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
&& !reply_disabled(id
) && !is_reset(cmd
)) {
153 fprintf(stderr
, "Out of sync, id %d, move %d\n", id
, b
->moves
);
154 static char buf
[128];
155 snprintf(buf
, sizeof(buf
), "out of sync, move %d expected", b
->moves
);
159 return reply_disabled(id
) ? P_NOREPLY
: P_OK
;
163 uct_printhook_ownermap(struct board
*board
, coord_t c
, FILE *f
)
165 struct uct
*u
= board
->es
;
167 const char chr
[] = ":XO,"; // dame, black, white, unclear
168 const char chm
[] = ":xo,";
169 char ch
= chr
[board_ownermap_judge_point(&u
->ownermap
, c
, GJ_THRES
)];
170 if (ch
== ',') { // less precise estimate then?
171 ch
= chm
[board_ownermap_judge_point(&u
->ownermap
, c
, 0.67)];
173 fprintf(f
, "%c ", ch
);
177 uct_notify_play(struct engine
*e
, struct board
*b
, struct move
*m
)
179 struct uct
*u
= e
->data
;
181 /* No state, create one - this is probably game beginning
182 * and we need to load the opening book right now. */
183 prepare_move(e
, b
, m
->color
);
187 /* Stop pondering, required by tree_promote_at() */
188 uct_pondering_stop(u
);
190 if (is_resign(m
->coord
)) {
196 /* Promote node of the appropriate move to the tree root. */
198 if (!tree_promote_at(u
->t
, b
, m
->coord
)) {
200 fprintf(stderr
, "Warning: Cannot promote move node! Several play commands in row?\n");
205 /* If we are a slave in a distributed engine, start pondering once
206 * we know which move we actually played. See uct_genmove() about
207 * the check for pass. */
208 if (u
->pondering_opt
&& u
->slave
&& m
->color
== u
->my_color
&& !is_pass(m
->coord
))
209 uct_pondering_start(u
, b
, u
->t
, stone_other(m
->color
));
215 uct_chat(struct engine
*e
, struct board
*b
, char *cmd
)
217 struct uct
*u
= e
->data
;
218 static char reply
[1024];
220 cmd
+= strspn(cmd
, " \n\t");
221 if (!strncasecmp(cmd
, "winrate", 7)) {
223 return "no game context (yet?)";
224 enum stone color
= u
->t
->root_color
;
225 struct tree_node
*n
= u
->t
->root
;
226 snprintf(reply
, 1024, "In %d playouts at %d threads, %s %s can win with %.2f%% probability",
227 n
->u
.playouts
, u
->threads
, stone2str(color
), coord2sstr(n
->coord
, b
),
228 tree_node_get_value(u
->t
, -1, n
->u
.value
) * 100);
229 if (u
->t
->use_extra_komi
&& abs(u
->t
->extra_komi
) >= 0.5) {
230 sprintf(reply
+ strlen(reply
), ", while self-imposing extra komi %.1f",
240 uct_dead_group_list(struct engine
*e
, struct board
*b
, struct move_queue
*mq
)
242 struct uct
*u
= e
->data
;
244 /* This means the game is probably over, no use pondering on. */
245 uct_pondering_stop(u
);
247 if (u
->pass_all_alive
)
248 return; // no dead groups
250 bool mock_state
= false;
253 /* No state, but we cannot just back out - we might
254 * have passed earlier, only assuming some stones are
255 * dead, and then re-connected, only to lose counting
256 * when all stones are assumed alive. */
257 /* Mock up some state and seed the ownermap by few
259 prepare_move(e
, b
, S_BLACK
); assert(u
->t
);
260 for (int i
= 0; i
< GJ_MINGAMES
; i
++)
261 uct_playout(u
, b
, S_BLACK
, u
->t
);
265 dead_group_list(u
, b
, mq
);
268 /* Clean up the mock state in case we will receive
269 * a genmove; we could get a non-alternating-move
270 * error from prepare_move() in that case otherwise. */
276 playout_policy_done(struct playout_policy
*p
)
278 if (p
->done
) p
->done(p
);
279 if (p
->data
) free(p
->data
);
284 uct_done(struct engine
*e
)
286 /* This is called on engine reset, especially when clear_board
287 * is received and new game should begin. */
288 struct uct
*u
= e
->data
;
289 uct_pondering_stop(u
);
290 if (u
->t
) reset_state(u
);
291 free(u
->ownermap
.map
);
294 free(u
->random_policy
);
295 playout_policy_done(u
->playout
);
296 uct_prior_done(u
->prior
);
300 /* Pachi threading structure (if uct_playouts_parallel() is used):
303 * | main(), GTP communication, ...
304 * | starts and stops the search managed by thread_manager
307 * | spawns and collects worker threads
313 * uct_playouts() loop, doing descend-playout until uct_halt
315 * Another way to look at it is by functions (lines denote thread boundaries):
318 * | uct_search() (uct_search_start() .. uct_search_stop())
319 * | -----------------------
320 * | spawn_thread_manager()
321 * | -----------------------
323 * V uct_playouts() */
325 /* Set in thread manager in case the workers should stop. */
326 volatile sig_atomic_t uct_halt
= 0;
327 /* ID of the running worker thread. */
328 __thread
int thread_id
= -1;
329 /* ID of the thread manager. */
330 static pthread_t thread_manager
;
331 static bool thread_manager_running
;
333 static pthread_mutex_t finish_mutex
= PTHREAD_MUTEX_INITIALIZER
;
334 static pthread_cond_t finish_cond
= PTHREAD_COND_INITIALIZER
;
335 static volatile int finish_thread
;
336 static pthread_mutex_t finish_serializer
= PTHREAD_MUTEX_INITIALIZER
;
349 spawn_worker(void *ctx_
)
351 struct spawn_ctx
*ctx
= ctx_
;
353 fast_srandom(ctx
->seed
);
354 thread_id
= ctx
->tid
;
356 ctx
->games
= uct_playouts(ctx
->u
, ctx
->b
, ctx
->color
, ctx
->t
);
358 pthread_mutex_lock(&finish_serializer
);
359 pthread_mutex_lock(&finish_mutex
);
360 finish_thread
= ctx
->tid
;
361 pthread_cond_signal(&finish_cond
);
362 pthread_mutex_unlock(&finish_mutex
);
366 /* Thread manager, controlling worker threads. It must be called with
367 * finish_mutex lock held, but it will unlock it itself before exiting;
368 * this is necessary to be completely deadlock-free. */
369 /* The finish_cond can be signalled for it to stop; in that case,
370 * the caller should set finish_thread = -1. */
371 /* After it is started, it will update mctx->t to point at some tree
372 * used for the actual search (matters only for TM_ROOT), on return
373 * it will set mctx->games to the number of performed simulations. */
375 spawn_thread_manager(void *ctx_
)
377 /* In thread_manager, we use only some of the ctx fields. */
378 struct spawn_ctx
*mctx
= ctx_
;
379 struct uct
*u
= mctx
->u
;
380 struct tree
*t
= mctx
->t
;
381 bool shared_tree
= u
->parallel_tree
;
382 fast_srandom(mctx
->seed
);
384 int played_games
= 0;
385 pthread_t threads
[u
->threads
];
390 /* Garbage collect the tree by preference when pondering. */
391 if (u
->pondering
&& t
->nodes
&& t
->nodes_size
> t
->max_tree_size
/2) {
392 unsigned long temp_size
= (MIN_FREE_MEM_PERCENT
* t
->max_tree_size
) / 100;
393 t
->root
= tree_garbage_collect(t
, temp_size
, t
->root
);
396 /* Spawn threads... */
397 for (int ti
= 0; ti
< u
->threads
; ti
++) {
398 struct spawn_ctx
*ctx
= malloc(sizeof(*ctx
));
399 ctx
->u
= u
; ctx
->b
= mctx
->b
; ctx
->color
= mctx
->color
;
400 mctx
->t
= ctx
->t
= shared_tree
? t
: tree_copy(t
);
401 ctx
->tid
= ti
; ctx
->seed
= fast_random(65536) + ti
;
402 pthread_create(&threads
[ti
], NULL
, spawn_worker
, ctx
);
404 fprintf(stderr
, "Spawned worker %d\n", ti
);
407 /* ...and collect them back: */
408 while (joined
< u
->threads
) {
409 /* Wait for some thread to finish... */
410 pthread_cond_wait(&finish_cond
, &finish_mutex
);
411 if (finish_thread
< 0) {
412 /* Stop-by-caller. Tell the workers to wrap up. */
416 /* ...and gather its remnants. */
417 struct spawn_ctx
*ctx
;
418 pthread_join(threads
[finish_thread
], (void **) &ctx
);
419 played_games
+= ctx
->games
;
422 if (ctx
->t
== mctx
->t
) mctx
->t
= t
;
423 tree_merge(t
, ctx
->t
);
428 fprintf(stderr
, "Joined worker %d\n", finish_thread
);
429 pthread_mutex_unlock(&finish_serializer
);
432 pthread_mutex_unlock(&finish_mutex
);
435 tree_normalize(mctx
->t
, u
->threads
);
437 mctx
->games
= played_games
;
441 static struct spawn_ctx
*
442 uct_search_start(struct uct
*u
, struct board
*b
, enum stone color
, struct tree
*t
)
444 assert(u
->threads
> 0);
445 assert(!thread_manager_running
);
447 struct spawn_ctx ctx
= { .u
= u
, .b
= b
, .color
= color
, .t
= t
, .seed
= fast_random(65536) };
448 static struct spawn_ctx mctx
; mctx
= ctx
;
449 pthread_mutex_lock(&finish_mutex
);
450 pthread_create(&thread_manager
, NULL
, spawn_thread_manager
, &mctx
);
451 thread_manager_running
= true;
455 static struct spawn_ctx
*
456 uct_search_stop(void)
458 assert(thread_manager_running
);
460 /* Signal thread manager to stop the workers. */
461 pthread_mutex_lock(&finish_mutex
);
463 pthread_cond_signal(&finish_cond
);
464 pthread_mutex_unlock(&finish_mutex
);
466 /* Collect the thread manager. */
467 struct spawn_ctx
*pctx
;
468 thread_manager_running
= false;
469 pthread_join(thread_manager
, (void **) &pctx
);
474 /* Determine whether we should terminate the search early. */
476 uct_search_stop_early(struct uct
*u
, struct tree
*t
, struct board
*b
,
477 struct time_info
*ti
, struct time_stop
*stop
,
478 struct tree_node
*best
, struct tree_node
*best2
,
479 int base_playouts
, int i
)
481 /* Early break in won situation. */
482 if (best
->u
.playouts
>= 2000 && tree_node_get_value(t
, 1, best
->u
.value
) >= u
->loss_threshold
)
484 /* Earlier break in super-won situation. */
485 if (best
->u
.playouts
>= 500 && tree_node_get_value(t
, 1, best
->u
.value
) >= 0.95)
488 /* Break early if we estimate the second-best move cannot
489 * catch up in assigned time anymore. We use all our time
490 * if we are in byoyomi with single stone remaining in our
491 * period, however - it's better to pre-ponder. */
492 bool time_indulgent
= (!ti
->len
.t
.main_time
&& ti
->len
.t
.byoyomi_stones
== 1);
493 if (best2
&& ti
->dim
== TD_WALLTIME
&& !time_indulgent
) {
494 double elapsed
= time_now() - ti
->len
.t
.timer_start
;
495 double remaining
= stop
->worst
.time
- elapsed
;
496 double pps
= ((double)i
- base_playouts
) / elapsed
;
497 double estplayouts
= remaining
* pps
+ PLAYOUT_DELTA_SAFEMARGIN
;
498 if (best
->u
.playouts
> best2
->u
.playouts
+ estplayouts
) {
500 fprintf(stderr
, "Early stop, result cannot change: "
501 "best %d, best2 %d, estimated %f simulations to go\n",
502 best
->u
.playouts
, best2
->u
.playouts
, estplayouts
);
510 /* Determine whether we should terminate the search later. */
512 uct_search_keep_looking(struct uct
*u
, struct tree
*t
, struct board
*b
,
513 struct tree_node
*best
, struct tree_node
*best2
,
514 struct tree_node
*bestr
, struct tree_node
*winner
, int i
)
518 fprintf(stderr
, "Did not find best move, still trying...\n");
522 if (u
->best2_ratio
> 0) {
523 /* Check best/best2 simulations ratio. If the
524 * two best moves give very similar results,
525 * keep simulating. */
526 if (best2
&& best2
->u
.playouts
527 && (double)best
->u
.playouts
/ best2
->u
.playouts
< u
->best2_ratio
) {
529 fprintf(stderr
, "Best2 ratio %f < threshold %f\n",
530 (double)best
->u
.playouts
/ best2
->u
.playouts
,
536 if (u
->bestr_ratio
> 0) {
537 /* Check best, best_best value difference. If the best move
538 * and its best child do not give similar enough results,
539 * keep simulating. */
540 if (bestr
&& bestr
->u
.playouts
541 && fabs((double)best
->u
.value
- bestr
->u
.value
) > u
->bestr_ratio
) {
543 fprintf(stderr
, "Bestr delta %f > threshold %f\n",
544 fabs((double)best
->u
.value
- bestr
->u
.value
),
550 if (winner
&& winner
!= best
) {
551 /* Keep simulating if best explored
552 * does not have also highest value. */
554 fprintf(stderr
, "[%d] best %3s [%d] %f != winner %3s [%d] %f\n", i
,
555 coord2sstr(best
->coord
, t
->board
),
556 best
->u
.playouts
, tree_node_get_value(t
, 1, best
->u
.value
),
557 coord2sstr(winner
->coord
, t
->board
),
558 winner
->u
.playouts
, tree_node_get_value(t
, 1, winner
->u
.value
));
562 /* No reason to keep simulating, bye. */
566 /* Run time-limited MCTS search on foreground. */
568 uct_search(struct uct
*u
, struct board
*b
, struct time_info
*ti
, enum stone color
, struct tree
*t
)
570 int base_playouts
= u
->t
->root
->u
.playouts
;
571 if (UDEBUGL(2) && base_playouts
> 0)
572 fprintf(stderr
, "<pre-simulated %d games skipped>\n", base_playouts
);
574 /* Set up time conditions. */
575 if (ti
->period
== TT_NULL
) *ti
= default_ti
;
576 struct time_stop stop
;
577 time_stop_conditions(ti
, b
, u
->fuseki_end
, u
->yose_start
, &stop
);
579 /* Number of last game with progress print. */
580 int last_print
= t
->root
->u
.playouts
;
581 /* Number of simulations to wait before next print. */
582 int print_interval
= TREE_SIMPROGRESS_INTERVAL
* (u
->thread_model
== TM_ROOT
? 1 : u
->threads
);
583 /* Printed notification about full memory? */
584 bool print_fullmem
= false;
586 struct spawn_ctx
*ctx
= uct_search_start(u
, b
, color
, t
);
588 /* The search tree is ctx->t. This is normally == t, but in case of
589 * TM_ROOT, it is one of the trees belonging to the independent
590 * workers. It is important to reference ctx->t directly since the
591 * thread manager will swap the tree pointer asynchronously. */
592 /* XXX: This means TM_ROOT support is suboptimal since single stalled
593 * thread can stall the others in case of limiting the search by game
594 * count. However, TM_ROOT just does not deserve any more extra code
597 struct tree_node
*best
= NULL
;
598 struct tree_node
*best2
= NULL
; // Second-best move.
599 struct tree_node
*bestr
= NULL
; // best's best child.
600 struct tree_node
*winner
= NULL
;
602 double busywait_interval
= TREE_BUSYWAIT_INTERVAL
;
604 /* Now, just periodically poll the search tree. */
606 time_sleep(busywait_interval
);
607 /* busywait_interval should never be less than desired time, or the
608 * time control is broken. But if it happens to be less, we still search
609 * at least 100ms otherwise the move is completely random. */
611 int i
= ctx
->t
->root
->u
.playouts
;
613 /* Print progress? */
614 if (i
- last_print
> print_interval
) {
615 last_print
+= print_interval
; // keep the numbers tidy
616 uct_progress_status(u
, ctx
->t
, color
, last_print
);
618 if (!print_fullmem
&& ctx
->t
->nodes_size
> u
->max_tree_size
) {
620 fprintf(stderr
, "memory limit hit (%lu > %lu)\n", ctx
->t
->nodes_size
, u
->max_tree_size
);
621 print_fullmem
= true;
624 best
= u
->policy
->choose(u
->policy
, ctx
->t
->root
, b
, color
, resign
);
625 if (best
) best2
= u
->policy
->choose(u
->policy
, ctx
->t
->root
, b
, color
, best
->coord
);
627 /* Possibly stop search early if it's no use to try on. */
628 if (best
&& uct_search_stop_early(u
, ctx
->t
, b
, ti
, &stop
, best
, best2
, base_playouts
, i
))
631 /* Check against time settings. */
632 bool desired_done
= false;
633 if (ti
->dim
== TD_WALLTIME
) {
634 double elapsed
= time_now() - ti
->len
.t
.timer_start
;
635 if (elapsed
> stop
.worst
.time
) break;
636 desired_done
= elapsed
> stop
.desired
.time
;
638 } else { assert(ti
->dim
== TD_GAMES
);
639 if (i
> stop
.worst
.playouts
) break;
640 desired_done
= i
> stop
.desired
.playouts
;
643 /* We want to stop simulating, but are willing to keep trying
644 * if we aren't completely sure about the winner yet. */
646 if (u
->policy
->winner
&& u
->policy
->evaluate
) {
647 struct uct_descent descent
= { .node
= ctx
->t
->root
};
648 u
->policy
->winner(u
->policy
, ctx
->t
, &descent
);
649 winner
= descent
.node
;
652 bestr
= u
->policy
->choose(u
->policy
, best
, b
, stone_other(color
), resign
);
653 if (!uct_search_keep_looking(u
, ctx
->t
, b
, best
, best2
, bestr
, winner
, i
))
657 /* TODO: Early break if best->variance goes under threshold and we already
658 * have enough playouts (possibly thanks to book or to pondering)? */
661 ctx
= uct_search_stop();
664 tree_dump(t
, u
->dumpthres
);
666 uct_progress_status(u
, t
, color
, ctx
->games
);
672 /* Start pondering background with @color to play. */
674 uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
)
677 fprintf(stderr
, "Starting to ponder with color %s\n", stone2str(stone_other(color
)));
680 /* We need a local board copy to ponder upon. */
681 struct board
*b
= malloc(sizeof(*b
)); board_copy(b
, b0
);
683 /* *b0 did not have the genmove'd move played yet. */
684 struct move m
= { t
->root
->coord
, t
->root_color
};
685 int res
= board_play(b
, &m
);
687 setup_dynkomi(u
, b
, stone_other(m
.color
));
689 /* Start MCTS manager thread "headless". */
690 uct_search_start(u
, b
, color
, t
);
693 /* uct_search_stop() frontend for the pondering (non-genmove) mode. */
695 uct_pondering_stop(struct uct
*u
)
697 u
->pondering
= false;
698 if (!thread_manager_running
)
701 /* Stop the thread manager. */
702 struct spawn_ctx
*ctx
= uct_search_stop();
704 fprintf(stderr
, "(pondering) ");
705 uct_progress_status(u
, ctx
->t
, ctx
->color
, ctx
->games
);
712 uct_genmove(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
714 double start_time
= time_now();
715 struct uct
*u
= e
->data
;
717 if (b
->superko_violation
) {
718 fprintf(stderr
, "!!! WARNING: SUPERKO VIOLATION OCCURED BEFORE THIS MOVE\n");
719 fprintf(stderr
, "Maybe you play with situational instead of positional superko?\n");
720 fprintf(stderr
, "I'm going to ignore the violation, but note that I may miss\n");
721 fprintf(stderr
, "some moves valid under this ruleset because of this.\n");
722 b
->superko_violation
= false;
726 uct_pondering_stop(u
);
727 prepare_move(e
, b
, color
);
731 /* How to decide whether to use dynkomi in this game? Since we use
732 * pondering, it's not simple "who-to-play" matter. Decide based on
733 * the last genmove issued. */
734 u
->t
->use_extra_komi
= !!(u
->dynkomi_mask
& color
);
735 setup_dynkomi(u
, b
, color
);
737 /* Make pessimistic assumption about komi for Japanese rules to
738 * avoid losing by 0.5 when winning by 0.5 with Chinese rules.
739 * The rules usually give the same winner if the integer part of komi
740 * is odd so we adjust the komi only if it is even (for a board of
741 * odd size). We are not trying to get an exact evaluation for rare
742 * cases of seki. For details see http://home.snafu.de/jasiek/parity.html
743 * TODO: Support the kgs-rules command once available. */
744 if (u
->territory_scoring
&& (((int)floor(b
->komi
) + b
->size
) & 1)) {
745 b
->komi
+= (color
== S_BLACK
? 1.0 : -1.0);
747 fprintf(stderr
, "Setting komi to %.1f assuming Japanese rules\n",
751 int base_playouts
= u
->t
->root
->u
.playouts
;
752 /* Perform the Monte Carlo Tree Search! */
753 int played_games
= uct_search(u
, b
, ti
, color
, u
->t
);
755 /* Choose the best move from the tree. */
756 struct tree_node
*best
= u
->policy
->choose(u
->policy
, u
->t
->root
, b
, color
, resign
);
759 return coord_copy(pass
);
762 fprintf(stderr
, "*** WINNER is %s (%d,%d) with score %1.4f (%d/%d:%d/%d games)\n",
763 coord2sstr(best
->coord
, b
), coord_x(best
->coord
, b
), coord_y(best
->coord
, b
),
764 tree_node_get_value(u
->t
, 1, best
->u
.value
), best
->u
.playouts
,
765 u
->t
->root
->u
.playouts
, u
->t
->root
->u
.playouts
- base_playouts
, played_games
);
767 /* Do not resign if we're so short of time that evaluation of best move is completely
768 * unreliable, we might be winning actually. In this case best is almost random but
769 * still better than resign. */
770 if (tree_node_get_value(u
->t
, 1, best
->u
.value
) < u
->resign_ratio
&& !is_pass(best
->coord
)
771 && best
->u
.playouts
> GJ_MINGAMES
) {
773 return coord_copy(resign
);
776 /* If the opponent just passed and we win counting, always
778 if (b
->moves
> 1 && is_pass(b
->last_move
.coord
)) {
779 /* Make sure enough playouts are simulated. */
780 while (u
->ownermap
.playouts
< GJ_MINGAMES
)
781 uct_playout(u
, b
, color
, u
->t
);
782 if (uct_pass_is_safe(u
, b
, color
, u
->pass_all_alive
|| pass_all_alive
)) {
784 fprintf(stderr
, "<Will rather pass, looks safe enough.>\n");
789 /* If we are a slave in the distributed engine, we'll soon get
790 * a "play" command later telling us which move was chosen,
791 * and pondering now will not gain much. */
793 tree_promote_node(u
->t
, &best
);
795 /* After a pass, pondering is harmful for two reasons:
796 * (i) We might keep pondering even when the game is over.
797 * Of course this is the case for opponent resign as well.
798 * (ii) More importantly, the ownermap will get skewed since
799 * the UCT will start cutting off any playouts. */
800 if (u
->pondering_opt
&& !is_pass(best
->coord
)) {
801 uct_pondering_start(u
, b
, u
->t
, stone_other(color
));
805 double time
= time_now() - start_time
+ 0.000001; /* avoid divide by zero */
806 fprintf(stderr
, "genmove in %0.2fs (%d games/s, %d games/s/thread)\n",
807 time
, (int)(played_games
/time
), (int)(played_games
/time
/u
->threads
));
809 return coord_copy(best
->coord
);
814 uct_genmoves(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
816 struct uct
*u
= e
->data
;
819 coord_t
*c
= uct_genmove(e
, b
, ti
, color
, pass_all_alive
);
821 /* Return a buffer with one line "total_playouts threads" then a list of lines
822 * "coord playouts value". Keep this code in sync with select_best_move(). */
823 static char reply
[10240];
825 char *end
= reply
+ sizeof(reply
);
826 struct tree_node
*root
= u
->t
->root
;
827 r
+= snprintf(r
, end
- r
, "%d %d", root
->u
.playouts
, u
->threads
);
828 int min_playouts
= root
->u
.playouts
/ 100;
830 // Give a large weight to pass or resign, but still allow other moves.
831 if (is_pass(*c
) || is_resign(*c
))
832 r
+= snprintf(r
, end
- r
, "\n%s %d %.1f", coord2sstr(*c
, b
), root
->u
.playouts
,
836 for (struct tree_node
*ni
= root
->children
; ni
; ni
= ni
->sibling
) {
837 if (ni
->u
.playouts
<= min_playouts
838 || ni
->hints
& TREE_HINT_INVALID
839 || is_pass(ni
->coord
))
841 char *coord
= coord2sstr(ni
->coord
, b
);
842 // We return the values as stored in the tree, so from black's view.
843 r
+= snprintf(r
, end
- r
, "\n%s %d %.7f", coord
, ni
->u
.playouts
, ni
->u
.value
);
850 uct_genbook(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
)
852 struct uct
*u
= e
->data
;
853 if (!u
->t
) prepare_move(e
, b
, color
);
856 if (ti
->dim
== TD_GAMES
) {
857 /* Don't count in games that already went into the book. */
858 ti
->len
.games
+= u
->t
->root
->u
.playouts
;
860 uct_search(u
, b
, ti
, color
, u
->t
);
862 assert(ti
->dim
== TD_GAMES
);
863 tree_save(u
->t
, b
, ti
->len
.games
/ 100);
869 uct_dumpbook(struct engine
*e
, struct board
*b
, enum stone color
)
871 struct uct
*u
= e
->data
;
872 struct tree
*t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0, u
->local_tree_aging
);
880 uct_state_init(char *arg
, struct board
*b
)
882 struct uct
*u
= calloc(1, sizeof(struct uct
));
883 bool using_elo
= false;
885 u
->debug_level
= debug_level
;
886 u
->gamelen
= MC_GAMELEN
;
890 u
->playout_amaf
= true;
891 u
->playout_amaf_nakade
= false;
892 u
->amaf_prior
= false;
893 u
->max_tree_size
= 3072ULL * 1048576;
895 u
->dynkomi_mask
= S_BLACK
;
898 u
->thread_model
= TM_TREEVL
;
899 u
->parallel_tree
= true;
900 u
->virtual_loss
= true;
902 u
->fuseki_end
= 20; // max time at 361*20% = 72 moves (our 36th move, still 99 to play)
903 u
->yose_start
= 40; // (100-40-25)*361/100/2 = 63 moves still to play by us then
904 u
->bestr_ratio
= 0.02;
905 // 2.5 is clearly too much, but seems to compensate well for overly stern time allocations.
906 // TODO: Further tuning and experiments with better time allocation schemes.
907 u
->best2_ratio
= 2.5;
909 u
->val_scale
= 0.04; u
->val_points
= 40;
912 u
->local_tree_aging
= 2;
915 char *optspec
, *next
= arg
;
918 next
+= strcspn(next
, ",");
919 if (*next
) { *next
++ = 0; } else { *next
= 0; }
921 char *optname
= optspec
;
922 char *optval
= strchr(optspec
, '=');
923 if (optval
) *optval
++ = 0;
925 if (!strcasecmp(optname
, "debug")) {
927 u
->debug_level
= atoi(optval
);
930 } else if (!strcasecmp(optname
, "mercy") && optval
) {
931 /* Minimal difference of black/white captures
932 * to stop playout - "Mercy Rule". Speeds up
933 * hopeless playouts at the expense of some
935 u
->mercymin
= atoi(optval
);
936 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
937 u
->gamelen
= atoi(optval
);
938 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
939 u
->expand_p
= atoi(optval
);
940 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
941 u
->dumpthres
= atoi(optval
);
942 } else if (!strcasecmp(optname
, "best2_ratio") && optval
) {
943 /* If set, prolong simulating while
944 * first_best/second_best playouts ratio
945 * is less than best2_ratio. */
946 u
->best2_ratio
= atof(optval
);
947 } else if (!strcasecmp(optname
, "bestr_ratio") && optval
) {
948 /* If set, prolong simulating while
949 * best,best_best_child values delta
950 * is more than bestr_ratio. */
951 u
->bestr_ratio
= atof(optval
);
952 } else if (!strcasecmp(optname
, "playout_amaf")) {
953 /* Whether to include random playout moves in
954 * AMAF as well. (Otherwise, only tree moves
955 * are included in AMAF. Of course makes sense
956 * only in connection with an AMAF policy.) */
957 /* with-without: 55.5% (+-4.1) */
958 if (optval
&& *optval
== '0')
959 u
->playout_amaf
= false;
961 u
->playout_amaf
= true;
962 } else if (!strcasecmp(optname
, "playout_amaf_nakade")) {
963 /* Whether to include nakade moves from playouts
964 * in the AMAF statistics; this tends to nullify
965 * the playout_amaf effect by adding too much
967 if (optval
&& *optval
== '0')
968 u
->playout_amaf_nakade
= false;
970 u
->playout_amaf_nakade
= true;
971 } else if (!strcasecmp(optname
, "playout_amaf_cutoff") && optval
) {
972 /* Keep only first N% of playout stage AMAF
974 u
->playout_amaf_cutoff
= atoi(optval
);
975 } else if ((!strcasecmp(optname
, "policy") || !strcasecmp(optname
, "random_policy")) && optval
) {
976 char *policyarg
= strchr(optval
, ':');
977 struct uct_policy
**p
= !strcasecmp(optname
, "policy") ? &u
->policy
: &u
->random_policy
;
980 if (!strcasecmp(optval
, "ucb1")) {
981 *p
= policy_ucb1_init(u
, policyarg
);
982 } else if (!strcasecmp(optval
, "ucb1amaf")) {
983 *p
= policy_ucb1amaf_init(u
, policyarg
);
985 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
988 } else if (!strcasecmp(optname
, "playout") && optval
) {
989 char *playoutarg
= strchr(optval
, ':');
992 if (!strcasecmp(optval
, "moggy")) {
993 u
->playout
= playout_moggy_init(playoutarg
, b
);
994 } else if (!strcasecmp(optval
, "light")) {
995 u
->playout
= playout_light_init(playoutarg
, b
);
996 } else if (!strcasecmp(optval
, "elo")) {
997 u
->playout
= playout_elo_init(playoutarg
, b
);
1000 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
1003 } else if (!strcasecmp(optname
, "prior") && optval
) {
1004 u
->prior
= uct_prior_init(optval
, b
);
1005 } else if (!strcasecmp(optname
, "amaf_prior") && optval
) {
1006 u
->amaf_prior
= atoi(optval
);
1007 } else if (!strcasecmp(optname
, "threads") && optval
) {
1008 /* By default, Pachi will run with only single
1009 * tree search thread! */
1010 u
->threads
= atoi(optval
);
1011 } else if (!strcasecmp(optname
, "thread_model") && optval
) {
1012 if (!strcasecmp(optval
, "root")) {
1013 /* Root parallelization - each thread
1014 * does independent search, trees are
1015 * merged at the end. */
1016 u
->thread_model
= TM_ROOT
;
1017 u
->parallel_tree
= false;
1018 u
->virtual_loss
= false;
1019 } else if (!strcasecmp(optval
, "tree")) {
1020 /* Tree parallelization - all threads
1021 * grind on the same tree. */
1022 u
->thread_model
= TM_TREE
;
1023 u
->parallel_tree
= true;
1024 u
->virtual_loss
= false;
1025 } else if (!strcasecmp(optval
, "treevl")) {
1026 /* Tree parallelization, but also
1027 * with virtual losses - this discou-
1028 * rages most threads choosing the
1029 * same tree branches to read. */
1030 u
->thread_model
= TM_TREEVL
;
1031 u
->parallel_tree
= true;
1032 u
->virtual_loss
= true;
1034 fprintf(stderr
, "UCT: Invalid thread model %s\n", optval
);
1037 } else if (!strcasecmp(optname
, "pondering")) {
1038 /* Keep searching even during opponent's turn. */
1039 u
->pondering_opt
= !optval
|| atoi(optval
);
1040 } else if (!strcasecmp(optname
, "fuseki_end") && optval
) {
1041 /* At the very beginning it's not worth thinking
1042 * too long because the playout evaluations are
1043 * very noisy. So gradually increase the thinking
1044 * time up to maximum when fuseki_end percent
1045 * of the board has been played.
1046 * This only applies if we are not in byoyomi. */
1047 u
->fuseki_end
= atoi(optval
);
1048 } else if (!strcasecmp(optname
, "yose_start") && optval
) {
1049 /* When yose_start percent of the board has been
1050 * played, or if we are in byoyomi, stop spending
1051 * more time and spread the remaining time
1053 * Between fuseki_end and yose_start, we spend
1054 * a constant proportion of the remaining time
1055 * on each move. (yose_start should actually
1056 * be much earlier than when real yose start,
1057 * but "yose" is a good short name to convey
1059 u
->yose_start
= atoi(optval
);
1060 } else if (!strcasecmp(optname
, "force_seed") && optval
) {
1061 u
->force_seed
= atoi(optval
);
1062 } else if (!strcasecmp(optname
, "no_book")) {
1064 } else if (!strcasecmp(optname
, "dynkomi") && optval
) {
1065 /* Dynamic komi approach; there are multiple
1066 * ways to adjust komi dynamically throughout
1067 * play. We currently support two: */
1068 char *dynkomiarg
= strchr(optval
, ':');
1071 if (!strcasecmp(optval
, "none")) {
1072 u
->dynkomi
= uct_dynkomi_init_none(u
, dynkomiarg
, b
);
1073 } else if (!strcasecmp(optval
, "linear")) {
1074 u
->dynkomi
= uct_dynkomi_init_linear(u
, dynkomiarg
, b
);
1075 } else if (!strcasecmp(optval
, "adaptive")) {
1076 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, dynkomiarg
, b
);
1078 fprintf(stderr
, "UCT: Invalid dynkomi mode %s\n", optval
);
1081 } else if (!strcasecmp(optname
, "dynkomi_mask") && optval
) {
1082 /* Bitmask of colors the player must be
1083 * for dynkomi be applied; you may want
1084 * to use dynkomi_mask=3 to allow dynkomi
1085 * even in games where Pachi is white. */
1086 u
->dynkomi_mask
= atoi(optval
);
1087 } else if (!strcasecmp(optname
, "val_scale") && optval
) {
1088 /* How much of the game result value should be
1089 * influenced by win size. Zero means it isn't. */
1090 u
->val_scale
= atof(optval
);
1091 } else if (!strcasecmp(optname
, "val_points") && optval
) {
1092 /* Maximum size of win to be scaled into game
1093 * result value. Zero means boardsize^2. */
1094 u
->val_points
= atoi(optval
) * 2; // result values are doubled
1095 } else if (!strcasecmp(optname
, "val_extra")) {
1096 /* If false, the score coefficient will be simply
1097 * added to the value, instead of scaling the result
1098 * coefficient because of it. */
1099 u
->val_extra
= !optval
|| atoi(optval
);
1100 } else if (!strcasecmp(optname
, "local_tree") && optval
) {
1101 /* Whether to bias exploration by local tree values
1102 * (must be supported by the used policy).
1104 * 1: Do, value = result.
1105 * Try to temper the result:
1106 * 2: Do, value = 0.5+(result-expected)/2.
1107 * 3: Do, value = 0.5+bzz((result-expected)^2).
1108 * 4: Do, value = 0.5+sqrt(result-expected)/2. */
1109 u
->local_tree
= atoi(optval
);
1110 } else if (!strcasecmp(optname
, "tenuki_d") && optval
) {
1111 /* Tenuki distance at which to break the local tree. */
1112 u
->tenuki_d
= atoi(optval
);
1113 if (u
->tenuki_d
> TREE_NODE_D_MAX
+ 1) {
1114 fprintf(stderr
, "uct: tenuki_d must not be larger than TREE_NODE_D_MAX+1 %d\n", TREE_NODE_D_MAX
+ 1);
1117 } else if (!strcasecmp(optname
, "local_tree_aging") && optval
) {
1118 /* How much to reduce local tree values between moves. */
1119 u
->local_tree_aging
= atof(optval
);
1120 } else if (!strcasecmp(optname
, "local_tree_allseq")) {
1121 /* By default, only complete sequences are stored
1122 * in the local tree. If this is on, also
1123 * subsequences starting at each move are stored. */
1124 u
->local_tree_allseq
= !optval
|| atoi(optval
);
1125 } else if (!strcasecmp(optname
, "local_tree_playout")) {
1126 /* Whether to adjust ELO playout probability
1127 * distributions according to matched localtree
1129 u
->local_tree_playout
= !optval
|| atoi(optval
);
1130 } else if (!strcasecmp(optname
, "local_tree_pseqroot")) {
1131 /* By default, when we have no sequence move
1132 * to suggest in-playout, we give up. If this
1133 * is on, we make probability distribution from
1134 * sequences first moves instead. */
1135 u
->local_tree_pseqroot
= !optval
|| atoi(optval
);
1136 } else if (!strcasecmp(optname
, "pass_all_alive")) {
1137 /* Whether to consider all stones alive at the game
1138 * end instead of marking dead groupd. */
1139 u
->pass_all_alive
= !optval
|| atoi(optval
);
1140 } else if (!strcasecmp(optname
, "territory_scoring")) {
1141 /* Use territory scoring (default is area scoring).
1142 * An explicit kgs-rules command overrides this. */
1143 u
->territory_scoring
= !optval
|| atoi(optval
);
1144 } else if (!strcasecmp(optname
, "random_policy_chance") && optval
) {
1145 /* If specified (N), with probability 1/N, random_policy policy
1146 * descend is used instead of main policy descend; useful
1147 * if specified policy (e.g. UCB1AMAF) can make unduly biased
1148 * choices sometimes, you can fall back to e.g.
1149 * random_policy=UCB1. */
1150 u
->random_policy_chance
= atoi(optval
);
1151 } else if (!strcasecmp(optname
, "max_tree_size") && optval
) {
1152 /* Maximum amount of memory [MiB] consumed by the move tree.
1153 * For fast_alloc it includes the temp tree used for pruning.
1154 * Default is 3072 (3 GiB). Note that if you use TM_ROOT,
1155 * this limits size of only one of the trees, not all of them
1157 u
->max_tree_size
= atol(optval
) * 1048576;
1158 } else if (!strcasecmp(optname
, "fast_alloc")) {
1159 u
->fast_alloc
= !optval
|| atoi(optval
);
1160 } else if (!strcasecmp(optname
, "slave")) {
1161 /* Act as slave for the distributed engine. */
1162 u
->slave
= !optval
|| atoi(optval
);
1164 } else if (!strcasecmp(optname
, "banner") && optval
) {
1165 /* Additional banner string. This must come as the
1166 * last engine parameter. */
1167 if (*next
) *--next
= ',';
1168 u
->banner
= strdup(optval
);
1171 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
1177 u
->resign_ratio
= 0.2; /* Resign when most games are lost. */
1178 u
->loss_threshold
= 0.85; /* Stop reading if after at least 5000 playouts this is best value. */
1180 u
->policy
= policy_ucb1amaf_init(u
, NULL
);
1182 if (!!u
->random_policy_chance
^ !!u
->random_policy
) {
1183 fprintf(stderr
, "uct: Only one of random_policy and random_policy_chance is set\n");
1187 if (!u
->local_tree
) {
1188 /* No ltree aging. */
1189 u
->local_tree_aging
= 1.0f
;
1192 u
->local_tree_playout
= false;
1194 if (u
->fast_alloc
&& !u
->parallel_tree
) {
1195 fprintf(stderr
, "fast_alloc not supported with root parallelization.\n");
1199 u
->max_tree_size
= (100ULL * u
->max_tree_size
) / (100 + MIN_FREE_MEM_PERCENT
);
1202 u
->prior
= uct_prior_init(NULL
, b
);
1205 u
->playout
= playout_moggy_init(NULL
, b
);
1206 u
->playout
->debug_level
= u
->debug_level
;
1208 u
->ownermap
.map
= malloc(board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
1211 u
->dynkomi
= uct_dynkomi_init_linear(u
, NULL
, b
);
1213 /* Some things remain uninitialized for now - the opening book
1214 * is not loaded and the tree not set up. */
1215 /* This will be initialized in setup_state() at the first move
1216 * received/requested. This is because right now we are not aware
1217 * about any komi or handicap setup and such. */
1223 engine_uct_init(char *arg
, struct board
*b
)
1225 struct uct
*u
= uct_state_init(arg
, b
);
1226 struct engine
*e
= calloc(1, sizeof(struct engine
));
1227 e
->name
= "UCT Engine";
1228 e
->printhook
= uct_printhook_ownermap
;
1229 e
->notify_play
= uct_notify_play
;
1231 e
->genmove
= uct_genmove
;
1232 e
->genmoves
= uct_genmoves
;
1233 e
->dead_group_list
= uct_dead_group_list
;
1237 e
->notify
= uct_notify
;
1239 const char banner
[] = "I'm playing UCT. When I'm losing, I will resign, "
1240 "if I think I win, I play until you pass. "
1241 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
1242 if (!u
->banner
) u
->banner
= "";
1243 e
->comment
= malloc(sizeof(banner
) + strlen(u
->banner
) + 1);
1244 sprintf(e
->comment
, "%s %s", banner
, u
->banner
);