17 #include "playout/elo.h"
18 #include "playout/moggy.h"
19 #include "playout/light.h"
23 #include "uct/internal.h"
24 #include "uct/prior.h"
29 struct uct_policy
*policy_ucb1_init(struct uct
*u
, char *arg
);
30 struct uct_policy
*policy_ucb1amaf_init(struct uct
*u
, char *arg
);
31 static void uct_pondering_stop(struct uct
*u
);
34 /* Default number of simulations to perform per move.
35 * Note that this is now in total over all threads! (Unless TM_ROOT.) */
36 #define MC_GAMES 80000
37 #define MC_GAMELEN MAX_GAMELEN
38 static const struct time_info default_ti
= {
41 .len
= { .games
= MC_GAMES
},
44 /* How big proportion of ownermap counts must be of one color to consider
47 /* How many games to consider at minimum before judging groups. */
48 #define GJ_MINGAMES 500
50 /* How often to inspect the tree from the main thread to check for playout
51 * stop, progress reports, etc. (in seconds) */
52 #define TREE_BUSYWAIT_INTERVAL 0.1 /* 100ms */
54 /* Once per how many simulations (per thread) to show a progress report line. */
55 #define TREE_SIMPROGRESS_INTERVAL 10000
59 setup_state(struct uct
*u
, struct board
*b
, enum stone color
)
61 u
->t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0);
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 setup_dynkomi(struct uct
*u
, struct board
*b
, enum stone to_play
)
83 if (u
->dynkomi
> b
->moves
&& u
->t
->use_extra_komi
)
84 u
->t
->extra_komi
= uct_get_extra_komi(u
, b
);
90 prepare_move(struct engine
*e
, struct board
*b
, enum stone color
)
92 struct uct
*u
= e
->data
;
95 /* Verify that we have sane state. */
97 assert(u
->t
&& b
->moves
);
98 if (color
!= stone_other(u
->t
->root_color
)) {
99 fprintf(stderr
, "Fatal: Non-alternating play detected %d %d\n",
100 color
, u
->t
->root_color
);
105 /* We need fresh state. */
107 setup_state(u
, b
, color
);
110 u
->ownermap
.playouts
= 0;
111 memset(u
->ownermap
.map
, 0, board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
115 dead_group_list(struct uct
*u
, struct board
*b
, struct move_queue
*mq
)
117 struct group_judgement gj
;
119 gj
.gs
= alloca(board_size2(b
) * sizeof(gj
.gs
[0]));
120 board_ownermap_judge_group(b
, &u
->ownermap
, &gj
);
121 groups_of_status(b
, &gj
, GS_DEAD
, mq
);
125 uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
)
127 if (u
->ownermap
.playouts
< GJ_MINGAMES
)
130 struct move_queue mq
= { .moves
= 0 };
132 dead_group_list(u
, b
, &mq
);
133 return pass_is_safe(b
, color
, &mq
);
138 uct_printhook_ownermap(struct board
*board
, coord_t c
, FILE *f
)
140 struct uct
*u
= board
->es
;
142 const char chr
[] = ":XO,"; // dame, black, white, unclear
143 const char chm
[] = ":xo,";
144 char ch
= chr
[board_ownermap_judge_point(&u
->ownermap
, c
, GJ_THRES
)];
145 if (ch
== ',') { // less precise estimate then?
146 ch
= chm
[board_ownermap_judge_point(&u
->ownermap
, c
, 0.67)];
148 fprintf(f
, "%c ", ch
);
152 uct_notify_play(struct engine
*e
, struct board
*b
, struct move
*m
)
154 struct uct
*u
= e
->data
;
156 /* No state, create one - this is probably game beginning
157 * and we need to load the opening book right now. */
158 prepare_move(e
, b
, m
->color
);
162 /* Stop pondering. */
163 /* XXX: If we are about to receive multiple 'play' commands,
164 * e.g. in a rengo, we will not ponder during the rest of them. */
165 uct_pondering_stop(u
);
167 if (is_resign(m
->coord
)) {
173 /* Promote node of the appropriate move to the tree root. */
175 if (!tree_promote_at(u
->t
, b
, m
->coord
)) {
177 fprintf(stderr
, "Warning: Cannot promote move node! Several play commands in row?\n");
181 /* Setting up dynkomi is not necessary here, probably, but we
182 * better do it anyway for consistency reasons. */
183 setup_dynkomi(u
, b
, stone_other(m
->color
));
188 uct_chat(struct engine
*e
, struct board
*b
, char *cmd
)
190 struct uct
*u
= e
->data
;
191 static char reply
[1024];
193 cmd
+= strspn(cmd
, " \n\t");
194 if (!strncasecmp(cmd
, "winrate", 7)) {
196 return "no game context (yet?)";
197 enum stone color
= u
->t
->root_color
;
198 struct tree_node
*n
= u
->t
->root
;
199 snprintf(reply
, 1024, "In %d playouts at %d threads, %s %s can win with %.2f%% probability",
200 n
->u
.playouts
, u
->threads
, stone2str(color
), coord2sstr(n
->coord
, b
),
201 tree_node_get_value(u
->t
, -1, n
->u
.value
) * 100);
202 if (u
->t
->use_extra_komi
&& abs(u
->t
->extra_komi
) >= 0.5) {
203 sprintf(reply
+ strlen(reply
), ", while self-imposing extra komi %.1f",
213 uct_dead_group_list(struct engine
*e
, struct board
*b
, struct move_queue
*mq
)
215 struct uct
*u
= e
->data
;
217 /* This means the game is probably over, no use pondering on. */
218 uct_pondering_stop(u
);
220 if (u
->pass_all_alive
)
221 return; // no dead groups
223 bool mock_state
= false;
226 /* No state, but we cannot just back out - we might
227 * have passed earlier, only assuming some stones are
228 * dead, and then re-connected, only to lose counting
229 * when all stones are assumed alive. */
230 /* Mock up some state and seed the ownermap by few
232 prepare_move(e
, b
, S_BLACK
); assert(u
->t
);
233 for (int i
= 0; i
< GJ_MINGAMES
; i
++)
234 uct_playout(u
, b
, S_BLACK
, u
->t
);
238 dead_group_list(u
, b
, mq
);
241 /* Clean up the mock state in case we will receive
242 * a genmove; we could get a non-alternating-move
243 * error from prepare_move() in that case otherwise. */
249 playout_policy_done(struct playout_policy
*p
)
251 if (p
->done
) p
->done(p
);
252 if (p
->data
) free(p
->data
);
257 uct_done(struct engine
*e
)
259 /* This is called on engine reset, especially when clear_board
260 * is received and new game should begin. */
261 struct uct
*u
= e
->data
;
262 uct_pondering_stop(u
);
263 if (u
->t
) reset_state(u
);
264 free(u
->ownermap
.map
);
267 free(u
->random_policy
);
268 playout_policy_done(u
->playout
);
269 uct_prior_done(u
->prior
);
273 /* Pachi threading structure (if uct_playouts_parallel() is used):
276 * | main(), GTP communication, ...
277 * | starts and stops the search managed by thread_manager
280 * | spawns and collects worker threads
286 * uct_playouts() loop, doing descend-playout until uct_halt
288 * Another way to look at it is by functions (lines denote thread boundaries):
291 * | uct_search() (uct_search_start() .. uct_search_stop())
292 * | -----------------------
293 * | spawn_thread_manager()
294 * | -----------------------
296 * V uct_playouts() */
298 /* Set in thread manager in case the workers should stop. */
299 volatile sig_atomic_t uct_halt
= 0;
300 /* ID of the running worker thread. */
301 __thread
int thread_id
= -1;
302 /* ID of the thread manager. */
303 static pthread_t thread_manager
;
304 static bool thread_manager_running
;
306 static pthread_mutex_t finish_mutex
= PTHREAD_MUTEX_INITIALIZER
;
307 static pthread_cond_t finish_cond
= PTHREAD_COND_INITIALIZER
;
308 static volatile int finish_thread
;
309 static pthread_mutex_t finish_serializer
= PTHREAD_MUTEX_INITIALIZER
;
322 spawn_worker(void *ctx_
)
324 struct spawn_ctx
*ctx
= ctx_
;
326 fast_srandom(ctx
->seed
);
327 thread_id
= ctx
->tid
;
329 ctx
->games
= uct_playouts(ctx
->u
, ctx
->b
, ctx
->color
, ctx
->t
);
331 pthread_mutex_lock(&finish_serializer
);
332 pthread_mutex_lock(&finish_mutex
);
333 finish_thread
= ctx
->tid
;
334 pthread_cond_signal(&finish_cond
);
335 pthread_mutex_unlock(&finish_mutex
);
339 /* Thread manager, controlling worker threads. It must be called with
340 * finish_mutex lock held, but it will unlock it itself before exiting;
341 * this is necessary to be completely deadlock-free. */
342 /* The finish_cond can be signalled for it to stop; in that case,
343 * the caller should set finish_thread = -1. */
344 /* After it is started, it will update mctx->t to point at some tree
345 * used for the actual search (matters only for TM_ROOT), on return
346 * it will set mctx->games to the number of performed simulations. */
348 spawn_thread_manager(void *ctx_
)
350 /* In thread_manager, we use only some of the ctx fields. */
351 struct spawn_ctx
*mctx
= ctx_
;
352 struct uct
*u
= mctx
->u
;
353 struct tree
*t
= mctx
->t
;
354 bool shared_tree
= u
->parallel_tree
;
355 fast_srandom(mctx
->seed
);
357 int played_games
= 0;
358 pthread_t threads
[u
->threads
];
363 /* Spawn threads... */
364 for (int ti
= 0; ti
< u
->threads
; ti
++) {
365 struct spawn_ctx
*ctx
= malloc(sizeof(*ctx
));
366 ctx
->u
= u
; ctx
->b
= mctx
->b
; ctx
->color
= mctx
->color
;
367 mctx
->t
= ctx
->t
= shared_tree
? t
: tree_copy(t
);
368 ctx
->tid
= ti
; ctx
->seed
= fast_random(65536) + ti
;
369 pthread_create(&threads
[ti
], NULL
, spawn_worker
, ctx
);
371 fprintf(stderr
, "Spawned worker %d\n", ti
);
374 /* ...and collect them back: */
375 while (joined
< u
->threads
) {
376 /* Wait for some thread to finish... */
377 pthread_cond_wait(&finish_cond
, &finish_mutex
);
378 if (finish_thread
< 0) {
379 /* Stop-by-caller. Tell the workers to wrap up. */
383 /* ...and gather its remnants. */
384 struct spawn_ctx
*ctx
;
385 pthread_join(threads
[finish_thread
], (void **) &ctx
);
386 played_games
+= ctx
->games
;
389 if (ctx
->t
== mctx
->t
) mctx
->t
= t
;
390 tree_merge(t
, ctx
->t
);
395 fprintf(stderr
, "Joined worker %d\n", finish_thread
);
396 pthread_mutex_unlock(&finish_serializer
);
399 pthread_mutex_unlock(&finish_mutex
);
402 tree_normalize(mctx
->t
, u
->threads
);
404 mctx
->games
= played_games
;
408 static struct spawn_ctx
*
409 uct_search_start(struct uct
*u
, struct board
*b
, enum stone color
, struct tree
*t
)
411 assert(u
->threads
> 0);
412 assert(!thread_manager_running
);
414 struct spawn_ctx ctx
= { .u
= u
, .b
= b
, .color
= color
, .t
= t
, .seed
= fast_random(65536) };
415 static struct spawn_ctx mctx
; mctx
= ctx
;
416 pthread_mutex_lock(&finish_mutex
);
417 pthread_create(&thread_manager
, NULL
, spawn_thread_manager
, &mctx
);
418 thread_manager_running
= true;
422 static struct spawn_ctx
*
423 uct_search_stop(void)
425 assert(thread_manager_running
);
427 /* Signal thread manager to stop the workers. */
428 pthread_mutex_lock(&finish_mutex
);
430 pthread_cond_signal(&finish_cond
);
431 pthread_mutex_unlock(&finish_mutex
);
433 /* Collect the thread manager. */
434 struct spawn_ctx
*pctx
;
435 thread_manager_running
= false;
436 pthread_join(thread_manager
, (void **) &pctx
);
441 /* Run time-limited MCTS search on foreground. */
443 uct_search(struct uct
*u
, struct board
*b
, struct time_info
*ti
, enum stone color
, struct tree
*t
)
445 if (UDEBUGL(2) && u
->t
->root
->u
.playouts
> 0)
446 fprintf(stderr
, "<pre-simulated %d games skipped>\n", u
->t
->root
->u
.playouts
);
448 /* Set up time conditions. */
449 if (ti
->period
== TT_NULL
) *ti
= default_ti
;
450 struct time_stop stop
;
451 time_stop_conditions(ti
, b
, u
->fuseki_end
, u
->yose_start
, &stop
);
453 /* Number of last game with progress print. */
454 int last_print
= t
->root
->u
.playouts
;
455 /* Number of simulations to wait before next print. */
456 int print_interval
= TREE_SIMPROGRESS_INTERVAL
* (u
->thread_model
== TM_ROOT
? 1 : u
->threads
);
457 /* Printed notification about full memory? */
458 bool print_fullmem
= false;
460 struct spawn_ctx
*ctx
= uct_search_start(u
, b
, color
, t
);
462 /* The search tree is ctx->t. This is normally == t, but in case of
463 * TM_ROOT, it is one of the trees belonging to the independent
464 * workers. It is important to reference ctx->t directly since the
465 * thread manager will swap the tree pointer asynchronously. */
466 /* XXX: This means TM_ROOT support is suboptimal since single stalled
467 * thread can stall the others in case of limiting the search by game
468 * count. However, TM_ROOT just does not deserve any more extra code
471 struct tree_node
*best
= NULL
, *prev_best
;
472 struct tree_node
*winner
= NULL
, *prev_winner
;
474 double busywait_interval
= TREE_BUSYWAIT_INTERVAL
;
476 /* Now, just periodically poll the search tree. */
478 time_sleep(busywait_interval
);
479 /* busywait_interval should never be less than desired time, or the
480 * time control is broken. But if it happens to be less, we still search
481 * at least 100ms otherwise the move is completely random. */
483 int i
= ctx
->t
->root
->u
.playouts
;
485 /* Print progress? */
486 if (i
- last_print
> print_interval
) {
487 last_print
+= print_interval
; // keep the numbers tidy
488 uct_progress_status(u
, ctx
->t
, color
, last_print
);
490 if (!print_fullmem
&& ctx
->t
->nodes_size
> u
->max_tree_size
) {
492 fprintf(stderr
, "memory limit hit (%ld > %lu)\n", ctx
->t
->nodes_size
, u
->max_tree_size
);
493 print_fullmem
= true;
496 /* Check against time settings. */
497 bool desired_done
= false;
498 if (ti
->dim
== TD_WALLTIME
) {
499 double elapsed
= time_now() - ti
->len
.t
.timer_start
;
500 if (elapsed
> stop
.worst
.time
) break;
501 desired_done
= elapsed
> stop
.desired
.time
;
503 assert(ti
->dim
== TD_GAMES
);
504 if (i
> stop
.worst
.playouts
) break;
505 desired_done
= i
> stop
.desired
.playouts
;
508 /* Early break in won situation. */
510 best
= u
->policy
->choose(u
->policy
, ctx
->t
->root
, b
, color
);
511 if (best
&& ((best
->u
.playouts
>= 2000 && tree_node_get_value(ctx
->t
, 1, best
->u
.value
) >= u
->loss_threshold
)
512 || (best
->u
.playouts
>= 500 && tree_node_get_value(ctx
->t
, 1, best
->u
.value
) >= 0.95)))
516 if (!u
->policy
->winner
|| !u
->policy
->evaluate
)
518 /* Stop only if best explored has also highest value: */
519 prev_winner
= winner
;
520 winner
= u
->policy
->winner(u
->policy
, ctx
->t
, ctx
->t
->root
);
521 if (best
&& best
== winner
)
523 if (UDEBUGL(3) && (best
!= prev_best
|| winner
!= prev_winner
)) {
524 fprintf(stderr
, "[%d] best", i
);
526 fprintf(stderr
, " %3s [%d] %f", coord2sstr(best
->coord
, ctx
->t
->board
),
527 best
->u
.playouts
, tree_node_get_value(ctx
->t
, 1, best
->u
.value
));
528 fprintf(stderr
, " != winner");
530 fprintf(stderr
, " %3s [%d] %f ", coord2sstr(winner
->coord
, ctx
->t
->board
),
531 winner
->u
.playouts
, tree_node_get_value(ctx
->t
, 1, winner
->u
.value
));
532 fprintf(stderr
, "\n");
536 /* TODO: Early break if best->variance goes under threshold and we already
537 * have enough playouts (possibly thanks to book or to pondering). */
538 /* TODO: Early break if second best has no chance to catch up. */
541 ctx
= uct_search_stop();
544 tree_dump(t
, u
->dumpthres
);
546 uct_progress_status(u
, t
, color
, ctx
->games
);
552 /* Start pondering background with @color to play. */
554 uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
)
557 fprintf(stderr
, "Starting to ponder with color %s\n", stone2str(stone_other(color
)));
560 /* We need a local board copy to ponder upon. */
561 struct board
*b
= malloc(sizeof(*b
)); board_copy(b
, b0
);
563 /* *b0 did not have the genmove'd move played yet. */
564 struct move m
= { t
->root
->coord
, t
->root_color
};
565 int res
= board_play(b
, &m
);
567 setup_dynkomi(u
, b
, stone_other(m
.color
));
569 /* Start MCTS manager thread "headless". */
570 uct_search_start(u
, b
, color
, t
);
573 /* uct_search_stop() frontend for the pondering (non-genmove) mode. */
575 uct_pondering_stop(struct uct
*u
)
577 u
->pondering
= false;
578 if (!thread_manager_running
)
581 /* Stop the thread manager. */
582 struct spawn_ctx
*ctx
= uct_search_stop();
584 fprintf(stderr
, "(pondering) ");
585 uct_progress_status(u
, ctx
->t
, ctx
->color
, ctx
->games
);
592 uct_genmove(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
594 double start_time
= time_now();
595 struct uct
*u
= e
->data
;
597 if (b
->superko_violation
) {
598 fprintf(stderr
, "!!! WARNING: SUPERKO VIOLATION OCCURED BEFORE THIS MOVE\n");
599 fprintf(stderr
, "Maybe you play with situational instead of positional superko?\n");
600 fprintf(stderr
, "I'm going to ignore the violation, but note that I may miss\n");
601 fprintf(stderr
, "some moves valid under this ruleset because of this.\n");
602 b
->superko_violation
= false;
606 uct_pondering_stop(u
);
607 prepare_move(e
, b
, color
);
610 /* How to decide whether to use dynkomi in this game? Since we use
611 * pondering, it's not simple "who-to-play" matter. Decide based on
612 * the last genmove issued. */
613 u
->t
->use_extra_komi
= !!(u
->dynkomi_mask
& color
);
614 setup_dynkomi(u
, b
, color
);
616 /* Perform the Monte Carlo Tree Search! */
617 int played_games
= uct_search(u
, b
, ti
, color
, u
->t
);
619 /* Choose the best move from the tree. */
620 struct tree_node
*best
= u
->policy
->choose(u
->policy
, u
->t
->root
, b
, color
);
623 return coord_copy(pass
);
626 fprintf(stderr
, "*** WINNER is %s (%d,%d) with score %1.4f (%d/%d:%d games)\n",
627 coord2sstr(best
->coord
, b
), coord_x(best
->coord
, b
), coord_y(best
->coord
, b
),
628 tree_node_get_value(u
->t
, 1, best
->u
.value
),
629 best
->u
.playouts
, u
->t
->root
->u
.playouts
, played_games
);
631 /* Do not resign if we're so short of time that evaluation of best move is completely
632 * unreliable, we might be winning actually. In this case best is almost random but
633 * still better than resign. */
634 if (tree_node_get_value(u
->t
, 1, best
->u
.value
) < u
->resign_ratio
&& !is_pass(best
->coord
)
635 && best
->u
.playouts
> GJ_MINGAMES
) {
637 return coord_copy(resign
);
640 /* If the opponent just passed and we win counting, always
642 if (b
->moves
> 1 && is_pass(b
->last_move
.coord
)) {
643 /* Make sure enough playouts are simulated. */
644 while (u
->ownermap
.playouts
< GJ_MINGAMES
)
645 uct_playout(u
, b
, color
, u
->t
);
646 if (uct_pass_is_safe(u
, b
, color
, u
->pass_all_alive
|| pass_all_alive
)) {
648 fprintf(stderr
, "<Will rather pass, looks safe enough.>\n");
653 tree_promote_node(u
->t
, &best
);
654 /* After a pass, pondering is harmful for two reasons:
655 * (i) We might keep pondering even when the game is over.
656 * Of course this is the case for opponent resign as well.
657 * (ii) More importantly, the ownermap will get skewed since
658 * the UCT will start cutting off any playouts. */
659 if (u
->pondering_opt
&& !is_pass(best
->coord
)) {
660 uct_pondering_start(u
, b
, u
->t
, stone_other(color
));
663 double time
= time_now() - start_time
+ 0.000001; /* avoid divide by zero */
664 fprintf(stderr
, "genmove in %0.2fs (%d games/s, %d games/s/thread)\n",
665 time
, (int)(played_games
/time
), (int)(played_games
/time
/u
->threads
));
667 return coord_copy(best
->coord
);
672 uct_genbook(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
)
674 struct uct
*u
= e
->data
;
675 if (!u
->t
) prepare_move(e
, b
, color
);
678 if (ti
->dim
== TD_GAMES
) {
679 /* Don't count in games that already went into the book. */
680 ti
->len
.games
+= u
->t
->root
->u
.playouts
;
682 uct_search(u
, b
, ti
, color
, u
->t
);
684 assert(ti
->dim
== TD_GAMES
);
685 tree_save(u
->t
, b
, ti
->len
.games
/ 100);
691 uct_dumpbook(struct engine
*e
, struct board
*b
, enum stone color
)
693 struct uct
*u
= e
->data
;
694 struct tree
*t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0);
702 uct_state_init(char *arg
, struct board
*b
)
704 struct uct
*u
= calloc(1, sizeof(struct uct
));
707 u
->gamelen
= MC_GAMELEN
;
711 u
->playout_amaf
= true;
712 u
->playout_amaf_nakade
= false;
713 u
->amaf_prior
= false;
714 u
->max_tree_size
= 3072ULL * 1048576;
716 if (board_size(b
) - 2 >= 19)
718 u
->dynkomi_mask
= S_BLACK
;
721 u
->thread_model
= TM_TREEVL
;
722 u
->parallel_tree
= true;
723 u
->virtual_loss
= true;
724 u
->fuseki_end
= 20; // max time at 361*20% = 72 moves (our 36th move, still 99 to play)
725 u
->yose_start
= 40; // (100-40-25)*361/100/2 = 63 moves still to play by us then
727 u
->val_scale
= 0.04; u
->val_points
= 40;
730 char *optspec
, *next
= arg
;
733 next
+= strcspn(next
, ",");
734 if (*next
) { *next
++ = 0; } else { *next
= 0; }
736 char *optname
= optspec
;
737 char *optval
= strchr(optspec
, '=');
738 if (optval
) *optval
++ = 0;
740 if (!strcasecmp(optname
, "debug")) {
742 u
->debug_level
= atoi(optval
);
745 } else if (!strcasecmp(optname
, "mercy") && optval
) {
746 /* Minimal difference of black/white captures
747 * to stop playout - "Mercy Rule". Speeds up
748 * hopeless playouts at the expense of some
750 u
->mercymin
= atoi(optval
);
751 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
752 u
->gamelen
= atoi(optval
);
753 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
754 u
->expand_p
= atoi(optval
);
755 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
756 u
->dumpthres
= atoi(optval
);
757 } else if (!strcasecmp(optname
, "playout_amaf")) {
758 /* Whether to include random playout moves in
759 * AMAF as well. (Otherwise, only tree moves
760 * are included in AMAF. Of course makes sense
761 * only in connection with an AMAF policy.) */
762 /* with-without: 55.5% (+-4.1) */
763 if (optval
&& *optval
== '0')
764 u
->playout_amaf
= false;
766 u
->playout_amaf
= true;
767 } else if (!strcasecmp(optname
, "playout_amaf_nakade")) {
768 /* Whether to include nakade moves from playouts
769 * in the AMAF statistics; this tends to nullify
770 * the playout_amaf effect by adding too much
772 if (optval
&& *optval
== '0')
773 u
->playout_amaf_nakade
= false;
775 u
->playout_amaf_nakade
= true;
776 } else if (!strcasecmp(optname
, "playout_amaf_cutoff") && optval
) {
777 /* Keep only first N% of playout stage AMAF
779 u
->playout_amaf_cutoff
= atoi(optval
);
780 } else if ((!strcasecmp(optname
, "policy") || !strcasecmp(optname
, "random_policy")) && optval
) {
781 char *policyarg
= strchr(optval
, ':');
782 struct uct_policy
**p
= !strcasecmp(optname
, "policy") ? &u
->policy
: &u
->random_policy
;
785 if (!strcasecmp(optval
, "ucb1")) {
786 *p
= policy_ucb1_init(u
, policyarg
);
787 } else if (!strcasecmp(optval
, "ucb1amaf")) {
788 *p
= policy_ucb1amaf_init(u
, policyarg
);
790 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
793 } else if (!strcasecmp(optname
, "playout") && optval
) {
794 char *playoutarg
= strchr(optval
, ':');
797 if (!strcasecmp(optval
, "moggy")) {
798 u
->playout
= playout_moggy_init(playoutarg
);
799 } else if (!strcasecmp(optval
, "light")) {
800 u
->playout
= playout_light_init(playoutarg
);
801 } else if (!strcasecmp(optval
, "elo")) {
802 u
->playout
= playout_elo_init(playoutarg
);
804 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
807 } else if (!strcasecmp(optname
, "prior") && optval
) {
808 u
->prior
= uct_prior_init(optval
, b
);
809 } else if (!strcasecmp(optname
, "amaf_prior") && optval
) {
810 u
->amaf_prior
= atoi(optval
);
811 } else if (!strcasecmp(optname
, "threads") && optval
) {
812 /* By default, Pachi will run with only single
813 * tree search thread! */
814 u
->threads
= atoi(optval
);
815 } else if (!strcasecmp(optname
, "thread_model") && optval
) {
816 if (!strcasecmp(optval
, "root")) {
817 /* Root parallelization - each thread
818 * does independent search, trees are
819 * merged at the end. */
820 u
->thread_model
= TM_ROOT
;
821 u
->parallel_tree
= false;
822 u
->virtual_loss
= false;
823 } else if (!strcasecmp(optval
, "tree")) {
824 /* Tree parallelization - all threads
825 * grind on the same tree. */
826 u
->thread_model
= TM_TREE
;
827 u
->parallel_tree
= true;
828 u
->virtual_loss
= false;
829 } else if (!strcasecmp(optval
, "treevl")) {
830 /* Tree parallelization, but also
831 * with virtual losses - this discou-
832 * rages most threads choosing the
833 * same tree branches to read. */
834 u
->thread_model
= TM_TREEVL
;
835 u
->parallel_tree
= true;
836 u
->virtual_loss
= true;
838 fprintf(stderr
, "UCT: Invalid thread model %s\n", optval
);
841 } else if (!strcasecmp(optname
, "pondering")) {
842 /* Keep searching even during opponent's turn. */
843 u
->pondering_opt
= !optval
|| atoi(optval
);
844 } else if (!strcasecmp(optname
, "fuseki_end") && optval
) {
845 /* At the very beginning it's not worth thinking
846 * too long because the playout evaluations are
847 * very noisy. So gradually increase the thinking
848 * time up to maximum when fuseki_end percent
849 * of the board has been played.
850 * This only applies if we are not in byoyomi. */
851 u
->fuseki_end
= atoi(optval
);
852 } else if (!strcasecmp(optname
, "yose_start") && optval
) {
853 /* When yose_start percent of the board has been
854 * played, or if we are in byoyomi, stop spending
855 * more time and spread the remaining time
857 * Between fuseki_end and yose_start, we spend
858 * a constant proportion of the remaining time
859 * on each move. (yose_start should actually
860 * be much earlier than when real yose start,
861 * but "yose" is a good short name to convey
863 u
->yose_start
= atoi(optval
);
864 } else if (!strcasecmp(optname
, "force_seed") && optval
) {
865 u
->force_seed
= atoi(optval
);
866 } else if (!strcasecmp(optname
, "no_book")) {
868 } else if (!strcasecmp(optname
, "dynkomi")) {
869 /* Dynamic komi in handicap game; linearly
870 * decreases to basic settings until move
872 u
->dynkomi
= optval
? atoi(optval
) : 150;
873 } else if (!strcasecmp(optname
, "dynkomi_mask") && optval
) {
874 /* Bitmask of colors the player must be
875 * for dynkomi be applied; you may want
876 * to use dynkomi_mask=3 to allow dynkomi
877 * even in games where Pachi is white. */
878 u
->dynkomi_mask
= atoi(optval
);
879 } else if (!strcasecmp(optname
, "val_scale") && optval
) {
880 /* How much of the game result value should be
881 * influenced by win size. Zero means it isn't. */
882 u
->val_scale
= atof(optval
);
883 } else if (!strcasecmp(optname
, "val_points") && optval
) {
884 /* Maximum size of win to be scaled into game
885 * result value. Zero means boardsize^2. */
886 u
->val_points
= atoi(optval
) * 2; // result values are doubled
887 } else if (!strcasecmp(optname
, "val_extra")) {
888 /* If false, the score coefficient will be simply
889 * added to the value, instead of scaling the result
890 * coefficient because of it. */
891 u
->val_extra
= !optval
|| atoi(optval
);
892 } else if (!strcasecmp(optname
, "root_heuristic") && optval
) {
893 /* Whether to bias exploration by root node values
894 * (must be supported by the used policy).
896 * 1: Do, value = result.
897 * Try to temper the result:
898 * 2: Do, value = 0.5+(result-expected)/2.
899 * 3: Do, value = 0.5+bzz((result-expected)^2). */
900 u
->root_heuristic
= atoi(optval
);
901 } else if (!strcasecmp(optname
, "pass_all_alive")) {
902 /* Whether to consider all stones alive at the game
903 * end instead of marking dead groupd. */
904 u
->pass_all_alive
= !optval
|| atoi(optval
);
905 } else if (!strcasecmp(optname
, "random_policy_chance") && optval
) {
906 /* If specified (N), with probability 1/N, random_policy policy
907 * descend is used instead of main policy descend; useful
908 * if specified policy (e.g. UCB1AMAF) can make unduly biased
909 * choices sometimes, you can fall back to e.g.
910 * random_policy=UCB1. */
911 u
->random_policy_chance
= atoi(optval
);
912 } else if (!strcasecmp(optname
, "max_tree_size") && optval
) {
913 /* Maximum amount of memory [MiB] consumed by the move tree.
914 * Default is 3072 (3 GiB). Note that if you use TM_ROOT,
915 * this limits size of only one of the trees, not all of them
917 u
->max_tree_size
= atol(optval
) * 1048576;
918 } else if (!strcasecmp(optname
, "banner") && optval
) {
919 /* Additional banner string. This must come as the
920 * last engine parameter. */
921 if (*next
) *--next
= ',';
922 u
->banner
= strdup(optval
);
925 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
931 u
->resign_ratio
= 0.2; /* Resign when most games are lost. */
932 u
->loss_threshold
= 0.85; /* Stop reading if after at least 5000 playouts this is best value. */
934 u
->policy
= policy_ucb1amaf_init(u
, NULL
);
936 if (!!u
->random_policy_chance
^ !!u
->random_policy
) {
937 fprintf(stderr
, "uct: Only one of random_policy and random_policy_chance is set\n");
941 if (u
->fast_alloc
&& !u
->parallel_tree
) {
942 fprintf(stderr
, "fast_alloc not supported with root parallelization.\n");
947 u
->prior
= uct_prior_init(NULL
, b
);
950 u
->playout
= playout_moggy_init(NULL
);
951 u
->playout
->debug_level
= u
->debug_level
;
953 u
->ownermap
.map
= malloc(board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
955 /* Some things remain uninitialized for now - the opening book
956 * is not loaded and the tree not set up. */
957 /* This will be initialized in setup_state() at the first move
958 * received/requested. This is because right now we are not aware
959 * about any komi or handicap setup and such. */
965 engine_uct_init(char *arg
, struct board
*b
)
967 struct uct
*u
= uct_state_init(arg
, b
);
968 struct engine
*e
= calloc(1, sizeof(struct engine
));
969 e
->name
= "UCT Engine";
970 e
->printhook
= uct_printhook_ownermap
;
971 e
->notify_play
= uct_notify_play
;
973 e
->genmove
= uct_genmove
;
974 e
->dead_group_list
= uct_dead_group_list
;
978 const char banner
[] = "I'm playing UCT. When I'm losing, I will resign, "
979 "if I think I win, I play until you pass. "
980 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
981 if (!u
->banner
) u
->banner
= "";
982 e
->comment
= malloc(sizeof(banner
) + strlen(u
->banner
) + 1);
983 sprintf(e
->comment
, "%s %s", banner
, u
->banner
);