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
39 /* How big proportion of ownermap counts must be of one color to consider
42 /* How many games to consider at minimum before judging groups. */
43 #define GJ_MINGAMES 500
45 /* How often to inspect the tree from the main thread to check for playout
46 * stop, progress reports, etc. (in seconds) */
47 #define TREE_BUSYWAIT_INTERVAL 0.1 /* 100ms */
49 /* For safety, use at most 3 times the desired time on a single move
50 * in main time, and 1.1 times in byoyomi. */
51 #define MAX_MAIN_TIME_EXTENSION 3.0
52 #define MAX_BYOYOMI_TIME_EXTENSION 1.1
54 /* Once per how many simulations (per thread) to show a progress report line. */
55 #define TREE_SIMPROGRESS_INTERVAL 10000
59 setup_state(struct uct
*u
, struct board
*b
, enum stone color
)
61 u
->t
= tree_init(b
, color
);
63 fast_srandom(u
->force_seed
);
65 fprintf(stderr
, "Fresh board with random seed %lu\n", fast_getseed());
66 //board_print(b, stderr);
67 if (!u
->no_book
&& b
->moves
== 0) {
68 assert(color
== S_BLACK
);
74 reset_state(struct uct
*u
)
77 tree_done(u
->t
); u
->t
= NULL
;
81 prepare_move(struct engine
*e
, struct board
*b
, enum stone color
)
83 struct uct
*u
= e
->data
;
86 /* Verify that we have sane state. */
88 assert(u
->t
&& b
->moves
);
89 if (color
!= stone_other(u
->t
->root_color
)) {
90 fprintf(stderr
, "Fatal: Non-alternating play detected %d %d\n",
91 color
, u
->t
->root_color
);
96 /* We need fresh state. */
98 setup_state(u
, b
, color
);
101 if (u
->dynkomi
&& u
->dynkomi
> b
->moves
&& (color
& u
->dynkomi_mask
))
102 u
->t
->extra_komi
= uct_get_extra_komi(u
, b
);
104 u
->ownermap
.playouts
= 0;
105 memset(u
->ownermap
.map
, 0, board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
109 dead_group_list(struct uct
*u
, struct board
*b
, struct move_queue
*mq
)
111 struct group_judgement gj
;
113 gj
.gs
= alloca(board_size2(b
) * sizeof(gj
.gs
[0]));
114 board_ownermap_judge_group(b
, &u
->ownermap
, &gj
);
115 groups_of_status(b
, &gj
, GS_DEAD
, mq
);
119 uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
)
121 if (u
->ownermap
.playouts
< GJ_MINGAMES
)
124 struct move_queue mq
= { .moves
= 0 };
126 dead_group_list(u
, b
, &mq
);
127 return pass_is_safe(b
, color
, &mq
);
132 uct_printhook_ownermap(struct board
*board
, coord_t c
, FILE *f
)
134 struct uct
*u
= board
->es
;
136 const char chr
[] = ":XO,"; // dame, black, white, unclear
137 const char chm
[] = ":xo,";
138 char ch
= chr
[board_ownermap_judge_point(&u
->ownermap
, c
, GJ_THRES
)];
139 if (ch
== ',') { // less precise estimate then?
140 ch
= chm
[board_ownermap_judge_point(&u
->ownermap
, c
, 0.67)];
142 fprintf(f
, "%c ", ch
);
146 uct_notify_play(struct engine
*e
, struct board
*b
, struct move
*m
)
148 struct uct
*u
= e
->data
;
150 /* No state, create one - this is probably game beginning
151 * and we need to load the opening book right now. */
152 prepare_move(e
, b
, m
->color
);
156 /* Stop pondering. */
157 /* XXX: If we are about to receive multiple 'play' commands,
158 * e.g. in a rengo, we will not ponder during the rest of them. */
159 uct_pondering_stop(u
);
161 if (is_resign(m
->coord
)) {
167 /* Promote node of the appropriate move to the tree root. */
169 if (!tree_promote_at(u
->t
, b
, m
->coord
)) {
171 fprintf(stderr
, "Warning: Cannot promote move node! Several play commands in row?\n");
180 uct_chat(struct engine
*e
, struct board
*b
, char *cmd
)
182 struct uct
*u
= e
->data
;
183 static char reply
[1024];
185 cmd
+= strspn(cmd
, " \n\t");
186 if (!strncasecmp(cmd
, "winrate", 7)) {
188 return "no game context (yet?)";
189 enum stone color
= u
->t
->root_color
;
190 struct tree_node
*n
= u
->t
->root
;
191 snprintf(reply
, 1024, "In %d playouts at %d threads, %s %s can win with %.2f%% probability",
192 n
->u
.playouts
, u
->threads
, stone2str(color
), coord2sstr(n
->coord
, b
),
193 tree_node_get_value(u
->t
, -1, n
->u
.value
) * 100);
194 if (abs(u
->t
->extra_komi
) >= 0.5) {
195 sprintf(reply
+ strlen(reply
), ", while self-imposing extra komi %.1f",
205 uct_dead_group_list(struct engine
*e
, struct board
*b
, struct move_queue
*mq
)
207 struct uct
*u
= e
->data
;
209 /* This means the game is probably over, no use pondering on. */
210 uct_pondering_stop(u
);
212 if (u
->pass_all_alive
)
213 return; // no dead groups
215 bool mock_state
= false;
218 /* No state, but we cannot just back out - we might
219 * have passed earlier, only assuming some stones are
220 * dead, and then re-connected, only to lose counting
221 * when all stones are assumed alive. */
222 /* Mock up some state and seed the ownermap by few
224 prepare_move(e
, b
, S_BLACK
); assert(u
->t
);
225 for (int i
= 0; i
< GJ_MINGAMES
; i
++)
226 uct_playout(u
, b
, S_BLACK
, u
->t
);
230 dead_group_list(u
, b
, mq
);
233 /* Clean up the mock state in case we will receive
234 * a genmove; we could get a non-alternating-move
235 * error from prepare_move() in that case otherwise. */
241 playout_policy_done(struct playout_policy
*p
)
243 if (p
->done
) p
->done(p
);
244 if (p
->data
) free(p
->data
);
249 uct_done(struct engine
*e
)
251 /* This is called on engine reset, especially when clear_board
252 * is received and new game should begin. */
253 struct uct
*u
= e
->data
;
254 uct_pondering_stop(u
);
255 if (u
->t
) reset_state(u
);
256 free(u
->ownermap
.map
);
259 free(u
->random_policy
);
260 playout_policy_done(u
->playout
);
261 uct_prior_done(u
->prior
);
265 /* Pachi threading structure (if uct_playouts_parallel() is used):
268 * | main(), GTP communication, ...
269 * | starts and stops the search managed by thread_manager
272 * | spawns and collects worker threads
278 * uct_playouts() loop, doing descend-playout until uct_halt
280 * Another way to look at it is by functions (lines denote thread boundaries):
283 * | uct_search() (uct_search_start() .. uct_search_stop())
284 * | -----------------------
285 * | spawn_thread_manager()
286 * | -----------------------
288 * V uct_playouts() */
290 /* Set in thread manager in case the workers should stop. */
291 volatile sig_atomic_t uct_halt
= 0;
292 /* ID of the running worker thread. */
293 __thread
int thread_id
= -1;
294 /* ID of the thread manager. */
295 static pthread_t thread_manager
;
296 static bool thread_manager_running
;
298 static pthread_mutex_t finish_mutex
= PTHREAD_MUTEX_INITIALIZER
;
299 static pthread_cond_t finish_cond
= PTHREAD_COND_INITIALIZER
;
300 static volatile int finish_thread
;
301 static pthread_mutex_t finish_serializer
= PTHREAD_MUTEX_INITIALIZER
;
314 spawn_worker(void *ctx_
)
316 struct spawn_ctx
*ctx
= ctx_
;
318 fast_srandom(ctx
->seed
);
319 thread_id
= ctx
->tid
;
321 ctx
->games
= uct_playouts(ctx
->u
, ctx
->b
, ctx
->color
, ctx
->t
);
323 pthread_mutex_lock(&finish_serializer
);
324 pthread_mutex_lock(&finish_mutex
);
325 finish_thread
= ctx
->tid
;
326 pthread_cond_signal(&finish_cond
);
327 pthread_mutex_unlock(&finish_mutex
);
331 /* Thread manager, controlling worker threads. It must be called with
332 * finish_mutex lock held, but it will unlock it itself before exiting;
333 * this is necessary to be completely deadlock-free. */
334 /* The finish_cond can be signalled for it to stop; in that case,
335 * the caller should set finish_thread = -1. */
336 /* After it is started, it will update mctx->t to point at some tree
337 * used for the actual search (matters only for TM_ROOT), on return
338 * it will set mctx->games to the number of performed simulations. */
340 spawn_thread_manager(void *ctx_
)
342 /* In thread_manager, we use only some of the ctx fields. */
343 struct spawn_ctx
*mctx
= ctx_
;
344 struct uct
*u
= mctx
->u
;
345 struct tree
*t
= mctx
->t
;
346 bool shared_tree
= u
->parallel_tree
;
347 fast_srandom(mctx
->seed
);
349 int played_games
= 0;
350 pthread_t threads
[u
->threads
];
355 /* Spawn threads... */
356 for (int ti
= 0; ti
< u
->threads
; ti
++) {
357 struct spawn_ctx
*ctx
= malloc(sizeof(*ctx
));
358 ctx
->u
= u
; ctx
->b
= mctx
->b
; ctx
->color
= mctx
->color
;
359 mctx
->t
= ctx
->t
= shared_tree
? t
: tree_copy(t
);
360 ctx
->tid
= ti
; ctx
->seed
= fast_random(65536) + ti
;
361 pthread_create(&threads
[ti
], NULL
, spawn_worker
, ctx
);
363 fprintf(stderr
, "Spawned worker %d\n", ti
);
366 /* ...and collect them back: */
367 while (joined
< u
->threads
) {
368 /* Wait for some thread to finish... */
369 pthread_cond_wait(&finish_cond
, &finish_mutex
);
370 if (finish_thread
< 0) {
371 /* Stop-by-caller. Tell the workers to wrap up. */
375 /* ...and gather its remnants. */
376 struct spawn_ctx
*ctx
;
377 pthread_join(threads
[finish_thread
], (void **) &ctx
);
378 played_games
+= ctx
->games
;
381 if (ctx
->t
== mctx
->t
) mctx
->t
= t
;
382 tree_merge(t
, ctx
->t
);
387 fprintf(stderr
, "Joined worker %d\n", finish_thread
);
388 pthread_mutex_unlock(&finish_serializer
);
391 pthread_mutex_unlock(&finish_mutex
);
394 tree_normalize(mctx
->t
, u
->threads
);
396 mctx
->games
= played_games
;
400 static struct spawn_ctx
*
401 uct_search_start(struct uct
*u
, struct board
*b
, enum stone color
, struct tree
*t
)
403 assert(u
->threads
> 0);
404 assert(!thread_manager_running
);
406 struct spawn_ctx ctx
= { .u
= u
, .b
= b
, .color
= color
, .t
= t
, .seed
= fast_random(65536) };
407 static struct spawn_ctx mctx
; mctx
= ctx
;
408 pthread_mutex_lock(&finish_mutex
);
409 pthread_create(&thread_manager
, NULL
, spawn_thread_manager
, &mctx
);
410 thread_manager_running
= true;
414 static struct spawn_ctx
*
415 uct_search_stop(void)
417 assert(thread_manager_running
);
419 /* Signal thread manager to stop the workers. */
420 pthread_mutex_lock(&finish_mutex
);
422 pthread_cond_signal(&finish_cond
);
423 pthread_mutex_unlock(&finish_mutex
);
425 /* Collect the thread manager. */
426 struct spawn_ctx
*pctx
;
427 thread_manager_running
= false;
428 pthread_join(thread_manager
, (void **) &pctx
);
433 /* Search stopping conditions */
434 union stop_conditions
{
435 struct { // TD_WALLTIME
436 double desired_stop
; /* stop at that time if possible */
437 double worst_stop
; /* stop no later than this */
440 int desired_playouts
;
445 /* Pre-process time_info for search control and sets the desired stopping conditions. */
447 time_prep(struct time_info
*ti
, struct uct
*u
, struct board
*b
, union stop_conditions
*stop
)
449 assert(ti
->period
!= TT_TOTAL
);
451 if (ti
->period
== TT_NULL
) {
452 ti
->period
= TT_MOVE
;
454 ti
->len
.games
= MC_GAMES
;
456 if (ti
->dim
== TD_GAMES
) {
457 stop
->p
.desired_playouts
= ti
->len
.games
;
458 stop
->p
.worst_playouts
= ti
->len
.games
* MAX_MAIN_TIME_EXTENSION
;
460 double desired_time
= ti
->len
.t
.recommended_time
;
462 if (time_in_byoyomi(ti
)) {
463 worst_time
= desired_time
* MAX_BYOYOMI_TIME_EXTENSION
;
464 desired_time
*= (2 - MAX_BYOYOMI_TIME_EXTENSION
); // make average(desired, worst) == recommended
466 int bsize
= (board_size(b
)-2)*(board_size(b
)-2);
467 int fuseki_end
= u
->fuseki_end
* bsize
/ 100; // move nb at fuseki end
468 int yose_start
= u
->yose_start
* bsize
/ 100; // move nb at yose start
470 int left_at_yose_start
= (b
->moves
- yose_start
) / 2 + board_estimated_moves_left(b
);
471 /* /2 because we only consider the moves we have to play ourselves */
472 if (left_at_yose_start
< MIN_MOVES_LEFT
)
473 left_at_yose_start
= MIN_MOVES_LEFT
;
474 double longest_time
= ti
->len
.t
.max_time
/ left_at_yose_start
;
475 if (longest_time
< desired_time
) {
476 // Should rarely happen, but keep desired_time anyway
477 } else if (b
->moves
< fuseki_end
) {
478 desired_time
+= ((longest_time
- desired_time
) * b
->moves
) / fuseki_end
;
479 /* In this branch fuseki_end can't be 0 */
480 } else if (b
->moves
< yose_start
) {
481 desired_time
= longest_time
;
483 worst_time
= desired_time
* MAX_MAIN_TIME_EXTENSION
;
485 if (worst_time
> ti
->len
.t
.max_time
)
486 worst_time
= ti
->len
.t
.max_time
;
487 if (desired_time
> worst_time
)
488 desired_time
= worst_time
;
490 stop
->t
.desired_stop
= ti
->len
.t
.timer_start
+ desired_time
- ti
->len
.t
.net_lag
;
491 stop
->t
.worst_stop
= ti
->len
.t
.timer_start
+ worst_time
- ti
->len
.t
.net_lag
;
492 // Both stop points may be in the past if too much lag.
495 fprintf(stderr
, "desired time %.02f, worst %.02f\n", desired_time
, worst_time
);
499 /* Run time-limited MCTS search on foreground. */
501 uct_search(struct uct
*u
, struct board
*b
, struct time_info
*ti
, enum stone color
, struct tree
*t
)
503 union stop_conditions stop
;
504 time_prep(ti
, u
, b
, &stop
);
505 if (UDEBUGL(2) && u
->t
->root
->u
.playouts
> 0)
506 fprintf(stderr
, "<pre-simulated %d games skipped>\n", u
->t
->root
->u
.playouts
);
508 /* Number of last game with progress print. */
509 int last_print
= t
->root
->u
.playouts
;
510 /* Number of simulations to wait before next print. */
511 int print_interval
= TREE_SIMPROGRESS_INTERVAL
* (u
->thread_model
== TM_ROOT
? 1 : u
->threads
);
512 /* Printed notification about full memory? */
513 bool print_fullmem
= false;
515 struct spawn_ctx
*ctx
= uct_search_start(u
, b
, color
, t
);
517 /* The search tree is ctx->t. This is normally == t, but in case of
518 * TM_ROOT, it is one of the trees belonging to the independent
519 * workers. It is important to reference ctx->t directly since the
520 * thread manager will swap the tree pointer asynchronously. */
521 /* XXX: This means TM_ROOT support is suboptimal since single stalled
522 * thread can stall the others in case of limiting the search by game
523 * count. However, TM_ROOT just does not deserve any more extra code
526 struct tree_node
*best
= NULL
, *prev_best
;
527 struct tree_node
*winner
= NULL
, *prev_winner
;
529 double busywait_interval
= TREE_BUSYWAIT_INTERVAL
;
531 /* Now, just periodically poll the search tree. */
533 time_sleep(busywait_interval
);
534 /* busywait_interval should never be less than desired time, or the
535 * time control is broken. But if it happens to be less, we still search
536 * at least 100ms otherwise the move is completely random. */
538 int i
= ctx
->t
->root
->u
.playouts
;
540 /* Print progress? */
541 if (i
- last_print
> print_interval
) {
542 last_print
+= print_interval
; // keep the numbers tidy
543 uct_progress_status(u
, ctx
->t
, color
, last_print
);
545 if (!print_fullmem
&& ctx
->t
->nodes_size
> u
->max_tree_size
) {
547 fprintf(stderr
, "memory limit hit (%ld > %lu)\n", ctx
->t
->nodes_size
, u
->max_tree_size
);
548 print_fullmem
= true;
551 /* Check against time settings. */
552 bool desired_done
= false;
553 if (ti
->dim
== TD_WALLTIME
) {
554 double now
= time_now();
555 if (now
> stop
.t
.worst_stop
) break;
556 desired_done
= now
> stop
.t
.desired_stop
;
558 assert(ti
->dim
== TD_GAMES
);
559 if (i
> stop
.p
.worst_playouts
) break;
560 desired_done
= i
> stop
.p
.desired_playouts
;
563 /* Early break in won situation. */
565 best
= u
->policy
->choose(u
->policy
, ctx
->t
->root
, b
, color
);
566 if (best
&& ((best
->u
.playouts
>= 2000 && tree_node_get_value(ctx
->t
, 1, best
->u
.value
) >= u
->loss_threshold
)
567 || (best
->u
.playouts
>= 500 && tree_node_get_value(ctx
->t
, 1, best
->u
.value
) >= 0.95)))
571 if (!u
->policy
->winner
|| !u
->policy
->evaluate
)
573 /* Stop only if best explored has also highest value: */
574 prev_winner
= winner
;
575 winner
= u
->policy
->winner(u
->policy
, ctx
->t
, ctx
->t
->root
);
576 if (best
&& best
== winner
)
578 if (UDEBUGL(3) && (best
!= prev_best
|| winner
!= prev_winner
)) {
579 fprintf(stderr
, "[%d] best", i
);
581 fprintf(stderr
, " %3s [%d] %f", coord2sstr(best
->coord
, ctx
->t
->board
),
582 best
->u
.playouts
, tree_node_get_value(ctx
->t
, 1, best
->u
.value
));
583 fprintf(stderr
, " != winner");
585 fprintf(stderr
, " %3s [%d] %f ", coord2sstr(winner
->coord
, ctx
->t
->board
),
586 winner
->u
.playouts
, tree_node_get_value(ctx
->t
, 1, winner
->u
.value
));
587 fprintf(stderr
, "\n");
591 /* TODO: Early break if best->variance goes under threshold and we already
592 * have enough playouts (possibly thanks to book or to pondering). */
593 /* TODO: Early break if second best has no chance to catch up. */
596 ctx
= uct_search_stop();
599 tree_dump(t
, u
->dumpthres
);
601 uct_progress_status(u
, t
, color
, ctx
->games
);
607 /* Start pondering background with @color to play. */
609 uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
)
612 fprintf(stderr
, "Starting to ponder with color %s\n", stone2str(stone_other(color
)));
614 /* We need a local board copy to ponder upon. */
615 struct board
*b
= malloc(sizeof(*b
)); board_copy(b
, b0
);
617 /* *b0 did not have the genmove'd move played yet. */
618 struct move m
= { t
->root
->coord
, t
->root_color
};
619 int res
= board_play(b
, &m
);
622 /* Start MCTS manager thread "headless". */
623 uct_search_start(u
, b
, color
, t
);
626 /* uct_search_stop() frontend for the pondering (non-genmove) mode. */
628 uct_pondering_stop(struct uct
*u
)
630 if (!thread_manager_running
)
633 /* Stop the thread manager. */
634 struct spawn_ctx
*ctx
= uct_search_stop();
636 fprintf(stderr
, "(pondering) ");
637 uct_progress_status(u
, ctx
->t
, ctx
->color
, ctx
->games
);
644 uct_genmove(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
646 double start_time
= time_now();
647 struct uct
*u
= e
->data
;
649 if (b
->superko_violation
) {
650 fprintf(stderr
, "!!! WARNING: SUPERKO VIOLATION OCCURED BEFORE THIS MOVE\n");
651 fprintf(stderr
, "Maybe you play with situational instead of positional superko?\n");
652 fprintf(stderr
, "I'm going to ignore the violation, but note that I may miss\n");
653 fprintf(stderr
, "some moves valid under this ruleset because of this.\n");
654 b
->superko_violation
= false;
658 uct_pondering_stop(u
);
659 prepare_move(e
, b
, color
);
662 /* Perform the Monte Carlo Tree Search! */
663 int played_games
= uct_search(u
, b
, ti
, color
, u
->t
);
665 /* Choose the best move from the tree. */
666 struct tree_node
*best
= u
->policy
->choose(u
->policy
, u
->t
->root
, b
, color
);
669 return coord_copy(pass
);
672 fprintf(stderr
, "*** WINNER is %s (%d,%d) with score %1.4f (%d/%d:%d games)\n",
673 coord2sstr(best
->coord
, b
), coord_x(best
->coord
, b
), coord_y(best
->coord
, b
),
674 tree_node_get_value(u
->t
, 1, best
->u
.value
),
675 best
->u
.playouts
, u
->t
->root
->u
.playouts
, played_games
);
676 if (tree_node_get_value(u
->t
, 1, best
->u
.value
) < u
->resign_ratio
&& !is_pass(best
->coord
)) {
678 return coord_copy(resign
);
681 /* If the opponent just passed and we win counting, always
683 if (b
->moves
> 1 && is_pass(b
->last_move
.coord
)) {
684 /* Make sure enough playouts are simulated. */
685 while (u
->ownermap
.playouts
< GJ_MINGAMES
)
686 uct_playout(u
, b
, color
, u
->t
);
687 if (uct_pass_is_safe(u
, b
, color
, u
->pass_all_alive
|| pass_all_alive
)) {
689 fprintf(stderr
, "<Will rather pass, looks safe enough.>\n");
694 tree_promote_node(u
->t
, best
);
695 /* After a pass, pondering is harmful for two reasons:
696 * (i) We might keep pondering even when the game is over.
697 * Of course this is the case for opponent resign as well.
698 * (ii) More importantly, the ownermap will get skewed since
699 * the UCT will start cutting off any playouts. */
700 if (u
->pondering
&& !is_pass(best
->coord
)) {
701 uct_pondering_start(u
, b
, u
->t
, stone_other(color
));
704 double time
= time_now() - start_time
+ 0.000001; /* avoid divide by zero */
705 fprintf(stderr
, "genmove in %0.2fs (%d games/s, %d games/s/thread)\n",
706 time
, (int)(played_games
/time
), (int)(played_games
/time
/u
->threads
));
708 return coord_copy(best
->coord
);
713 uct_genbook(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
)
715 struct uct
*u
= e
->data
;
716 if (!u
->t
) prepare_move(e
, b
, color
);
719 if (ti
->dim
== TD_GAMES
) {
720 /* Don't count in games that already went into the book. */
721 ti
->len
.games
+= u
->t
->root
->u
.playouts
;
723 uct_search(u
, b
, ti
, color
, u
->t
);
725 assert(ti
->dim
== TD_GAMES
);
726 tree_save(u
->t
, b
, ti
->len
.games
/ 100);
732 uct_dumpbook(struct engine
*e
, struct board
*b
, enum stone color
)
734 struct tree
*t
= tree_init(b
, color
);
742 uct_state_init(char *arg
, struct board
*b
)
744 struct uct
*u
= calloc(1, sizeof(struct uct
));
747 u
->gamelen
= MC_GAMELEN
;
751 u
->playout_amaf
= true;
752 u
->playout_amaf_nakade
= false;
753 u
->amaf_prior
= false;
754 u
->max_tree_size
= 3072ULL * 1048576;
756 if (board_size(b
) - 2 >= 19)
758 u
->dynkomi_mask
= S_BLACK
;
761 u
->thread_model
= TM_TREEVL
;
762 u
->parallel_tree
= true;
763 u
->virtual_loss
= true;
764 u
->fuseki_end
= 20; // max time at 361*20% = 72 moves (our 36th move, still 99 to play)
765 u
->yose_start
= 40; // (100-40-25)*361/100/2 = 63 moves still to play by us then
767 u
->val_scale
= 0.02; u
->val_points
= 20;
770 char *optspec
, *next
= arg
;
773 next
+= strcspn(next
, ",");
774 if (*next
) { *next
++ = 0; } else { *next
= 0; }
776 char *optname
= optspec
;
777 char *optval
= strchr(optspec
, '=');
778 if (optval
) *optval
++ = 0;
780 if (!strcasecmp(optname
, "debug")) {
782 u
->debug_level
= atoi(optval
);
785 } else if (!strcasecmp(optname
, "mercy") && optval
) {
786 /* Minimal difference of black/white captures
787 * to stop playout - "Mercy Rule". Speeds up
788 * hopeless playouts at the expense of some
790 u
->mercymin
= atoi(optval
);
791 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
792 u
->gamelen
= atoi(optval
);
793 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
794 u
->expand_p
= atoi(optval
);
795 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
796 u
->dumpthres
= atoi(optval
);
797 } else if (!strcasecmp(optname
, "playout_amaf")) {
798 /* Whether to include random playout moves in
799 * AMAF as well. (Otherwise, only tree moves
800 * are included in AMAF. Of course makes sense
801 * only in connection with an AMAF policy.) */
802 /* with-without: 55.5% (+-4.1) */
803 if (optval
&& *optval
== '0')
804 u
->playout_amaf
= false;
806 u
->playout_amaf
= true;
807 } else if (!strcasecmp(optname
, "playout_amaf_nakade")) {
808 /* Whether to include nakade moves from playouts
809 * in the AMAF statistics; this tends to nullify
810 * the playout_amaf effect by adding too much
812 if (optval
&& *optval
== '0')
813 u
->playout_amaf_nakade
= false;
815 u
->playout_amaf_nakade
= true;
816 } else if (!strcasecmp(optname
, "playout_amaf_cutoff") && optval
) {
817 /* Keep only first N% of playout stage AMAF
819 u
->playout_amaf_cutoff
= atoi(optval
);
820 } else if ((!strcasecmp(optname
, "policy") || !strcasecmp(optname
, "random_policy")) && optval
) {
821 char *policyarg
= strchr(optval
, ':');
822 struct uct_policy
**p
= !strcasecmp(optname
, "policy") ? &u
->policy
: &u
->random_policy
;
825 if (!strcasecmp(optval
, "ucb1")) {
826 *p
= policy_ucb1_init(u
, policyarg
);
827 } else if (!strcasecmp(optval
, "ucb1amaf")) {
828 *p
= policy_ucb1amaf_init(u
, policyarg
);
830 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
833 } else if (!strcasecmp(optname
, "playout") && optval
) {
834 char *playoutarg
= strchr(optval
, ':');
837 if (!strcasecmp(optval
, "moggy")) {
838 u
->playout
= playout_moggy_init(playoutarg
);
839 } else if (!strcasecmp(optval
, "light")) {
840 u
->playout
= playout_light_init(playoutarg
);
841 } else if (!strcasecmp(optval
, "elo")) {
842 u
->playout
= playout_elo_init(playoutarg
);
844 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
847 } else if (!strcasecmp(optname
, "prior") && optval
) {
848 u
->prior
= uct_prior_init(optval
, b
);
849 } else if (!strcasecmp(optname
, "amaf_prior") && optval
) {
850 u
->amaf_prior
= atoi(optval
);
851 } else if (!strcasecmp(optname
, "threads") && optval
) {
852 /* By default, Pachi will run with only single
853 * tree search thread! */
854 u
->threads
= atoi(optval
);
855 } else if (!strcasecmp(optname
, "thread_model") && optval
) {
856 if (!strcasecmp(optval
, "root")) {
857 /* Root parallelization - each thread
858 * does independent search, trees are
859 * merged at the end. */
860 u
->thread_model
= TM_ROOT
;
861 u
->parallel_tree
= false;
862 u
->virtual_loss
= false;
863 } else if (!strcasecmp(optval
, "tree")) {
864 /* Tree parallelization - all threads
865 * grind on the same tree. */
866 u
->thread_model
= TM_TREE
;
867 u
->parallel_tree
= true;
868 u
->virtual_loss
= false;
869 } else if (!strcasecmp(optval
, "treevl")) {
870 /* Tree parallelization, but also
871 * with virtual losses - this discou-
872 * rages most threads choosing the
873 * same tree branches to read. */
874 u
->thread_model
= TM_TREEVL
;
875 u
->parallel_tree
= true;
876 u
->virtual_loss
= true;
878 fprintf(stderr
, "UCT: Invalid thread model %s\n", optval
);
881 } else if (!strcasecmp(optname
, "pondering")) {
882 /* Keep searching even during opponent's turn. */
883 u
->pondering
= !optval
|| atoi(optval
);
884 } else if (!strcasecmp(optname
, "fuseki_end") && optval
) {
885 /* At the very beginning it's not worth thinking too long because the
886 * playout evaluations are very noisy. So gradually increase the thinking
887 * time up to maximum when fuseki_end percent of the board has been played.
888 * This only applies if we are not in byoyomi. */
889 u
->fuseki_end
= atoi(optval
);
890 } else if (!strcasecmp(optname
, "yose_start") && optval
) {
891 /* When yose_start percent of the board has been played, or if we are in
892 * byoyomi, stop spending more time and spread the remaining time uniformly.
893 * Between fuseki_end and yose_start, we spend on each move a constant
894 * proportion of the remaining time. (yose_start should actually be much
895 * earlier than when real yose start, but "yose" is a good short name to
896 * convey the idea.) */
897 u
->yose_start
= atoi(optval
);
898 } else if (!strcasecmp(optname
, "force_seed") && optval
) {
899 u
->force_seed
= atoi(optval
);
900 } else if (!strcasecmp(optname
, "no_book")) {
902 } else if (!strcasecmp(optname
, "dynkomi")) {
903 /* Dynamic komi in handicap game; linearly
904 * decreases to basic settings until move
906 u
->dynkomi
= optval
? atoi(optval
) : 150;
907 } else if (!strcasecmp(optname
, "dynkomi_mask") && optval
) {
908 /* Bitmask of colors the player must be
909 * for dynkomi be applied; you may want
910 * to use dynkomi_mask=3 to allow dynkomi
911 * even in games where Pachi is white. */
912 u
->dynkomi_mask
= atoi(optval
);
913 } else if (!strcasecmp(optname
, "val_scale") && optval
) {
914 /* How much of the game result value should be
915 * influenced by win size. Zero means it isn't. */
916 u
->val_scale
= atof(optval
);
917 } else if (!strcasecmp(optname
, "val_points") && optval
) {
918 /* Maximum size of win to be scaled into game
919 * result value. Zero means boardsize^2. */
920 u
->val_points
= atoi(optval
) * 2; // result values are doubled
921 } else if (!strcasecmp(optname
, "val_extra")) {
922 /* If false, the score coefficient will be simply
923 * added to the value, instead of scaling the result
924 * coefficient because of it. */
925 u
->val_extra
= !optval
|| atoi(optval
);
926 } else if (!strcasecmp(optname
, "root_heuristic") && optval
) {
927 /* Whether to bias exploration by root node values
928 * (must be supported by the used policy).
930 * 1: Do, value = result.
931 * Try to temper the result:
932 * 2: Do, value = 0.5+(result-expected)/2.
933 * 3: Do, value = 0.5+bzz((result-expected)^2). */
934 u
->root_heuristic
= atoi(optval
);
935 } else if (!strcasecmp(optname
, "pass_all_alive")) {
936 /* Whether to consider all stones alive at the game
937 * end instead of marking dead groupd. */
938 u
->pass_all_alive
= !optval
|| atoi(optval
);
939 } else if (!strcasecmp(optname
, "random_policy_chance") && optval
) {
940 /* If specified (N), with probability 1/N, random_policy policy
941 * descend is used instead of main policy descend; useful
942 * if specified policy (e.g. UCB1AMAF) can make unduly biased
943 * choices sometimes, you can fall back to e.g.
944 * random_policy=UCB1. */
945 u
->random_policy_chance
= atoi(optval
);
946 } else if (!strcasecmp(optname
, "max_tree_size") && optval
) {
947 /* Maximum amount of memory [MiB] consumed by the move tree.
948 * Default is 3072 (3 GiB). Note that if you use TM_ROOT,
949 * this limits size of only one of the trees, not all of them
951 u
->max_tree_size
= atol(optval
) * 1048576;
952 } else if (!strcasecmp(optname
, "banner") && optval
) {
953 /* Additional banner string. This must come as the
954 * last engine parameter. */
955 if (*next
) *--next
= ',';
956 u
->banner
= strdup(optval
);
959 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
965 u
->resign_ratio
= 0.2; /* Resign when most games are lost. */
966 u
->loss_threshold
= 0.85; /* Stop reading if after at least 5000 playouts this is best value. */
968 u
->policy
= policy_ucb1amaf_init(u
, NULL
);
970 if (!!u
->random_policy_chance
^ !!u
->random_policy
) {
971 fprintf(stderr
, "uct: Only one of random_policy and random_policy_chance is set\n");
976 u
->prior
= uct_prior_init(NULL
, b
);
979 u
->playout
= playout_moggy_init(NULL
);
980 u
->playout
->debug_level
= u
->debug_level
;
982 u
->ownermap
.map
= malloc(board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
984 /* Some things remain uninitialized for now - the opening book
985 * is not loaded and the tree not set up. */
986 /* This will be initialized in setup_state() at the first move
987 * received/requested. This is because right now we are not aware
988 * about any komi or handicap setup and such. */
994 engine_uct_init(char *arg
, struct board
*b
)
996 struct uct
*u
= uct_state_init(arg
, b
);
997 struct engine
*e
= calloc(1, sizeof(struct engine
));
998 e
->name
= "UCT Engine";
999 e
->printhook
= uct_printhook_ownermap
;
1000 e
->notify_play
= uct_notify_play
;
1002 e
->genmove
= uct_genmove
;
1003 e
->dead_group_list
= uct_dead_group_list
;
1007 const char banner
[] = "I'm playing UCT. When I'm losing, I will resign, "
1008 "if I think I win, I play until you pass. "
1009 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
1010 if (!u
->banner
) u
->banner
= "";
1011 e
->comment
= malloc(sizeof(banner
) + strlen(u
->banner
) + 1);
1012 sprintf(e
->comment
, "%s %s", banner
, u
->banner
);