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 worst_time
= desired_time
* MAX_MAIN_TIME_EXTENSION
;
468 if (worst_time
> ti
->len
.t
.max_time
)
469 worst_time
= ti
->len
.t
.max_time
;
470 if (desired_time
> worst_time
)
471 desired_time
= worst_time
;
473 stop
->t
.desired_stop
= ti
->len
.t
.timer_start
+ desired_time
- ti
->len
.t
.net_lag
;
474 stop
->t
.worst_stop
= ti
->len
.t
.timer_start
+ worst_time
- ti
->len
.t
.net_lag
;
475 // Both stop points may be in the past if too much lag.
478 fprintf(stderr
, "desired time %.02f, worst %.02f\n", desired_time
, worst_time
);
482 /* Run time-limited MCTS search on foreground. */
484 uct_search(struct uct
*u
, struct board
*b
, struct time_info
*ti
, enum stone color
, struct tree
*t
)
486 union stop_conditions stop
;
487 time_prep(ti
, u
, b
, &stop
);
488 if (UDEBUGL(2) && u
->t
->root
->u
.playouts
> 0)
489 fprintf(stderr
, "<pre-simulated %d games skipped>\n", u
->t
->root
->u
.playouts
);
491 /* Number of last game with progress print. */
492 int last_print
= t
->root
->u
.playouts
;
493 /* Number of simulations to wait before next print. */
494 int print_interval
= TREE_SIMPROGRESS_INTERVAL
* (u
->thread_model
== TM_ROOT
? 1 : u
->threads
);
495 /* Printed notification about full memory? */
496 bool print_fullmem
= false;
498 struct spawn_ctx
*ctx
= uct_search_start(u
, b
, color
, t
);
500 /* The search tree is ctx->t. This is normally == t, but in case of
501 * TM_ROOT, it is one of the trees belonging to the independent
502 * workers. It is important to reference ctx->t directly since the
503 * thread manager will swap the tree pointer asynchronously. */
504 /* XXX: This means TM_ROOT support is suboptimal since single stalled
505 * thread can stall the others in case of limiting the search by game
506 * count. However, TM_ROOT just does not deserve any more extra code
509 struct tree_node
*best
= NULL
, *prev_best
;
510 struct tree_node
*winner
= NULL
, *prev_winner
;
512 double busywait_interval
= TREE_BUSYWAIT_INTERVAL
;
514 /* Now, just periodically poll the search tree. */
516 time_sleep(busywait_interval
);
517 /* busywait_interval should never be less than desired time, or the
518 * time control is broken. But if it happens to be less, we still search
519 * at least 100ms otherwise the move is completely random. */
521 int i
= ctx
->t
->root
->u
.playouts
;
523 /* Print progress? */
524 if (i
- last_print
> print_interval
) {
525 last_print
+= print_interval
; // keep the numbers tidy
526 uct_progress_status(u
, ctx
->t
, color
, last_print
);
528 if (!print_fullmem
&& ctx
->t
->nodes_size
> u
->max_tree_size
) {
530 fprintf(stderr
, "memory limit hit (%ld > %lu)\n", ctx
->t
->nodes_size
, u
->max_tree_size
);
531 print_fullmem
= true;
534 /* Check against time settings. */
535 bool desired_done
= false;
536 if (ti
->dim
== TD_WALLTIME
) {
537 double now
= time_now();
538 if (now
> stop
.t
.worst_stop
) break;
539 desired_done
= now
> stop
.t
.desired_stop
;
541 assert(ti
->dim
== TD_GAMES
);
542 if (i
> stop
.p
.worst_playouts
) break;
543 desired_done
= i
> stop
.p
.desired_playouts
;
546 /* Early break in won situation. */
548 best
= u
->policy
->choose(u
->policy
, ctx
->t
->root
, b
, color
);
549 if (best
&& ((best
->u
.playouts
>= 2000 && tree_node_get_value(ctx
->t
, 1, best
->u
.value
) >= u
->loss_threshold
)
550 || (best
->u
.playouts
>= 500 && tree_node_get_value(ctx
->t
, 1, best
->u
.value
) >= 0.95)))
554 if (!u
->policy
->winner
|| !u
->policy
->evaluate
)
556 /* Stop only if best explored has also highest value: */
557 prev_winner
= winner
;
558 winner
= u
->policy
->winner(u
->policy
, ctx
->t
, ctx
->t
->root
);
559 if (best
&& best
== winner
)
561 if (UDEBUGL(3) && (best
!= prev_best
|| winner
!= prev_winner
)) {
562 fprintf(stderr
, "[%d] best", i
);
564 fprintf(stderr
, " %3s [%d] %f", coord2sstr(best
->coord
, ctx
->t
->board
),
565 best
->u
.playouts
, tree_node_get_value(ctx
->t
, 1, best
->u
.value
));
566 fprintf(stderr
, " != winner");
568 fprintf(stderr
, " %3s [%d] %f ", coord2sstr(winner
->coord
, ctx
->t
->board
),
569 winner
->u
.playouts
, tree_node_get_value(ctx
->t
, 1, winner
->u
.value
));
570 fprintf(stderr
, "\n");
574 /* TODO: Early break if best->variance goes under threshold and we already
575 * have enough playouts (possibly thanks to book or to pondering). */
576 /* TODO: Early break if second best has no chance to catch up. */
579 ctx
= uct_search_stop();
582 tree_dump(t
, u
->dumpthres
);
584 uct_progress_status(u
, t
, color
, ctx
->games
);
590 /* Start pondering background with @color to play. */
592 uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
)
595 fprintf(stderr
, "Starting to ponder with color %s\n", stone2str(stone_other(color
)));
597 /* We need a local board copy to ponder upon. */
598 struct board
*b
= malloc(sizeof(*b
)); board_copy(b
, b0
);
600 /* *b0 did not have the genmove'd move played yet. */
601 struct move m
= { t
->root
->coord
, t
->root_color
};
602 int res
= board_play(b
, &m
);
605 /* Start MCTS manager thread "headless". */
606 uct_search_start(u
, b
, color
, t
);
609 /* uct_search_stop() frontend for the pondering (non-genmove) mode. */
611 uct_pondering_stop(struct uct
*u
)
613 if (!thread_manager_running
)
616 /* Stop the thread manager. */
617 struct spawn_ctx
*ctx
= uct_search_stop();
619 fprintf(stderr
, "(pondering) ");
620 uct_progress_status(u
, ctx
->t
, ctx
->color
, ctx
->games
);
627 uct_genmove(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
629 struct uct
*u
= e
->data
;
631 if (b
->superko_violation
) {
632 fprintf(stderr
, "!!! WARNING: SUPERKO VIOLATION OCCURED BEFORE THIS MOVE\n");
633 fprintf(stderr
, "Maybe you play with situational instead of positional superko?\n");
634 fprintf(stderr
, "I'm going to ignore the violation, but note that I may miss\n");
635 fprintf(stderr
, "some moves valid under this ruleset because of this.\n");
636 b
->superko_violation
= false;
640 uct_pondering_stop(u
);
641 prepare_move(e
, b
, color
);
644 /* Perform the Monte Carlo Tree Search! */
645 int played_games
= uct_search(u
, b
, ti
, color
, u
->t
);
647 /* Choose the best move from the tree. */
648 struct tree_node
*best
= u
->policy
->choose(u
->policy
, u
->t
->root
, b
, color
);
651 return coord_copy(pass
);
654 fprintf(stderr
, "*** WINNER is %s (%d,%d) with score %1.4f (%d/%d:%d games)\n",
655 coord2sstr(best
->coord
, b
), coord_x(best
->coord
, b
), coord_y(best
->coord
, b
),
656 tree_node_get_value(u
->t
, 1, best
->u
.value
),
657 best
->u
.playouts
, u
->t
->root
->u
.playouts
, played_games
);
658 if (tree_node_get_value(u
->t
, 1, best
->u
.value
) < u
->resign_ratio
&& !is_pass(best
->coord
)) {
660 return coord_copy(resign
);
663 /* If the opponent just passed and we win counting, always
665 if (b
->moves
> 1 && is_pass(b
->last_move
.coord
)) {
666 /* Make sure enough playouts are simulated. */
667 while (u
->ownermap
.playouts
< GJ_MINGAMES
)
668 uct_playout(u
, b
, color
, u
->t
);
669 if (uct_pass_is_safe(u
, b
, color
, u
->pass_all_alive
|| pass_all_alive
)) {
671 fprintf(stderr
, "<Will rather pass, looks safe enough.>\n");
676 tree_promote_node(u
->t
, best
);
677 /* After a pass, pondering is harmful for two reasons:
678 * (i) We might keep pondering even when the game is over.
679 * Of course this is the case for opponent resign as well.
680 * (ii) More importantly, the ownermap will get skewed since
681 * the UCT will start cutting off any playouts. */
682 if (u
->pondering
&& !is_pass(best
->coord
)) {
683 uct_pondering_start(u
, b
, u
->t
, stone_other(color
));
685 return coord_copy(best
->coord
);
690 uct_genbook(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
)
692 struct uct
*u
= e
->data
;
693 if (!u
->t
) prepare_move(e
, b
, color
);
696 if (ti
->dim
== TD_GAMES
) {
697 /* Don't count in games that already went into the book. */
698 ti
->len
.games
+= u
->t
->root
->u
.playouts
;
700 uct_search(u
, b
, ti
, color
, u
->t
);
702 assert(ti
->dim
== TD_GAMES
);
703 tree_save(u
->t
, b
, ti
->len
.games
/ 100);
709 uct_dumpbook(struct engine
*e
, struct board
*b
, enum stone color
)
711 struct tree
*t
= tree_init(b
, color
);
719 uct_state_init(char *arg
, struct board
*b
)
721 struct uct
*u
= calloc(1, sizeof(struct uct
));
724 u
->gamelen
= MC_GAMELEN
;
728 u
->playout_amaf
= true;
729 u
->playout_amaf_nakade
= false;
730 u
->amaf_prior
= false;
731 u
->max_tree_size
= 3072ULL * 1048576;
733 if (board_size(b
) - 2 >= 19)
735 u
->dynkomi_mask
= S_BLACK
;
738 u
->thread_model
= TM_TREEVL
;
739 u
->parallel_tree
= true;
740 u
->virtual_loss
= true;
742 u
->val_scale
= 0.02; u
->val_points
= 20;
745 char *optspec
, *next
= arg
;
748 next
+= strcspn(next
, ",");
749 if (*next
) { *next
++ = 0; } else { *next
= 0; }
751 char *optname
= optspec
;
752 char *optval
= strchr(optspec
, '=');
753 if (optval
) *optval
++ = 0;
755 if (!strcasecmp(optname
, "debug")) {
757 u
->debug_level
= atoi(optval
);
760 } else if (!strcasecmp(optname
, "mercy") && optval
) {
761 /* Minimal difference of black/white captures
762 * to stop playout - "Mercy Rule". Speeds up
763 * hopeless playouts at the expense of some
765 u
->mercymin
= atoi(optval
);
766 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
767 u
->gamelen
= atoi(optval
);
768 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
769 u
->expand_p
= atoi(optval
);
770 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
771 u
->dumpthres
= atoi(optval
);
772 } else if (!strcasecmp(optname
, "playout_amaf")) {
773 /* Whether to include random playout moves in
774 * AMAF as well. (Otherwise, only tree moves
775 * are included in AMAF. Of course makes sense
776 * only in connection with an AMAF policy.) */
777 /* with-without: 55.5% (+-4.1) */
778 if (optval
&& *optval
== '0')
779 u
->playout_amaf
= false;
781 u
->playout_amaf
= true;
782 } else if (!strcasecmp(optname
, "playout_amaf_nakade")) {
783 /* Whether to include nakade moves from playouts
784 * in the AMAF statistics; this tends to nullify
785 * the playout_amaf effect by adding too much
787 if (optval
&& *optval
== '0')
788 u
->playout_amaf_nakade
= false;
790 u
->playout_amaf_nakade
= true;
791 } else if (!strcasecmp(optname
, "playout_amaf_cutoff") && optval
) {
792 /* Keep only first N% of playout stage AMAF
794 u
->playout_amaf_cutoff
= atoi(optval
);
795 } else if ((!strcasecmp(optname
, "policy") || !strcasecmp(optname
, "random_policy")) && optval
) {
796 char *policyarg
= strchr(optval
, ':');
797 struct uct_policy
**p
= !strcasecmp(optname
, "policy") ? &u
->policy
: &u
->random_policy
;
800 if (!strcasecmp(optval
, "ucb1")) {
801 *p
= policy_ucb1_init(u
, policyarg
);
802 } else if (!strcasecmp(optval
, "ucb1amaf")) {
803 *p
= policy_ucb1amaf_init(u
, policyarg
);
805 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
808 } else if (!strcasecmp(optname
, "playout") && optval
) {
809 char *playoutarg
= strchr(optval
, ':');
812 if (!strcasecmp(optval
, "moggy")) {
813 u
->playout
= playout_moggy_init(playoutarg
);
814 } else if (!strcasecmp(optval
, "light")) {
815 u
->playout
= playout_light_init(playoutarg
);
816 } else if (!strcasecmp(optval
, "elo")) {
817 u
->playout
= playout_elo_init(playoutarg
);
819 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
822 } else if (!strcasecmp(optname
, "prior") && optval
) {
823 u
->prior
= uct_prior_init(optval
, b
);
824 } else if (!strcasecmp(optname
, "amaf_prior") && optval
) {
825 u
->amaf_prior
= atoi(optval
);
826 } else if (!strcasecmp(optname
, "threads") && optval
) {
827 /* By default, Pachi will run with only single
828 * tree search thread! */
829 u
->threads
= atoi(optval
);
830 } else if (!strcasecmp(optname
, "thread_model") && optval
) {
831 if (!strcasecmp(optval
, "root")) {
832 /* Root parallelization - each thread
833 * does independent search, trees are
834 * merged at the end. */
835 u
->thread_model
= TM_ROOT
;
836 u
->parallel_tree
= false;
837 u
->virtual_loss
= false;
838 } else if (!strcasecmp(optval
, "tree")) {
839 /* Tree parallelization - all threads
840 * grind on the same tree. */
841 u
->thread_model
= TM_TREE
;
842 u
->parallel_tree
= true;
843 u
->virtual_loss
= false;
844 } else if (!strcasecmp(optval
, "treevl")) {
845 /* Tree parallelization, but also
846 * with virtual losses - this discou-
847 * rages most threads choosing the
848 * same tree branches to read. */
849 u
->thread_model
= TM_TREEVL
;
850 u
->parallel_tree
= true;
851 u
->virtual_loss
= true;
853 fprintf(stderr
, "UCT: Invalid thread model %s\n", optval
);
856 } else if (!strcasecmp(optname
, "pondering")) {
857 /* Keep searching even during opponent's turn. */
858 u
->pondering
= !optval
|| atoi(optval
);
859 } else if (!strcasecmp(optname
, "force_seed") && optval
) {
860 u
->force_seed
= atoi(optval
);
861 } else if (!strcasecmp(optname
, "no_book")) {
863 } else if (!strcasecmp(optname
, "dynkomi")) {
864 /* Dynamic komi in handicap game; linearly
865 * decreases to basic settings until move
867 u
->dynkomi
= optval
? atoi(optval
) : 150;
868 } else if (!strcasecmp(optname
, "dynkomi_mask") && optval
) {
869 /* Bitmask of colors the player must be
870 * for dynkomi be applied; you may want
871 * to use dynkomi_mask=3 to allow dynkomi
872 * even in games where Pachi is white. */
873 u
->dynkomi_mask
= atoi(optval
);
874 } else if (!strcasecmp(optname
, "val_scale") && optval
) {
875 /* How much of the game result value should be
876 * influenced by win size. Zero means it isn't. */
877 u
->val_scale
= atof(optval
);
878 } else if (!strcasecmp(optname
, "val_points") && optval
) {
879 /* Maximum size of win to be scaled into game
880 * result value. Zero means boardsize^2. */
881 u
->val_points
= atoi(optval
) * 2; // result values are doubled
882 } else if (!strcasecmp(optname
, "val_extra")) {
883 /* If false, the score coefficient will be simply
884 * added to the value, instead of scaling the result
885 * coefficient because of it. */
886 u
->val_extra
= !optval
|| atoi(optval
);
887 } else if (!strcasecmp(optname
, "root_heuristic") && optval
) {
888 /* Whether to bias exploration by root node values
889 * (must be supported by the used policy).
891 * 1: Do, value = result.
892 * Try to temper the result:
893 * 2: Do, value = 0.5+(result-expected)/2.
894 * 3: Do, value = 0.5+bzz((result-expected)^2). */
895 u
->root_heuristic
= atoi(optval
);
896 } else if (!strcasecmp(optname
, "pass_all_alive")) {
897 /* Whether to consider all stones alive at the game
898 * end instead of marking dead groupd. */
899 u
->pass_all_alive
= !optval
|| atoi(optval
);
900 } else if (!strcasecmp(optname
, "random_policy_chance") && optval
) {
901 /* If specified (N), with probability 1/N, random_policy policy
902 * descend is used instead of main policy descend; useful
903 * if specified policy (e.g. UCB1AMAF) can make unduly biased
904 * choices sometimes, you can fall back to e.g.
905 * random_policy=UCB1. */
906 u
->random_policy_chance
= atoi(optval
);
907 } else if (!strcasecmp(optname
, "max_tree_size") && optval
) {
908 /* Maximum amount of memory [MiB] consumed by the move tree.
909 * Default is 3072 (3 GiB). Note that if you use TM_ROOT,
910 * this limits size of only one of the trees, not all of them
912 u
->max_tree_size
= atol(optval
) * 1048576;
913 } else if (!strcasecmp(optname
, "banner") && optval
) {
914 /* Additional banner string. This must come as the
915 * last engine parameter. */
916 if (*next
) *--next
= ',';
917 u
->banner
= strdup(optval
);
920 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
926 u
->resign_ratio
= 0.2; /* Resign when most games are lost. */
927 u
->loss_threshold
= 0.85; /* Stop reading if after at least 5000 playouts this is best value. */
929 u
->policy
= policy_ucb1amaf_init(u
, NULL
);
931 if (!!u
->random_policy_chance
^ !!u
->random_policy
) {
932 fprintf(stderr
, "uct: Only one of random_policy and random_policy_chance is set\n");
937 u
->prior
= uct_prior_init(NULL
, b
);
940 u
->playout
= playout_moggy_init(NULL
);
941 u
->playout
->debug_level
= u
->debug_level
;
943 u
->ownermap
.map
= malloc(board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
945 /* Some things remain uninitialized for now - the opening book
946 * is not loaded and the tree not set up. */
947 /* This will be initialized in setup_state() at the first move
948 * received/requested. This is because right now we are not aware
949 * about any komi or handicap setup and such. */
955 engine_uct_init(char *arg
, struct board
*b
)
957 struct uct
*u
= uct_state_init(arg
, b
);
958 struct engine
*e
= calloc(1, sizeof(struct engine
));
959 e
->name
= "UCT Engine";
960 e
->printhook
= uct_printhook_ownermap
;
961 e
->notify_play
= uct_notify_play
;
963 e
->genmove
= uct_genmove
;
964 e
->dead_group_list
= uct_dead_group_list
;
968 const char banner
[] = "I'm playing UCT. When I'm losing, I will resign, "
969 "if I think I win, I play until you pass. "
970 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
971 if (!u
->banner
) u
->banner
= "";
972 e
->comment
= malloc(sizeof(banner
) + strlen(u
->banner
) + 1);
973 sprintf(e
->comment
, "%s %s", banner
, u
->banner
);