u
, struct board
*b
, enum stone color
)
70 u
->t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0, u
->local_tree_aging
);
72 fast_srandom(u
->force_seed
);
74 fprintf(stderr
, "Fresh board with random seed %lu\n", fast_getseed());
75 //board_print(b, stderr);
76 if (!u
->no_book
&& b
->moves
== 0) {
77 assert(color
== S_BLACK
);
83 reset_state(struct uct
*u
)
86 tree_done(u
->t
); u
->t
= NULL
;
90 setup_dynkomi(struct uct
*u
, struct board
*b
, enum stone to_play
)
92 if (u
->t
->use_extra_komi
&& u
->dynkomi
->permove
)
93 u
->t
->extra_komi
= u
->dynkomi
->permove(u
->dynkomi
, b
, u
->t
);
97 prepare_move(struct engine
*e
, struct board
*b
, enum stone color
)
99 struct uct
*u
= e
->data
;
102 /* Verify that we have sane state. */
104 assert(u
->t
&& b
->moves
);
105 if (color
!= stone_other(u
->t
->root_color
)) {
106 fprintf(stderr
, "Fatal: Non-alternating play detected %d %d\n",
107 color
, u
->t
->root_color
);
112 /* We need fresh state. */
114 setup_state(u
, b
, color
);
117 u
->ownermap
.playouts
= 0;
118 memset(u
->ownermap
.map
, 0, board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
122 dead_group_list(struct uct
*u
, struct board
*b
, struct move_queue
*mq
)
124 struct group_judgement gj
;
126 gj
.gs
= alloca(board_size2(b
) * sizeof(gj
.gs
[0]));
127 board_ownermap_judge_group(b
, &u
->ownermap
, &gj
);
128 groups_of_status(b
, &gj
, GS_DEAD
, mq
);
132 uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
)
134 if (u
->ownermap
.playouts
< GJ_MINGAMES
)
137 struct move_queue mq
= { .moves
= 0 };
139 dead_group_list(u
, b
, &mq
);
140 return pass_is_safe(b
, color
, &mq
);
143 /* This function is called only when running as slave in the distributed version. */
144 static enum parse_code
145 uct_notify(struct engine
*e
, struct board
*b
, int id
, char *cmd
, char *args
, char **reply
)
147 struct uct
*u
= e
->data
;
149 /* Force resending the whole command history if we are out of sync
150 * but do it only once, not if already getting the history. */
151 if ((move_number(id
) != b
->moves
|| !b
->size
)
152 && !reply_disabled(id
) && !is_reset(cmd
)) {
154 fprintf(stderr
, "Out of sync, id %d, move %d\n", id
, b
->moves
);
155 static char buf
[128];
156 snprintf(buf
, sizeof(buf
), "out of sync, move %d expected", b
->moves
);
160 return reply_disabled(id
) ? P_NOREPLY
: P_OK
;
164 uct_printhook_ownermap(struct board
*board
, coord_t c
, FILE *f
)
166 struct uct
*u
= board
->es
;
168 const char chr
[] = ":XO,"; // dame, black, white, unclear
169 const char chm
[] = ":xo,";
170 char ch
= chr
[board_ownermap_judge_point(&u
->ownermap
, c
, GJ_THRES
)];
171 if (ch
== ',') { // less precise estimate then?
172 ch
= chm
[board_ownermap_judge_point(&u
->ownermap
, c
, 0.67)];
174 fprintf(f
, "%c ", ch
);
178 uct_notify_play(struct engine
*e
, struct board
*b
, struct move
*m
)
180 struct uct
*u
= e
->data
;
182 /* No state, create one - this is probably game beginning
183 * and we need to load the opening book right now. */
184 prepare_move(e
, b
, m
->color
);
188 /* Stop pondering, required by tree_promote_at() */
189 uct_pondering_stop(u
);
191 if (is_resign(m
->coord
)) {
197 /* Promote node of the appropriate move to the tree root. */
199 if (!tree_promote_at(u
->t
, b
, m
->coord
)) {
201 fprintf(stderr
, "Warning: Cannot promote move node! Several play commands in row?\n");
206 /* If we are a slave in a distributed engine, start pondering once
207 * we know which move we actually played. See uct_genmove() about
208 * the check for pass. */
209 if (u
->pondering_opt
&& u
->slave
&& m
->color
== u
->my_color
&& !is_pass(m
->coord
))
210 uct_pondering_start(u
, b
, u
->t
, stone_other(m
->color
));
216 uct_chat(struct engine
*e
, struct board
*b
, char *cmd
)
218 struct uct
*u
= e
->data
;
219 static char reply
[1024];
221 cmd
+= strspn(cmd
, " \n\t");
222 if (!strncasecmp(cmd
, "winrate", 7)) {
224 return "no game context (yet?)";
225 enum stone color
= u
->t
->root_color
;
226 struct tree_node
*n
= u
->t
->root
;
227 snprintf(reply
, 1024, "In %d playouts at %d threads, %s %s can win with %.2f%% probability",
228 n
->u
.playouts
, u
->threads
, stone2str(color
), coord2sstr(n
->coord
, b
),
229 tree_node_get_value(u
->t
, -1, n
->u
.value
) * 100);
230 if (u
->t
->use_extra_komi
&& abs(u
->t
->extra_komi
) >= 0.5) {
231 sprintf(reply
+ strlen(reply
), ", while self-imposing extra komi %.1f",
241 uct_dead_group_list(struct engine
*e
, struct board
*b
, struct move_queue
*mq
)
243 struct uct
*u
= e
->data
;
245 /* This means the game is probably over, no use pondering on. */
246 uct_pondering_stop(u
);
248 if (u
->pass_all_alive
)
249 return; // no dead groups
251 bool mock_state
= false;
254 /* No state, but we cannot just back out - we might
255 * have passed earlier, only assuming some stones are
256 * dead, and then re-connected, only to lose counting
257 * when all stones are assumed alive. */
258 /* Mock up some state and seed the ownermap by few
260 prepare_move(e
, b
, S_BLACK
); assert(u
->t
);
261 for (int i
= 0; i
< GJ_MINGAMES
; i
++)
262 uct_playout(u
, b
, S_BLACK
, u
->t
);
266 dead_group_list(u
, b
, mq
);
269 /* Clean up the mock state in case we will receive
270 * a genmove; we could get a non-alternating-move
271 * error from prepare_move() in that case otherwise. */
277 playout_policy_done(struct playout_policy
*p
)
279 if (p
->done
) p
->done(p
);
280 if (p
->data
) free(p
->data
);
285 uct_done(struct engine
*e
)
287 /* This is called on engine reset, especially when clear_board
288 * is received and new game should begin. */
289 struct uct
*u
= e
->data
;
290 uct_pondering_stop(u
);
291 if (u
->t
) reset_state(u
);
292 free(u
->ownermap
.map
);
295 free(u
->random_policy
);
296 playout_policy_done(u
->playout
);
297 uct_prior_done(u
->prior
);
301 /* Pachi threading structure (if uct_playouts_parallel() is used):
304 * | main(), GTP communication, ...
305 * | starts and stops the search managed by thread_manager
308 * | spawns and collects worker threads
314 * uct_playouts() loop, doing descend-playout until uct_halt
316 * Another way to look at it is by functions (lines denote thread boundaries):
319 * | uct_search() (uct_search_start() .. uct_search_stop())
320 * | -----------------------
321 * | spawn_thread_manager()
322 * | -----------------------
324 * V uct_playouts() */
326 /* Set in thread manager in case the workers should stop. */
327 volatile sig_atomic_t uct_halt
= 0;
328 /* ID of the running worker thread. */
329 __thread
int thread_id
= -1;
330 /* ID of the thread manager. */
331 static pthread_t thread_manager
;
332 static bool thread_manager_running
;
334 static pthread_mutex_t finish_mutex
= PTHREAD_MUTEX_INITIALIZER
;
335 static pthread_cond_t finish_cond
= PTHREAD_COND_INITIALIZER
;
336 static volatile int finish_thread
;
337 static pthread_mutex_t finish_serializer
= PTHREAD_MUTEX_INITIALIZER
;
350 spawn_worker(void *ctx_
)
352 struct spawn_ctx
*ctx
= ctx_
;
354 fast_srandom(ctx
->seed
);
355 thread_id
= ctx
->tid
;
357 ctx
->games
= uct_playouts(ctx
->u
, ctx
->b
, ctx
->color
, ctx
->t
);
359 pthread_mutex_lock(&finish_serializer
);
360 pthread_mutex_lock(&finish_mutex
);
361 finish_thread
= ctx
->tid
;
362 pthread_cond_signal(&finish_cond
);
363 pthread_mutex_unlock(&finish_mutex
);
367 /* Thread manager, controlling worker threads. It must be called with
368 * finish_mutex lock held, but it will unlock it itself before exiting;
369 * this is necessary to be completely deadlock-free. */
370 /* The finish_cond can be signalled for it to stop; in that case,
371 * the caller should set finish_thread = -1. */
372 /* After it is started, it will update mctx->t to point at some tree
373 * used for the actual search (matters only for TM_ROOT), on return
374 * it will set mctx->games to the number of performed simulations. */
376 spawn_thread_manager(void *ctx_
)
378 /* In thread_manager, we use only some of the ctx fields. */
379 struct spawn_ctx
*mctx
= ctx_
;
380 struct uct
*u
= mctx
->u
;
381 struct tree
*t
= mctx
->t
;
382 bool shared_tree
= u
->parallel_tree
;
383 fast_srandom(mctx
->seed
);
385 int played_games
= 0;
386 pthread_t threads
[u
->threads
];
391 /* Garbage collect the tree by preference when pondering. */
392 if (u
->pondering
&& t
->nodes
&& t
->nodes_size
> t
->max_tree_size
/2) {
393 unsigned long temp_size
= (MIN_FREE_MEM_PERCENT
* t
->max_tree_size
) / 100;
394 t
->root
= tree_garbage_collect(t
, temp_size
, t
->root
);
397 /* Spawn threads... */
398 for (int ti
= 0; ti
< u
->threads
; ti
++) {
399 struct spawn_ctx
*ctx
= malloc(sizeof(*ctx
));
400 ctx
->u
= u
; ctx
->b
= mctx
->b
; ctx
->color
= mctx
->color
;
401 mctx
->t
= ctx
->t
= shared_tree
? t
: tree_copy(t
);
402 ctx
->tid
= ti
; ctx
->seed
= fast_random(65536) + ti
;
403 pthread_create(&threads
[ti
], NULL
, spawn_worker
, ctx
);
405 fprintf(stderr
, "Spawned worker %d\n", ti
);
408 /* ...and collect them back: */
409 while (joined
< u
->threads
) {
410 /* Wait for some thread to finish... */
411 pthread_cond_wait(&finish_cond
, &finish_mutex
);
412 if (finish_thread
< 0) {
413 /* Stop-by-caller. Tell the workers to wrap up. */
417 /* ...and gather its remnants. */
418 struct spawn_ctx
*ctx
;
419 pthread_join(threads
[finish_thread
], (void **) &ctx
);
420 played_games
+= ctx
->games
;
423 if (ctx
->t
== mctx
->t
) mctx
->t
= t
;
424 tree_merge(t
, ctx
->t
);
429 fprintf(stderr
, "Joined worker %d\n", finish_thread
);
430 pthread_mutex_unlock(&finish_serializer
);
433 pthread_mutex_unlock(&finish_mutex
);
436 tree_normalize(mctx
->t
, u
->threads
);
438 mctx
->games
= played_games
;
442 static struct spawn_ctx
*
443 uct_search_start(struct uct
*u
, struct board
*b
, enum stone color
, struct tree
*t
)
445 assert(u
->threads
> 0);
446 assert(!thread_manager_running
);
448 struct spawn_ctx ctx
= { .u
= u
, .b
= b
, .color
= color
, .t
= t
, .seed
= fast_random(65536) };
449 static struct spawn_ctx mctx
; mctx
= ctx
;
450 pthread_mutex_lock(&finish_mutex
);
451 pthread_create(&thread_manager
, NULL
, spawn_thread_manager
, &mctx
);
452 thread_manager_running
= true;
456 static struct spawn_ctx
*
457 uct_search_stop(void)
459 assert(thread_manager_running
);
461 /* Signal thread manager to stop the workers. */
462 pthread_mutex_lock(&finish_mutex
);
464 pthread_cond_signal(&finish_cond
);
465 pthread_mutex_unlock(&finish_mutex
);
467 /* Collect the thread manager. */
468 struct spawn_ctx
*pctx
;
469 thread_manager_running
= false;
470 pthread_join(thread_manager
, (void **) &pctx
);
475 /* Determine whether we should terminate the search early. */
477 uct_search_stop_early(struct uct
*u
, struct tree
*t
, struct board
*b
,
478 struct time_info
*ti
, struct time_stop
*stop
,
479 struct tree_node
*best
, struct tree_node
*best2
,
480 int base_playouts
, int i
)
482 /* Always use at least half the desired time. It is silly
483 * to lose a won game because we played a bad move in 0.1s. */
485 if (ti
->dim
== TD_WALLTIME
) {
486 elapsed
= time_now() - ti
->len
.t
.timer_start
;
487 if (elapsed
< 0.5 * stop
->desired
.time
) return false;
490 /* Early break in won situation. */
491 if (best
->u
.playouts
>= 2000 && tree_node_get_value(t
, 1, best
->u
.value
) >= u
->loss_threshold
)
493 /* Earlier break in super-won situation. */
494 if (best
->u
.playouts
>= 500 && tree_node_get_value(t
, 1, best
->u
.value
) >= 0.95)
497 /* Break early if we estimate the second-best move cannot
498 * catch up in assigned time anymore. We use all our time
499 * if we are in byoyomi with single stone remaining in our
500 * period, however - it's better to pre-ponder. */
501 bool time_indulgent
= (!ti
->len
.t
.main_time
&& ti
->len
.t
.byoyomi_stones
== 1);
502 if (best2
&& ti
->dim
== TD_WALLTIME
&& !time_indulgent
) {
503 double remaining
= stop
->worst
.time
- elapsed
;
504 double pps
= ((double)i
- base_playouts
) / elapsed
;
505 double estplayouts
= remaining
* pps
+ PLAYOUT_DELTA_SAFEMARGIN
;
506 if (best
->u
.playouts
> best2
->u
.playouts
+ estplayouts
) {
508 fprintf(stderr
, "Early stop, result cannot change: "
509 "best %d, best2 %d, estimated %f simulations to go\n",
510 best
->u
.playouts
, best2
->u
.playouts
, estplayouts
);
518 /* Determine whether we should terminate the search later. */
520 uct_search_keep_looking(struct uct
*u
, struct tree
*t
, struct board
*b
,
521 struct time_info
*ti
, struct time_stop
*stop
,
522 struct tree_node
*best
, struct tree_node
*best2
,
523 struct tree_node
*bestr
, struct tree_node
*winner
, int i
)
527 fprintf(stderr
, "Did not find best move, still trying...\n");
531 /* Do not waste time if we are winning. Spend up to worst time if
532 * we are unsure, but only desired time if we are sure of winning. */
533 float beta
= 2 * (tree_node_get_value(t
, 1, best
->u
.value
) - 0.5);
534 if (ti
->dim
== TD_WALLTIME
&& beta
> 0) {
535 double good_enough
= stop
->desired
.time
* beta
+ stop
->worst
.time
* (1 - beta
);
536 double elapsed
= time_now() - ti
->len
.t
.timer_start
;
537 if (elapsed
> good_enough
) return false;
540 if (u
->best2_ratio
> 0) {
541 /* Check best/best2 simulations ratio. If the
542 * two best moves give very similar results,
543 * keep simulating. */
544 if (best2
&& best2
->u
.playouts
545 && (double)best
->u
.playouts
/ best2
->u
.playouts
< u
->best2_ratio
) {
547 fprintf(stderr
, "Best2 ratio %f < threshold %f\n",
548 (double)best
->u
.playouts
/ best2
->u
.playouts
,
554 if (u
->bestr_ratio
> 0) {
555 /* Check best, best_best value difference. If the best move
556 * and its best child do not give similar enough results,
557 * keep simulating. */
558 if (bestr
&& bestr
->u
.playouts
559 && fabs((double)best
->u
.value
- bestr
->u
.value
) > u
->bestr_ratio
) {
561 fprintf(stderr
, "Bestr delta %f > threshold %f\n",
562 fabs((double)best
->u
.value
- bestr
->u
.value
),
568 if (winner
&& winner
!= best
) {
569 /* Keep simulating if best explored
570 * does not have also highest value. */
572 fprintf(stderr
, "[%d] best %3s [%d] %f != winner %3s [%d] %f\n", i
,
573 coord2sstr(best
->coord
, t
->board
),
574 best
->u
.playouts
, tree_node_get_value(t
, 1, best
->u
.value
),
575 coord2sstr(winner
->coord
, t
->board
),
576 winner
->u
.playouts
, tree_node_get_value(t
, 1, winner
->u
.value
));
580 /* No reason to keep simulating, bye. */
584 /* Run time-limited MCTS search on foreground. */
586 uct_search(struct uct
*u
, struct board
*b
, struct time_info
*ti
, enum stone color
, struct tree
*t
)
588 int base_playouts
= u
->t
->root
->u
.playouts
;
589 if (UDEBUGL(2) && base_playouts
> 0)
590 fprintf(stderr
, "<pre-simulated %d games skipped>\n", base_playouts
);
592 /* Set up time conditions. */
593 if (ti
->period
== TT_NULL
) *ti
= default_ti
;
594 struct time_stop stop
;
595 time_stop_conditions(ti
, b
, u
->fuseki_end
, u
->yose_start
, &stop
);
597 /* Number of last dynkomi adjustment. */
598 int last_dynkomi
= t
->root
->u
.playouts
;
599 /* Number of last game with progress print. */
600 int last_print
= t
->root
->u
.playouts
;
601 /* Number of simulations to wait before next print. */
602 int print_interval
= TREE_SIMPROGRESS_INTERVAL
* (u
->thread_model
== TM_ROOT
? 1 : u
->threads
);
603 /* Printed notification about full memory? */
604 bool print_fullmem
= false;
606 struct spawn_ctx
*ctx
= uct_search_start(u
, b
, color
, t
);
608 /* The search tree is ctx->t. This is normally == t, but in case of
609 * TM_ROOT, it is one of the trees belonging to the independent
610 * workers. It is important to reference ctx->t directly since the
611 * thread manager will swap the tree pointer asynchronously. */
612 /* XXX: This means TM_ROOT support is suboptimal since single stalled
613 * thread can stall the others in case of limiting the search by game
614 * count. However, TM_ROOT just does not deserve any more extra code
617 struct tree_node
*best
= NULL
;
618 struct tree_node
*best2
= NULL
; // Second-best move.
619 struct tree_node
*bestr
= NULL
; // best's best child.
620 struct tree_node
*winner
= NULL
;
622 double busywait_interval
= TREE_BUSYWAIT_INTERVAL
;
624 /* Now, just periodically poll the search tree. */
626 time_sleep(busywait_interval
);
627 /* busywait_interval should never be less than desired time, or the
628 * time control is broken. But if it happens to be less, we still search
629 * at least 100ms otherwise the move is completely random. */
631 int i
= ctx
->t
->root
->u
.playouts
;
633 /* Adjust dynkomi? */
634 if (ctx
->t
->use_extra_komi
&& u
->dynkomi
->permove
635 && u
->dynkomi_interval
636 && i
> last_dynkomi
+ u
->dynkomi_interval
) {
637 float old_dynkomi
= ctx
->t
->extra_komi
;
638 ctx
->t
->extra_komi
= u
->dynkomi
->permove(u
->dynkomi
, b
, ctx
->t
);
639 if (UDEBUGL(3) && old_dynkomi
!= ctx
->t
->extra_komi
)
640 fprintf(stderr
, "dynkomi adjusted (%f -> %f)\n", old_dynkomi
, ctx
->t
->extra_komi
);
643 /* Print progress? */
644 if (i
- last_print
> print_interval
) {
645 last_print
+= print_interval
; // keep the numbers tidy
646 uct_progress_status(u
, ctx
->t
, color
, last_print
);
648 if (!print_fullmem
&& ctx
->t
->nodes_size
> u
->max_tree_size
) {
650 fprintf(stderr
, "memory limit hit (%lu > %lu)\n", ctx
->t
->nodes_size
, u
->max_tree_size
);
651 print_fullmem
= true;
654 /* Never consider stopping if we played too few simulations.
655 * Maybe we risk losing on time when playing in super-extreme
656 * time pressure but the tree is going to be just too messed
657 * up otherwise - we might even play invalid suicides or pass
658 * when we mustn't. */
662 best
= u
->policy
->choose(u
->policy
, ctx
->t
->root
, b
, color
, resign
);
663 if (best
) best2
= u
->policy
->choose(u
->policy
, ctx
->t
->root
, b
, color
, best
->coord
);
665 /* Possibly stop search early if it's no use to try on. */
666 if (best
&& uct_search_stop_early(u
, ctx
->t
, b
, ti
, &stop
, best
, best2
, base_playouts
, i
))
669 /* Check against time settings. */
670 bool desired_done
= false;
671 if (ti
->dim
== TD_WALLTIME
) {
672 double elapsed
= time_now() - ti
->len
.t
.timer_start
;
673 if (elapsed
> stop
.worst
.time
) break;
674 desired_done
= elapsed
> stop
.desired
.time
;
676 } else { assert(ti
->dim
== TD_GAMES
);
677 if (i
> stop
.worst
.playouts
) break;
678 desired_done
= i
> stop
.desired
.playouts
;
681 /* We want to stop simulating, but are willing to keep trying
682 * if we aren't completely sure about the winner yet. */
684 if (u
->policy
->winner
&& u
->policy
->evaluate
) {
685 struct uct_descent descent
= { .node
= ctx
->t
->root
};
686 u
->policy
->winner(u
->policy
, ctx
->t
, &descent
);
687 winner
= descent
.node
;
690 bestr
= u
->policy
->choose(u
->policy
, best
, b
, stone_other(color
), resign
);
691 if (!uct_search_keep_looking(u
, ctx
->t
, b
, ti
, &stop
, best
, best2
, bestr
, winner
, i
))
695 /* TODO: Early break if best->variance goes under threshold and we already
696 * have enough playouts (possibly thanks to book or to pondering)? */
699 ctx
= uct_search_stop();
702 tree_dump(t
, u
->dumpthres
);
704 uct_progress_status(u
, t
, color
, ctx
->games
);
710 /* Start pondering background with @color to play. */
712 uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
)
715 fprintf(stderr
, "Starting to ponder with color %s\n", stone2str(stone_other(color
)));
718 /* We need a local board copy to ponder upon. */
719 struct board
*b
= malloc(sizeof(*b
)); board_copy(b
, b0
);
721 /* *b0 did not have the genmove'd move played yet. */
722 struct move m
= { t
->root
->coord
, t
->root_color
};
723 int res
= board_play(b
, &m
);
725 setup_dynkomi(u
, b
, stone_other(m
.color
));
727 /* Start MCTS manager thread "headless". */
728 uct_search_start(u
, b
, color
, t
);
731 /* uct_search_stop() frontend for the pondering (non-genmove) mode. */
733 uct_pondering_stop(struct uct
*u
)
735 u
->pondering
= false;
736 if (!thread_manager_running
)
739 /* Stop the thread manager. */
740 struct spawn_ctx
*ctx
= uct_search_stop();
742 fprintf(stderr
, "(pondering) ");
743 uct_progress_status(u
, ctx
->t
, ctx
->color
, ctx
->games
);
750 uct_genmove(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
752 double start_time
= time_now();
753 struct uct
*u
= e
->data
;
755 if (b
->superko_violation
) {
756 fprintf(stderr
, "!!! WARNING: SUPERKO VIOLATION OCCURED BEFORE THIS MOVE\n");
757 fprintf(stderr
, "Maybe you play with situational instead of positional superko?\n");
758 fprintf(stderr
, "I'm going to ignore the violation, but note that I may miss\n");
759 fprintf(stderr
, "some moves valid under this ruleset because of this.\n");
760 b
->superko_violation
= false;
764 uct_pondering_stop(u
);
765 prepare_move(e
, b
, color
);
769 /* How to decide whether to use dynkomi in this game? Since we use
770 * pondering, it's not simple "who-to-play" matter. Decide based on
771 * the last genmove issued. */
772 u
->t
->use_extra_komi
= !!(u
->dynkomi_mask
& color
);
773 setup_dynkomi(u
, b
, color
);
775 /* Make pessimistic assumption about komi for Japanese rules to
776 * avoid losing by 0.5 when winning by 0.5 with Chinese rules.
777 * The rules usually give the same winner if the integer part of komi
778 * is odd so we adjust the komi only if it is even (for a board of
779 * odd size). We are not trying to get an exact evaluation for rare
780 * cases of seki. For details see http://home.snafu.de/jasiek/parity.html
781 * TODO: Support the kgs-rules command once available. */
782 if (u
->territory_scoring
&& (((int)floor(b
->komi
) + board_size(b
)) & 1)) {
783 b
->komi
+= (color
== S_BLACK
? 1.0 : -1.0);
785 fprintf(stderr
, "Setting komi to %.1f assuming Japanese rules\n",
789 int base_playouts
= u
->t
->root
->u
.playouts
;
790 /* Perform the Monte Carlo Tree Search! */
791 int played_games
= uct_search(u
, b
, ti
, color
, u
->t
);
793 /* Choose the best move from the tree. */
794 struct tree_node
*best
= u
->policy
->choose(u
->policy
, u
->t
->root
, b
, color
, resign
);
796 if (!u
->slave
) reset_state(u
);
797 return coord_copy(pass
);
800 fprintf(stderr
, "*** WINNER is %s (%d,%d) with score %1.4f (%d/%d:%d/%d games), extra komi %f\n",
801 coord2sstr(best
->coord
, b
), coord_x(best
->coord
, b
), coord_y(best
->coord
, b
),
802 tree_node_get_value(u
->t
, 1, best
->u
.value
), best
->u
.playouts
,
803 u
->t
->root
->u
.playouts
, u
->t
->root
->u
.playouts
- base_playouts
, played_games
,
806 /* Do not resign if we're so short of time that evaluation of best
807 * move is completely unreliable, we might be winning actually.
808 * In this case best is almost random but still better than resign.
809 * Also do not resign if we are getting bad results while actually
810 * giving away extra komi points (dynkomi). */
811 if (tree_node_get_value(u
->t
, 1, best
->u
.value
) < u
->resign_ratio
812 && !is_pass(best
->coord
) && best
->u
.playouts
> GJ_MINGAMES
813 && u
->t
->extra_komi
<= 1 /* XXX we assume dynamic komi == we are black */) {
814 if (!u
->slave
) reset_state(u
);
815 return coord_copy(resign
);
818 /* If the opponent just passed and we win counting, always
820 if (b
->moves
> 1 && is_pass(b
->last_move
.coord
)) {
821 /* Make sure enough playouts are simulated. */
822 while (u
->ownermap
.playouts
< GJ_MINGAMES
)
823 uct_playout(u
, b
, color
, u
->t
);
824 if (uct_pass_is_safe(u
, b
, color
, u
->pass_all_alive
|| pass_all_alive
)) {
826 fprintf(stderr
, "<Will rather pass, looks safe enough.>\n");
831 /* If we are a slave in the distributed engine, we'll soon get
832 * a "play" command later telling us which move was chosen,
833 * and pondering now will not gain much. */
835 tree_promote_node(u
->t
, &best
);
837 /* After a pass, pondering is harmful for two reasons:
838 * (i) We might keep pondering even when the game is over.
839 * Of course this is the case for opponent resign as well.
840 * (ii) More importantly, the ownermap will get skewed since
841 * the UCT will start cutting off any playouts. */
842 if (u
->pondering_opt
&& !is_pass(best
->coord
)) {
843 uct_pondering_start(u
, b
, u
->t
, stone_other(color
));
847 double time
= time_now() - start_time
+ 0.000001; /* avoid divide by zero */
848 fprintf(stderr
, "genmove in %0.2fs (%d games/s, %d games/s/thread)\n",
849 time
, (int)(played_games
/time
), (int)(played_games
/time
/u
->threads
));
851 return coord_copy(best
->coord
);
856 uct_genmoves(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
858 struct uct
*u
= e
->data
;
861 coord_t
*c
= uct_genmove(e
, b
, ti
, color
, pass_all_alive
);
863 /* Return a buffer with one line "total_playouts threads" then a list of lines
864 * "coord playouts value". Keep this code in sync with select_best_move(). */
865 static char reply
[10240];
867 char *end
= reply
+ sizeof(reply
);
868 struct tree_node
*root
= u
->t
->root
;
869 r
+= snprintf(r
, end
- r
, "%d %d", root
->u
.playouts
, u
->threads
);
870 int min_playouts
= root
->u
.playouts
/ 100;
872 // Give a large weight to pass or resign, but still allow other moves.
873 if (is_pass(*c
) || is_resign(*c
))
874 r
+= snprintf(r
, end
- r
, "\n%s %d %.1f", coord2sstr(*c
, b
), root
->u
.playouts
,
878 for (struct tree_node
*ni
= root
->children
; ni
; ni
= ni
->sibling
) {
879 if (ni
->u
.playouts
<= min_playouts
880 || ni
->hints
& TREE_HINT_INVALID
881 || is_pass(ni
->coord
))
883 char *coord
= coord2sstr(ni
->coord
, b
);
884 // We return the values as stored in the tree, so from black's view.
885 r
+= snprintf(r
, end
- r
, "\n%s %d %.7f", coord
, ni
->u
.playouts
, ni
->u
.value
);
892 uct_genbook(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
)
894 struct uct
*u
= e
->data
;
895 if (!u
->t
) prepare_move(e
, b
, color
);
898 if (ti
->dim
== TD_GAMES
) {
899 /* Don't count in games that already went into the book. */
900 ti
->len
.games
+= u
->t
->root
->u
.playouts
;
902 uct_search(u
, b
, ti
, color
, u
->t
);
904 assert(ti
->dim
== TD_GAMES
);
905 tree_save(u
->t
, b
, ti
->len
.games
/ 100);
911 uct_dumpbook(struct engine
*e
, struct board
*b
, enum stone color
)
913 struct uct
*u
= e
->data
;
914 struct tree
*t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0, u
->local_tree_aging
);
922 uct_state_init(char *arg
, struct board
*b
)
924 struct uct
*u
= calloc(1, sizeof(struct uct
));
925 bool using_elo
= false;
927 u
->debug_level
= debug_level
;
928 u
->gamelen
= MC_GAMELEN
;
932 u
->playout_amaf
= true;
933 u
->playout_amaf_nakade
= false;
934 u
->amaf_prior
= false;
935 u
->max_tree_size
= 3072ULL * 1048576;
937 u
->dynkomi_mask
= S_BLACK
;
940 u
->thread_model
= TM_TREEVL
;
941 u
->parallel_tree
= true;
942 u
->virtual_loss
= true;
944 u
->fuseki_end
= 20; // max time at 361*20% = 72 moves (our 36th move, still 99 to play)
945 u
->yose_start
= 40; // (100-40-25)*361/100/2 = 63 moves still to play by us then
946 u
->bestr_ratio
= 0.02;
947 // 2.5 is clearly too much, but seems to compensate well for overly stern time allocations.
948 // TODO: Further tuning and experiments with better time allocation schemes.
949 u
->best2_ratio
= 2.5;
951 u
->val_scale
= 0.04; u
->val_points
= 40;
954 u
->local_tree_aging
= 2;
957 char *optspec
, *next
= arg
;
960 next
+= strcspn(next
, ",");
961 if (*next
) { *next
++ = 0; } else { *next
= 0; }
963 char *optname
= optspec
;
964 char *optval
= strchr(optspec
, '=');
965 if (optval
) *optval
++ = 0;
967 if (!strcasecmp(optname
, "debug")) {
969 u
->debug_level
= atoi(optval
);
972 } else if (!strcasecmp(optname
, "mercy") && optval
) {
973 /* Minimal difference of black/white captures
974 * to stop playout - "Mercy Rule". Speeds up
975 * hopeless playouts at the expense of some
977 u
->mercymin
= atoi(optval
);
978 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
979 u
->gamelen
= atoi(optval
);
980 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
981 u
->expand_p
= atoi(optval
);
982 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
983 u
->dumpthres
= atoi(optval
);
984 } else if (!strcasecmp(optname
, "best2_ratio") && optval
) {
985 /* If set, prolong simulating while
986 * first_best/second_best playouts ratio
987 * is less than best2_ratio. */
988 u
->best2_ratio
= atof(optval
);
989 } else if (!strcasecmp(optname
, "bestr_ratio") && optval
) {
990 /* If set, prolong simulating while
991 * best,best_best_child values delta
992 * is more than bestr_ratio. */
993 u
->bestr_ratio
= atof(optval
);
994 } else if (!strcasecmp(optname
, "playout_amaf")) {
995 /* Whether to include random playout moves in
996 * AMAF as well. (Otherwise, only tree moves
997 * are included in AMAF. Of course makes sense
998 * only in connection with an AMAF policy.) */
999 /* with-without: 55.5% (+-4.1) */
1000 if (optval
&& *optval
== '0')
1001 u
->playout_amaf
= false;
1003 u
->playout_amaf
= true;
1004 } else if (!strcasecmp(optname
, "playout_amaf_nakade")) {
1005 /* Whether to include nakade moves from playouts
1006 * in the AMAF statistics; this tends to nullify
1007 * the playout_amaf effect by adding too much
1009 if (optval
&& *optval
== '0')
1010 u
->playout_amaf_nakade
= false;
1012 u
->playout_amaf_nakade
= true;
1013 } else if (!strcasecmp(optname
, "playout_amaf_cutoff") && optval
) {
1014 /* Keep only first N% of playout stage AMAF
1016 u
->playout_amaf_cutoff
= atoi(optval
);
1017 } else if ((!strcasecmp(optname
, "policy") || !strcasecmp(optname
, "random_policy")) && optval
) {
1018 char *policyarg
= strchr(optval
, ':');
1019 struct uct_policy
**p
= !strcasecmp(optname
, "policy") ? &u
->policy
: &u
->random_policy
;
1022 if (!strcasecmp(optval
, "ucb1")) {
1023 *p
= policy_ucb1_init(u
, policyarg
);
1024 } else if (!strcasecmp(optval
, "ucb1amaf")) {
1025 *p
= policy_ucb1amaf_init(u
, policyarg
);
1027 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
1030 } else if (!strcasecmp(optname
, "playout") && optval
) {
1031 char *playoutarg
= strchr(optval
, ':');
1034 if (!strcasecmp(optval
, "moggy")) {
1035 u
->playout
= playout_moggy_init(playoutarg
, b
);
1036 } else if (!strcasecmp(optval
, "light")) {
1037 u
->playout
= playout_light_init(playoutarg
, b
);
1038 } else if (!strcasecmp(optval
, "elo")) {
1039 u
->playout
= playout_elo_init(playoutarg
, b
);
1042 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
1045 } else if (!strcasecmp(optname
, "prior") && optval
) {
1046 u
->prior
= uct_prior_init(optval
, b
);
1047 } else if (!strcasecmp(optname
, "amaf_prior") && optval
) {
1048 u
->amaf_prior
= atoi(optval
);
1049 } else if (!strcasecmp(optname
, "threads") && optval
) {
1050 /* By default, Pachi will run with only single
1051 * tree search thread! */
1052 u
->threads
= atoi(optval
);
1053 } else if (!strcasecmp(optname
, "thread_model") && optval
) {
1054 if (!strcasecmp(optval
, "root")) {
1055 /* Root parallelization - each thread
1056 * does independent search, trees are
1057 * merged at the end. */
1058 u
->thread_model
= TM_ROOT
;
1059 u
->parallel_tree
= false;
1060 u
->virtual_loss
= false;
1061 } else if (!strcasecmp(optval
, "tree")) {
1062 /* Tree parallelization - all threads
1063 * grind on the same tree. */
1064 u
->thread_model
= TM_TREE
;
1065 u
->parallel_tree
= true;
1066 u
->virtual_loss
= false;
1067 } else if (!strcasecmp(optval
, "treevl")) {
1068 /* Tree parallelization, but also
1069 * with virtual losses - this discou-
1070 * rages most threads choosing the
1071 * same tree branches to read. */
1072 u
->thread_model
= TM_TREEVL
;
1073 u
->parallel_tree
= true;
1074 u
->virtual_loss
= true;
1076 fprintf(stderr
, "UCT: Invalid thread model %s\n", optval
);
1079 } else if (!strcasecmp(optname
, "pondering")) {
1080 /* Keep searching even during opponent's turn. */
1081 u
->pondering_opt
= !optval
|| atoi(optval
);
1082 } else if (!strcasecmp(optname
, "fuseki_end") && optval
) {
1083 /* At the very beginning it's not worth thinking
1084 * too long because the playout evaluations are
1085 * very noisy. So gradually increase the thinking
1086 * time up to maximum when fuseki_end percent
1087 * of the board has been played.
1088 * This only applies if we are not in byoyomi. */
1089 u
->fuseki_end
= atoi(optval
);
1090 } else if (!strcasecmp(optname
, "yose_start") && optval
) {
1091 /* When yose_start percent of the board has been
1092 * played, or if we are in byoyomi, stop spending
1093 * more time and spread the remaining time
1095 * Between fuseki_end and yose_start, we spend
1096 * a constant proportion of the remaining time
1097 * on each move. (yose_start should actually
1098 * be much earlier than when real yose start,
1099 * but "yose" is a good short name to convey
1101 u
->yose_start
= atoi(optval
);
1102 } else if (!strcasecmp(optname
, "force_seed") && optval
) {
1103 u
->force_seed
= atoi(optval
);
1104 } else if (!strcasecmp(optname
, "no_book")) {
1106 } else if (!strcasecmp(optname
, "dynkomi") && optval
) {
1107 /* Dynamic komi approach; there are multiple
1108 * ways to adjust komi dynamically throughout
1109 * play. We currently support two: */
1110 char *dynkomiarg
= strchr(optval
, ':');
1113 if (!strcasecmp(optval
, "none")) {
1114 u
->dynkomi
= uct_dynkomi_init_none(u
, dynkomiarg
, b
);
1115 } else if (!strcasecmp(optval
, "linear")) {
1116 u
->dynkomi
= uct_dynkomi_init_linear(u
, dynkomiarg
, b
);
1117 } else if (!strcasecmp(optval
, "adaptive")) {
1118 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, dynkomiarg
, b
);
1120 fprintf(stderr
, "UCT: Invalid dynkomi mode %s\n", optval
);
1123 } else if (!strcasecmp(optname
, "dynkomi_mask") && optval
) {
1124 /* Bitmask of colors the player must be
1125 * for dynkomi be applied; you may want
1126 * to use dynkomi_mask=3 to allow dynkomi
1127 * even in games where Pachi is white. */
1128 u
->dynkomi_mask
= atoi(optval
);
1129 } else if (!strcasecmp(optname
, "dynkomi_interval") && optval
) {
1130 /* If non-zero, re-adjust dynamic komi
1131 * throughout a single genmove reading,
1132 * roughly every N simulations. */
1133 u
->dynkomi_interval
= atoi(optval
);
1134 } else if (!strcasecmp(optname
, "val_scale") && optval
) {
1135 /* How much of the game result value should be
1136 * influenced by win size. Zero means it isn't. */
1137 u
->val_scale
= atof(optval
);
1138 } else if (!strcasecmp(optname
, "val_points") && optval
) {
1139 /* Maximum size of win to be scaled into game
1140 * result value. Zero means boardsize^2. */
1141 u
->val_points
= atoi(optval
) * 2; // result values are doubled
1142 } else if (!strcasecmp(optname
, "val_extra")) {
1143 /* If false, the score coefficient will be simply
1144 * added to the value, instead of scaling the result
1145 * coefficient because of it. */
1146 u
->val_extra
= !optval
|| atoi(optval
);
1147 } else if (!strcasecmp(optname
, "local_tree") && optval
) {
1148 /* Whether to bias exploration by local tree values
1149 * (must be supported by the used policy).
1151 * 1: Do, value = result.
1152 * Try to temper the result:
1153 * 2: Do, value = 0.5+(result-expected)/2.
1154 * 3: Do, value = 0.5+bzz((result-expected)^2).
1155 * 4: Do, value = 0.5+sqrt(result-expected)/2. */
1156 u
->local_tree
= atoi(optval
);
1157 } else if (!strcasecmp(optname
, "tenuki_d") && optval
) {
1158 /* Tenuki distance at which to break the local tree. */
1159 u
->tenuki_d
= atoi(optval
);
1160 if (u
->tenuki_d
> TREE_NODE_D_MAX
+ 1) {
1161 fprintf(stderr
, "uct: tenuki_d must not be larger than TREE_NODE_D_MAX+1 %d\n", TREE_NODE_D_MAX
+ 1);
1164 } else if (!strcasecmp(optname
, "local_tree_aging") && optval
) {
1165 /* How much to reduce local tree values between moves. */
1166 u
->local_tree_aging
= atof(optval
);
1167 } else if (!strcasecmp(optname
, "local_tree_allseq")) {
1168 /* By default, only complete sequences are stored
1169 * in the local tree. If this is on, also
1170 * subsequences starting at each move are stored. */
1171 u
->local_tree_allseq
= !optval
|| atoi(optval
);
1172 } else if (!strcasecmp(optname
, "local_tree_playout")) {
1173 /* Whether to adjust ELO playout probability
1174 * distributions according to matched localtree
1176 u
->local_tree_playout
= !optval
|| atoi(optval
);
1177 } else if (!strcasecmp(optname
, "local_tree_pseqroot")) {
1178 /* By default, when we have no sequence move
1179 * to suggest in-playout, we give up. If this
1180 * is on, we make probability distribution from
1181 * sequences first moves instead. */
1182 u
->local_tree_pseqroot
= !optval
|| atoi(optval
);
1183 } else if (!strcasecmp(optname
, "pass_all_alive")) {
1184 /* Whether to consider all stones alive at the game
1185 * end instead of marking dead groupd. */
1186 u
->pass_all_alive
= !optval
|| atoi(optval
);
1187 } else if (!strcasecmp(optname
, "territory_scoring")) {
1188 /* Use territory scoring (default is area scoring).
1189 * An explicit kgs-rules command overrides this. */
1190 u
->territory_scoring
= !optval
|| atoi(optval
);
1191 } else if (!strcasecmp(optname
, "random_policy_chance") && optval
) {
1192 /* If specified (N), with probability 1/N, random_policy policy
1193 * descend is used instead of main policy descend; useful
1194 * if specified policy (e.g. UCB1AMAF) can make unduly biased
1195 * choices sometimes, you can fall back to e.g.
1196 * random_policy=UCB1. */
1197 u
->random_policy_chance
= atoi(optval
);
1198 } else if (!strcasecmp(optname
, "max_tree_size") && optval
) {
1199 /* Maximum amount of memory [MiB] consumed by the move tree.
1200 * For fast_alloc it includes the temp tree used for pruning.
1201 * Default is 3072 (3 GiB). Note that if you use TM_ROOT,
1202 * this limits size of only one of the trees, not all of them
1204 u
->max_tree_size
= atol(optval
) * 1048576;
1205 } else if (!strcasecmp(optname
, "fast_alloc")) {
1206 u
->fast_alloc
= !optval
|| atoi(optval
);
1207 } else if (!strcasecmp(optname
, "slave")) {
1208 /* Act as slave for the distributed engine. */
1209 u
->slave
= !optval
|| atoi(optval
);
1210 } else if (!strcasecmp(optname
, "banner") && optval
) {
1211 /* Additional banner string. This must come as the
1212 * last engine parameter. */
1213 if (*next
) *--next
= ',';
1214 u
->banner
= strdup(optval
);
1217 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
1223 u
->resign_ratio
= 0.2; /* Resign when most games are lost. */
1224 u
->loss_threshold
= 0.85; /* Stop reading if after at least 2000 playouts this is best value. */
1226 u
->policy
= policy_ucb1amaf_init(u
, NULL
);
1228 if (!!u
->random_policy_chance
^ !!u
->random_policy
) {
1229 fprintf(stderr
, "uct: Only one of random_policy and random_policy_chance is set\n");
1233 if (!u
->local_tree
) {
1234 /* No ltree aging. */
1235 u
->local_tree_aging
= 1.0f
;
1238 u
->local_tree_playout
= false;
1240 if (u
->fast_alloc
&& !u
->parallel_tree
) {
1241 fprintf(stderr
, "fast_alloc not supported with root parallelization.\n");
1245 u
->max_tree_size
= (100ULL * u
->max_tree_size
) / (100 + MIN_FREE_MEM_PERCENT
);
1248 u
->prior
= uct_prior_init(NULL
, b
);
1251 u
->playout
= playout_moggy_init(NULL
, b
);
1252 u
->playout
->debug_level
= u
->debug_level
;
1254 u
->ownermap
.map
= malloc(board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
1257 u
->dynkomi
= uct_dynkomi_init_linear(u
, NULL
, b
);
1259 /* Some things remain uninitialized for now - the opening book
1260 * is not loaded and the tree not set up. */
1261 /* This will be initialized in setup_state() at the first move
1262 * received/requested. This is because right now we are not aware
1263 * about any komi or handicap setup and such. */
1269 engine_uct_init(char *arg
, struct board
*b
)
1271 struct uct
*u
= uct_state_init(arg
, b
);
1272 struct engine
*e
= calloc(1, sizeof(struct engine
));
1273 e
->name
= "UCT Engine";
1274 e
->printhook
= uct_printhook_ownermap
;
1275 e
->notify_play
= uct_notify_play
;
1277 e
->genmove
= uct_genmove
;
1278 e
->genmoves
= uct_genmoves
;
1279 e
->dead_group_list
= uct_dead_group_list
;
1283 e
->notify
= uct_notify
;
1285 const char banner
[] = "I'm playing UCT. When I'm losing, I will resign, "
1286 "if I think I win, I play until you pass. "
1287 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
1288 if (!u
->banner
) u
->banner
= "";
1289 e
->comment
= malloc(sizeof(banner
) + strlen(u
->banner
) + 1);
1290 sprintf(e
->comment
, "%s %s", banner
, u
->banner
);