17 #include "joseki/base.h"
19 #include "playout/moggy.h"
20 #include "playout/light.h"
21 #include "tactics/util.h"
23 #include "uct/dynkomi.h"
24 #include "uct/internal.h"
25 #include "uct/plugins.h"
26 #include "uct/prior.h"
27 #include "uct/search.h"
28 #include "uct/slave.h"
33 struct uct_policy
*policy_ucb1_init(struct uct
*u
, char *arg
);
34 struct uct_policy
*policy_ucb1amaf_init(struct uct
*u
, char *arg
, struct board
*board
);
35 static void uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
);
37 /* Maximal simulation length. */
38 #define MC_GAMELEN MAX_GAMELEN
42 setup_state(struct uct
*u
, struct board
*b
, enum stone color
)
44 u
->t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0,
45 u
->max_pruned_size
, u
->pruning_threshold
, u
->local_tree_aging
, u
->stats_hbits
);
46 if (u
->initial_extra_komi
)
47 u
->t
->extra_komi
= u
->initial_extra_komi
;
49 fast_srandom(u
->force_seed
);
51 fprintf(stderr
, "Fresh board with random seed %lu\n", fast_getseed());
52 if (!u
->no_tbook
&& b
->moves
== 0) {
53 if (color
== S_BLACK
) {
55 } else if (DEBUGL(0)) {
56 fprintf(stderr
, "Warning: First move appears to be white\n");
62 reset_state(struct uct
*u
)
65 tree_done(u
->t
); u
->t
= NULL
;
69 setup_dynkomi(struct uct
*u
, struct board
*b
, enum stone to_play
)
71 if (u
->t
->use_extra_komi
&& !u
->pondering
&& u
->dynkomi
->permove
)
72 u
->t
->extra_komi
= u
->dynkomi
->permove(u
->dynkomi
, b
, u
->t
);
73 else if (!u
->t
->use_extra_komi
)
78 uct_prepare_move(struct uct
*u
, struct board
*b
, enum stone color
)
81 /* Verify that we have sane state. */
83 assert(u
->t
&& b
->moves
);
84 if (color
!= stone_other(u
->t
->root_color
)) {
85 fprintf(stderr
, "Fatal: Non-alternating play detected %d %d\n",
86 color
, u
->t
->root_color
);
89 uct_htable_reset(u
->t
);
92 /* We need fresh state. */
94 setup_state(u
, b
, color
);
97 u
->ownermap
.playouts
= 0;
98 memset(u
->ownermap
.map
, 0, board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
99 u
->played_own
= u
->played_all
= 0;
103 dead_group_list(struct uct
*u
, struct board
*b
, struct move_queue
*mq
)
105 enum gj_state gs_array
[board_size2(b
)];
106 struct group_judgement gj
= { .thres
= GJ_THRES
, .gs
= gs_array
};
107 board_ownermap_judge_groups(b
, &u
->ownermap
, &gj
);
108 groups_of_status(b
, &gj
, GS_DEAD
, mq
);
112 uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
)
114 /* Make sure enough playouts are simulated to get a reasonable dead group list. */
115 while (u
->ownermap
.playouts
< GJ_MINGAMES
)
116 uct_playout(u
, b
, color
, u
->t
);
118 struct move_queue mq
= { .moves
= 0 };
119 dead_group_list(u
, b
, &mq
);
120 if (pass_all_alive
) {
121 for (unsigned int i
= 0; i
< mq
.moves
; i
++) {
122 if (board_at(b
, mq
.move
[i
]) == stone_other(color
)) {
123 return false; // We need to remove opponent dead groups first.
126 mq
.moves
= 0; // our dead stones are alive when pass_all_alive is true
128 if (u
->allow_losing_pass
) {
130 if (board_at(b
, c
) == S_OFFBOARD
)
132 if (board_ownermap_judge_point(&u
->ownermap
, c
, GJ_THRES
) == PJ_UNKNOWN
) {
134 fprintf(stderr
, "uct_pass_is_safe fails at %s[%d]\n", coord2sstr(c
, b
), c
);
135 return false; // Unclear point, clarify first.
140 return pass_is_safe(b
, color
, &mq
);
144 uct_printhook_ownermap(struct board
*board
, coord_t c
, char *s
, char *end
)
146 struct uct
*u
= board
->es
;
151 const char chr
[] = ":XO,"; // dame, black, white, unclear
152 const char chm
[] = ":xo,";
153 char ch
= chr
[board_ownermap_judge_point(&u
->ownermap
, c
, GJ_THRES
)];
154 if (ch
== ',') { // less precise estimate then?
155 ch
= chm
[board_ownermap_judge_point(&u
->ownermap
, c
, 0.67)];
157 s
+= snprintf(s
, end
- s
, "%c ", ch
);
162 uct_notify_play(struct engine
*e
, struct board
*b
, struct move
*m
, char *enginearg
)
164 struct uct
*u
= e
->data
;
166 /* No state, create one - this is probably game beginning
167 * and we need to load the opening tbook right now. */
168 uct_prepare_move(u
, b
, m
->color
);
172 /* Stop pondering, required by tree_promote_at() */
173 uct_pondering_stop(u
);
174 if (UDEBUGL(2) && u
->slave
)
175 tree_dump(u
->t
, u
->dumpthres
);
177 if (is_resign(m
->coord
)) {
183 /* Promote node of the appropriate move to the tree root. */
185 if (!tree_promote_at(u
->t
, b
, m
->coord
)) {
187 fprintf(stderr
, "Warning: Cannot promote move node! Several play commands in row?\n");
188 /* Preserve dynamic komi information, though, that is important. */
189 u
->initial_extra_komi
= u
->t
->extra_komi
;
194 /* If we are a slave in a distributed engine, start pondering once
195 * we know which move we actually played. See uct_genmove() about
196 * the check for pass. */
197 if (u
->pondering_opt
&& u
->slave
&& m
->color
== u
->my_color
&& !is_pass(m
->coord
))
198 uct_pondering_start(u
, b
, u
->t
, stone_other(m
->color
));
204 uct_undo(struct engine
*e
, struct board
*b
)
206 struct uct
*u
= e
->data
;
208 if (!u
->t
) return NULL
;
209 uct_pondering_stop(u
);
210 u
->initial_extra_komi
= u
->t
->extra_komi
;
216 uct_result(struct engine
*e
, struct board
*b
)
218 struct uct
*u
= e
->data
;
219 static char reply
[1024];
223 enum stone color
= u
->t
->root_color
;
224 struct tree_node
*n
= u
->t
->root
;
225 snprintf(reply
, 1024, "%s %s %d %.2f %.1f",
226 stone2str(color
), coord2sstr(node_coord(n
), b
),
227 n
->u
.playouts
, tree_node_get_value(u
->t
, -1, n
->u
.value
),
228 u
->t
->use_extra_komi
? u
->t
->extra_komi
: 0);
233 uct_chat(struct engine
*e
, struct board
*b
, bool opponent
, char *from
, char *cmd
)
235 struct uct
*u
= e
->data
;
238 return generic_chat(b
, opponent
, from
, cmd
, S_NONE
, pass
, 0, 1, u
->threads
, 0.0, 0.0);
240 struct tree_node
*n
= u
->t
->root
;
241 double winrate
= tree_node_get_value(u
->t
, -1, n
->u
.value
);
242 double extra_komi
= u
->t
->use_extra_komi
&& abs(u
->t
->extra_komi
) >= 0.5 ? u
->t
->extra_komi
: 0;
244 return generic_chat(b
, opponent
, from
, cmd
, u
->t
->root_color
, node_coord(n
), n
->u
.playouts
, 1,
245 u
->threads
, winrate
, extra_komi
);
249 uct_dead_group_list(struct engine
*e
, struct board
*b
, struct move_queue
*mq
)
251 struct uct
*u
= e
->data
;
253 /* This means the game is probably over, no use pondering on. */
254 uct_pondering_stop(u
);
256 if (u
->pass_all_alive
)
257 return; // no dead groups
259 bool mock_state
= false;
262 /* No state, but we cannot just back out - we might
263 * have passed earlier, only assuming some stones are
264 * dead, and then re-connected, only to lose counting
265 * when all stones are assumed alive. */
266 uct_prepare_move(u
, b
, S_BLACK
); assert(u
->t
);
269 /* Make sure the ownermap is well-seeded. */
270 while (u
->ownermap
.playouts
< GJ_MINGAMES
)
271 uct_playout(u
, b
, S_BLACK
, u
->t
);
272 /* Show the ownermap: */
274 board_print_custom(b
, stderr
, uct_printhook_ownermap
);
276 dead_group_list(u
, b
, mq
);
279 /* Clean up the mock state in case we will receive
280 * a genmove; we could get a non-alternating-move
281 * error from uct_prepare_move() in that case otherwise. */
287 playout_policy_done(struct playout_policy
*p
)
289 if (p
->done
) p
->done(p
);
290 if (p
->data
) free(p
->data
);
295 uct_done(struct engine
*e
)
297 /* This is called on engine reset, especially when clear_board
298 * is received and new game should begin. */
299 struct uct
*u
= e
->data
;
300 uct_pondering_stop(u
);
301 if (u
->t
) reset_state(u
);
302 free(u
->ownermap
.map
);
305 free(u
->random_policy
);
306 playout_policy_done(u
->playout
);
307 uct_prior_done(u
->prior
);
308 joseki_done(u
->jdict
);
309 pluginset_done(u
->plugins
);
314 /* Run time-limited MCTS search on foreground. */
316 uct_search(struct uct
*u
, struct board
*b
, struct time_info
*ti
, enum stone color
, struct tree
*t
, bool print_progress
)
318 struct uct_search_state s
;
319 uct_search_start(u
, b
, color
, t
, ti
, &s
);
320 if (UDEBUGL(2) && s
.base_playouts
> 0)
321 fprintf(stderr
, "<pre-simulated %d games>\n", s
.base_playouts
);
323 /* The search tree is ctx->t. This is currently == . It is important
324 * to reference ctx->t directly since the
325 * thread manager will swap the tree pointer asynchronously. */
327 /* Now, just periodically poll the search tree. */
328 /* Note that in case of TD_GAMES, threads will not wait for
329 * the uct_search_check_stop() signalization. */
331 time_sleep(TREE_BUSYWAIT_INTERVAL
);
332 /* TREE_BUSYWAIT_INTERVAL should never be less than desired time, or the
333 * time control is broken. But if it happens to be less, we still search
334 * at least 100ms otherwise the move is completely random. */
336 int i
= uct_search_games(&s
);
337 /* Print notifications etc. */
338 uct_search_progress(u
, b
, color
, t
, ti
, &s
, i
);
339 /* Check if we should stop the search. */
340 if (uct_search_check_stop(u
, b
, color
, t
, ti
, &s
, i
))
344 struct uct_thread_ctx
*ctx
= uct_search_stop();
345 if (UDEBUGL(2)) tree_dump(t
, u
->dumpthres
);
347 fprintf(stderr
, "(avg score %f/%d; dynkomi's %f/%d value %f/%d)\n",
348 t
->avg_score
.value
, t
->avg_score
.playouts
,
349 u
->dynkomi
->score
.value
, u
->dynkomi
->score
.playouts
,
350 u
->dynkomi
->value
.value
, u
->dynkomi
->value
.playouts
);
352 uct_progress_status(u
, t
, color
, ctx
->games
, NULL
);
354 u
->played_own
+= ctx
->games
;
358 /* Start pondering background with @color to play. */
360 uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
)
363 fprintf(stderr
, "Starting to ponder with color %s\n", stone2str(stone_other(color
)));
366 /* We need a local board copy to ponder upon. */
367 struct board
*b
= malloc2(sizeof(*b
)); board_copy(b
, b0
);
369 /* *b0 did not have the genmove'd move played yet. */
370 struct move m
= { node_coord(t
->root
), t
->root_color
};
371 int res
= board_play(b
, &m
);
373 setup_dynkomi(u
, b
, stone_other(m
.color
));
375 /* Start MCTS manager thread "headless". */
376 static struct uct_search_state s
;
377 uct_search_start(u
, b
, color
, t
, NULL
, &s
);
380 /* uct_search_stop() frontend for the pondering (non-genmove) mode, and
381 * to stop the background search for a slave in the distributed engine. */
383 uct_pondering_stop(struct uct
*u
)
385 if (!thread_manager_running
)
388 /* Stop the thread manager. */
389 struct uct_thread_ctx
*ctx
= uct_search_stop();
391 if (u
->pondering
) fprintf(stderr
, "(pondering) ");
392 uct_progress_status(u
, ctx
->t
, ctx
->color
, ctx
->games
, NULL
);
396 u
->pondering
= false;
402 uct_genmove_setup(struct uct
*u
, struct board
*b
, enum stone color
)
404 if (b
->superko_violation
) {
405 fprintf(stderr
, "!!! WARNING: SUPERKO VIOLATION OCCURED BEFORE THIS MOVE\n");
406 fprintf(stderr
, "Maybe you play with situational instead of positional superko?\n");
407 fprintf(stderr
, "I'm going to ignore the violation, but note that I may miss\n");
408 fprintf(stderr
, "some moves valid under this ruleset because of this.\n");
409 b
->superko_violation
= false;
412 uct_prepare_move(u
, b
, color
);
417 /* How to decide whether to use dynkomi in this game? Since we use
418 * pondering, it's not simple "who-to-play" matter. Decide based on
419 * the last genmove issued. */
420 u
->t
->use_extra_komi
= !!(u
->dynkomi_mask
& color
);
421 setup_dynkomi(u
, b
, color
);
423 if (b
->rules
== RULES_JAPANESE
)
424 u
->territory_scoring
= true;
426 /* Make pessimistic assumption about komi for Japanese rules to
427 * avoid losing by 0.5 when winning by 0.5 with Chinese rules.
428 * The rules usually give the same winner if the integer part of komi
429 * is odd so we adjust the komi only if it is even (for a board of
430 * odd size). We are not trying to get an exact evaluation for rare
431 * cases of seki. For details see http://home.snafu.de/jasiek/parity.html */
432 if (u
->territory_scoring
&& (((int)floor(b
->komi
) + board_size(b
)) & 1)) {
433 b
->komi
+= (color
== S_BLACK
? 1.0 : -1.0);
435 fprintf(stderr
, "Setting komi to %.1f assuming Japanese rules\n",
441 uct_genmove(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
443 double start_time
= time_now();
444 struct uct
*u
= e
->data
;
445 u
->pass_all_alive
|= pass_all_alive
;
446 uct_pondering_stop(u
);
447 uct_genmove_setup(u
, b
, color
);
449 /* Start the Monte Carlo Tree Search! */
450 int base_playouts
= u
->t
->root
->u
.playouts
;
451 int played_games
= uct_search(u
, b
, ti
, color
, u
->t
, false);
454 struct tree_node
*best
;
455 best
= uct_search_result(u
, b
, color
, u
->pass_all_alive
, played_games
, base_playouts
, &best_coord
);
458 double time
= time_now() - start_time
+ 0.000001; /* avoid divide by zero */
459 fprintf(stderr
, "genmove in %0.2fs (%d games/s, %d games/s/thread)\n",
460 time
, (int)(played_games
/time
), (int)(played_games
/time
/u
->threads
));
463 uct_progress_status(u
, u
->t
, color
, played_games
, &best_coord
);
466 /* Pass or resign. */
467 if (is_pass(best_coord
))
468 u
->initial_extra_komi
= u
->t
->extra_komi
;
470 return coord_copy(best_coord
);
472 tree_promote_node(u
->t
, &best
);
474 /* After a pass, pondering is harmful for two reasons:
475 * (i) We might keep pondering even when the game is over.
476 * Of course this is the case for opponent resign as well.
477 * (ii) More importantly, the ownermap will get skewed since
478 * the UCT will start cutting off any playouts. */
479 if (u
->pondering_opt
&& !is_pass(node_coord(best
))) {
480 uct_pondering_start(u
, b
, u
->t
, stone_other(color
));
482 return coord_copy(best_coord
);
487 uct_gentbook(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
)
489 struct uct
*u
= e
->data
;
490 if (!u
->t
) uct_prepare_move(u
, b
, color
);
493 if (ti
->dim
== TD_GAMES
) {
494 /* Don't count in games that already went into the tbook. */
495 ti
->len
.games
+= u
->t
->root
->u
.playouts
;
497 uct_search(u
, b
, ti
, color
, u
->t
, true);
499 assert(ti
->dim
== TD_GAMES
);
500 tree_save(u
->t
, b
, ti
->len
.games
/ 100);
506 uct_dumptbook(struct engine
*e
, struct board
*b
, enum stone color
)
508 struct uct
*u
= e
->data
;
509 struct tree
*t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0,
510 u
->max_pruned_size
, u
->pruning_threshold
, u
->local_tree_aging
, 0);
518 uct_evaluate_one(struct engine
*e
, struct board
*b
, struct time_info
*ti
, coord_t c
, enum stone color
)
520 struct uct
*u
= e
->data
;
524 struct move m
= { c
, color
};
525 int res
= board_play(&b2
, &m
);
528 color
= stone_other(color
);
530 if (u
->t
) reset_state(u
);
531 uct_prepare_move(u
, &b2
, color
);
535 uct_search(u
, &b2
, ti
, color
, u
->t
, true);
536 struct tree_node
*best
= u
->policy
->choose(u
->policy
, u
->t
->root
, &b2
, color
, resign
);
538 bestval
= NAN
; // the opponent has no reply!
540 bestval
= tree_node_get_value(u
->t
, 1, best
->u
.value
);
543 reset_state(u
); // clean our junk
545 return isnan(bestval
) ? NAN
: 1.0f
- bestval
;
549 uct_evaluate(struct engine
*e
, struct board
*b
, struct time_info
*ti
, floating_t
*vals
, enum stone color
)
551 for (int i
= 0; i
< b
->flen
; i
++) {
552 if (is_pass(b
->f
[i
]))
555 vals
[i
] = uct_evaluate_one(e
, b
, ti
, b
->f
[i
], color
);
561 uct_state_init(char *arg
, struct board
*b
)
563 struct uct
*u
= calloc2(1, sizeof(struct uct
));
564 bool pat_setup
= false;
566 u
->debug_level
= debug_level
;
567 u
->reportfreq
= 10000;
568 u
->gamelen
= MC_GAMELEN
;
569 u
->resign_threshold
= 0.2;
570 u
->sure_win_threshold
= 0.95;
572 u
->significant_threshold
= 50;
575 u
->playout_amaf
= true;
576 u
->amaf_prior
= false;
577 u
->max_tree_size
= 1408ULL * 1048576;
578 u
->fast_alloc
= true;
579 u
->pruning_threshold
= 0;
582 u
->thread_model
= TM_TREEVL
;
585 u
->pondering_opt
= true;
587 u
->fuseki_end
= 20; // max time at 361*20% = 72 moves (our 36th move, still 99 to play)
588 u
->yose_start
= 40; // (100-40-25)*361/100/2 = 63 moves still to play by us then
589 u
->bestr_ratio
= 0.02;
590 // 2.5 is clearly too much, but seems to compensate well for overly stern time allocations.
591 // TODO: Further tuning and experiments with better time allocation schemes.
592 u
->best2_ratio
= 2.5;
593 // Higher values of max_maintime_ratio sometimes cause severe time trouble in tournaments
594 // It might be necessary to reduce it to 1.5 on large board, but more tuning is needed.
595 u
->max_maintime_ratio
= 2.0;
597 u
->val_scale
= 0; u
->val_points
= 40;
598 u
->dynkomi_interval
= 1000;
599 u
->dynkomi_mask
= S_BLACK
| S_WHITE
;
602 u
->local_tree_aging
= 80;
603 u
->local_tree_depth_decay
= 1.5;
604 u
->local_tree_eval
= LTE_ROOT
;
605 u
->local_tree_neival
= true;
609 u
->stats_delay
= 0.01; // 10 ms
610 u
->shared_levels
= 1;
612 u
->plugins
= pluginset_init(b
);
614 u
->jdict
= joseki_load(b
->size
);
617 char *optspec
, *next
= arg
;
620 next
+= strcspn(next
, ",");
621 if (*next
) { *next
++ = 0; } else { *next
= 0; }
623 char *optname
= optspec
;
624 char *optval
= strchr(optspec
, '=');
625 if (optval
) *optval
++ = 0;
629 if (!strcasecmp(optname
, "debug")) {
631 u
->debug_level
= atoi(optval
);
634 } else if (!strcasecmp(optname
, "reporting") && optval
) {
635 /* The format of output for detailed progress
636 * information (such as current best move and
637 * its value, etc.). */
638 if (!strcasecmp(optval
, "text")) {
639 /* Plaintext traditional output. */
640 u
->reporting
= UR_TEXT
;
641 } else if (!strcasecmp(optval
, "json")) {
642 /* JSON output. Implies debug=0. */
643 u
->reporting
= UR_JSON
;
645 } else if (!strcasecmp(optval
, "jsonbig")) {
646 /* JSON output, but much more detailed.
647 * Implies debug=0. */
648 u
->reporting
= UR_JSON_BIG
;
651 fprintf(stderr
, "UCT: Invalid reporting format %s\n", optval
);
654 } else if (!strcasecmp(optname
, "reportfreq") && optval
) {
655 /* The progress information line will be shown
656 * every <reportfreq> simulations. */
657 u
->reportfreq
= atoi(optval
);
658 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
659 /* When dumping the UCT tree on output, include
660 * nodes with at least this many playouts.
661 * (This value is re-scaled "intelligently"
662 * in case of very large trees.) */
663 u
->dumpthres
= atoi(optval
);
664 } else if (!strcasecmp(optname
, "resign_threshold") && optval
) {
665 /* Resign when this ratio of games is lost
666 * after GJ_MINGAMES sample is taken. */
667 u
->resign_threshold
= atof(optval
);
668 } else if (!strcasecmp(optname
, "sure_win_threshold") && optval
) {
669 /* Stop reading when this ratio of games is won
670 * after PLAYOUT_EARLY_BREAK_MIN sample is
671 * taken. (Prevents stupid time losses,
672 * friendly to human opponents.) */
673 u
->sure_win_threshold
= atof(optval
);
674 } else if (!strcasecmp(optname
, "force_seed") && optval
) {
675 /* Set RNG seed at the tree setup. */
676 u
->force_seed
= atoi(optval
);
677 } else if (!strcasecmp(optname
, "no_tbook")) {
678 /* Disable UCT opening tbook. */
680 } else if (!strcasecmp(optname
, "pass_all_alive")) {
681 /* Whether to consider passing only after all
682 * dead groups were removed from the board;
683 * this is like all genmoves are in fact
684 * kgs-genmove_cleanup. */
685 u
->pass_all_alive
= !optval
|| atoi(optval
);
686 } else if (!strcasecmp(optname
, "allow_losing_pass")) {
687 /* Whether to consider passing in a clear
688 * but losing situation, to be scored as a loss
690 u
->allow_losing_pass
= !optval
|| atoi(optval
);
691 } else if (!strcasecmp(optname
, "territory_scoring")) {
692 /* Use territory scoring (default is area scoring).
693 * An explicit kgs-rules command overrides this. */
694 u
->territory_scoring
= !optval
|| atoi(optval
);
695 } else if (!strcasecmp(optname
, "stones_only")) {
696 /* Do not count eyes. Nice to teach go to kids.
697 * http://strasbourg.jeudego.org/regle_strasbourgeoise.htm */
698 b
->rules
= RULES_STONES_ONLY
;
699 u
->pass_all_alive
= true;
700 } else if (!strcasecmp(optname
, "banner") && optval
) {
701 /* Additional banner string. This must come as the
702 * last engine parameter. */
703 if (*next
) *--next
= ',';
704 u
->banner
= strdup(optval
);
706 } else if (!strcasecmp(optname
, "plugin") && optval
) {
707 /* Load an external plugin; filename goes before the colon,
708 * extra arguments after the colon. */
709 char *pluginarg
= strchr(optval
, ':');
712 plugin_load(u
->plugins
, optval
, pluginarg
);
714 /** UCT behavior and policies */
716 } else if ((!strcasecmp(optname
, "policy")
717 /* Node selection policy. ucb1amaf is the
718 * default policy implementing RAVE, while
719 * ucb1 is the simple exploration/exploitation
720 * policy. Policies can take further extra
722 || !strcasecmp(optname
, "random_policy")) && optval
) {
723 /* A policy to be used randomly with small
724 * chance instead of the default policy. */
725 char *policyarg
= strchr(optval
, ':');
726 struct uct_policy
**p
= !strcasecmp(optname
, "policy") ? &u
->policy
: &u
->random_policy
;
729 if (!strcasecmp(optval
, "ucb1")) {
730 *p
= policy_ucb1_init(u
, policyarg
);
731 } else if (!strcasecmp(optval
, "ucb1amaf")) {
732 *p
= policy_ucb1amaf_init(u
, policyarg
, b
);
734 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
737 } else if (!strcasecmp(optname
, "playout") && optval
) {
738 /* Random simulation (playout) policy.
739 * moggy is the default policy with large
740 * amount of domain-specific knowledge and
741 * heuristics. light is a simple uniformly
742 * random move selection policy. */
743 char *playoutarg
= strchr(optval
, ':');
746 if (!strcasecmp(optval
, "moggy")) {
747 u
->playout
= playout_moggy_init(playoutarg
, b
, u
->jdict
);
748 } else if (!strcasecmp(optval
, "light")) {
749 u
->playout
= playout_light_init(playoutarg
, b
);
751 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
754 } else if (!strcasecmp(optname
, "prior") && optval
) {
755 /* Node priors policy. When expanding a node,
756 * it will seed node values heuristically
757 * (most importantly, based on playout policy
758 * opinion, but also with regard to other
759 * things). See uct/prior.c for details.
760 * Use prior=eqex=0 to disable priors. */
761 u
->prior
= uct_prior_init(optval
, b
, u
);
762 } else if (!strcasecmp(optname
, "mercy") && optval
) {
763 /* Minimal difference of black/white captures
764 * to stop playout - "Mercy Rule". Speeds up
765 * hopeless playouts at the expense of some
767 u
->mercymin
= atoi(optval
);
768 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
769 /* Maximum length of single simulation
771 u
->gamelen
= atoi(optval
);
772 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
773 /* Expand UCT nodes after it has been
774 * visited this many times. */
775 u
->expand_p
= atoi(optval
);
776 } else if (!strcasecmp(optname
, "random_policy_chance") && optval
) {
777 /* If specified (N), with probability 1/N, random_policy policy
778 * descend is used instead of main policy descend; useful
779 * if specified policy (e.g. UCB1AMAF) can make unduly biased
780 * choices sometimes, you can fall back to e.g.
781 * random_policy=UCB1. */
782 u
->random_policy_chance
= atoi(optval
);
784 /** General AMAF behavior */
785 /* (Only relevant if the policy supports AMAF.
786 * More variables can be tuned as policy
789 } else if (!strcasecmp(optname
, "playout_amaf")) {
790 /* Whether to include random playout moves in
791 * AMAF as well. (Otherwise, only tree moves
792 * are included in AMAF. Of course makes sense
793 * only in connection with an AMAF policy.) */
794 /* with-without: 55.5% (+-4.1) */
795 if (optval
&& *optval
== '0')
796 u
->playout_amaf
= false;
798 u
->playout_amaf
= true;
799 } else if (!strcasecmp(optname
, "playout_amaf_cutoff") && optval
) {
800 /* Keep only first N% of playout stage AMAF
802 u
->playout_amaf_cutoff
= atoi(optval
);
803 } else if (!strcasecmp(optname
, "amaf_prior") && optval
) {
804 /* In node policy, consider prior values
805 * part of the real result term or part
806 * of the AMAF term? */
807 u
->amaf_prior
= atoi(optval
);
809 /** Performance and memory management */
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
, "tree")) {
817 /* Tree parallelization - all threads
818 * grind on the same tree. */
819 u
->thread_model
= TM_TREE
;
821 } else if (!strcasecmp(optval
, "treevl")) {
822 /* Tree parallelization, but also
823 * with virtual losses - this discou-
824 * rages most threads choosing the
825 * same tree branches to read. */
826 u
->thread_model
= TM_TREEVL
;
828 fprintf(stderr
, "UCT: Invalid thread model %s\n", optval
);
831 } else if (!strcasecmp(optname
, "virtual_loss") && optval
) {
832 /* Number of virtual losses added before evaluating a node. */
833 u
->virtual_loss
= atoi(optval
);
834 } else if (!strcasecmp(optname
, "pondering")) {
835 /* Keep searching even during opponent's turn. */
836 u
->pondering_opt
= !optval
|| atoi(optval
);
837 } else if (!strcasecmp(optname
, "max_tree_size") && optval
) {
838 /* Maximum amount of memory [MiB] consumed by the move tree.
839 * For fast_alloc it includes the temp tree used for pruning.
840 * Default is 3072 (3 GiB). */
841 u
->max_tree_size
= atol(optval
) * 1048576;
842 } else if (!strcasecmp(optname
, "fast_alloc")) {
843 u
->fast_alloc
= !optval
|| atoi(optval
);
844 } else if (!strcasecmp(optname
, "pruning_threshold") && optval
) {
845 /* Force pruning at beginning of a move if the tree consumes
846 * more than this [MiB]. Default is 10% of max_tree_size.
847 * Increase to reduce pruning time overhead if memory is plentiful.
848 * This option is meaningful only for fast_alloc. */
849 u
->pruning_threshold
= atol(optval
) * 1048576;
853 } else if (!strcasecmp(optname
, "best2_ratio") && optval
) {
854 /* If set, prolong simulating while
855 * first_best/second_best playouts ratio
856 * is less than best2_ratio. */
857 u
->best2_ratio
= atof(optval
);
858 } else if (!strcasecmp(optname
, "bestr_ratio") && optval
) {
859 /* If set, prolong simulating while
860 * best,best_best_child values delta
861 * is more than bestr_ratio. */
862 u
->bestr_ratio
= atof(optval
);
863 } else if (!strcasecmp(optname
, "max_maintime_ratio") && optval
) {
864 /* If set and while not in byoyomi, prolong simulating no more than
865 * max_maintime_ratio times the normal desired thinking time. */
866 u
->max_maintime_ratio
= atof(optval
);
867 } else if (!strcasecmp(optname
, "fuseki_end") && optval
) {
868 /* At the very beginning it's not worth thinking
869 * too long because the playout evaluations are
870 * very noisy. So gradually increase the thinking
871 * time up to maximum when fuseki_end percent
872 * of the board has been played.
873 * This only applies if we are not in byoyomi. */
874 u
->fuseki_end
= atoi(optval
);
875 } else if (!strcasecmp(optname
, "yose_start") && optval
) {
876 /* When yose_start percent of the board has been
877 * played, or if we are in byoyomi, stop spending
878 * more time and spread the remaining time
880 * Between fuseki_end and yose_start, we spend
881 * a constant proportion of the remaining time
882 * on each move. (yose_start should actually
883 * be much earlier than when real yose start,
884 * but "yose" is a good short name to convey
886 u
->yose_start
= atoi(optval
);
890 } else if (!strcasecmp(optname
, "dynkomi") && optval
) {
891 /* Dynamic komi approach; there are multiple
892 * ways to adjust komi dynamically throughout
893 * play. We currently support two: */
894 char *dynkomiarg
= strchr(optval
, ':');
897 if (!strcasecmp(optval
, "none")) {
898 u
->dynkomi
= uct_dynkomi_init_none(u
, dynkomiarg
, b
);
899 } else if (!strcasecmp(optval
, "linear")) {
900 /* You should set dynkomi_mask=1 or a very low
901 * handicap_value for white. */
902 u
->dynkomi
= uct_dynkomi_init_linear(u
, dynkomiarg
, b
);
903 } else if (!strcasecmp(optval
, "adaptive")) {
904 /* There are many more knobs to
905 * crank - see uct/dynkomi.c. */
906 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, dynkomiarg
, b
);
908 fprintf(stderr
, "UCT: Invalid dynkomi mode %s\n", optval
);
911 } else if (!strcasecmp(optname
, "dynkomi_mask") && optval
) {
912 /* Bitmask of colors the player must be
913 * for dynkomi be applied; the default dynkomi_mask=3 allows
914 * dynkomi even in games where Pachi is white. */
915 u
->dynkomi_mask
= atoi(optval
);
916 } else if (!strcasecmp(optname
, "dynkomi_interval") && optval
) {
917 /* If non-zero, re-adjust dynamic komi
918 * throughout a single genmove reading,
919 * roughly every N simulations. */
920 /* XXX: Does not work with tree
921 * parallelization. */
922 u
->dynkomi_interval
= atoi(optval
);
923 } else if (!strcasecmp(optname
, "extra_komi") && optval
) {
924 /* Initial dynamic komi settings. This
925 * is useful for the adaptive dynkomi
926 * policy as the value to start with
927 * (this is NOT kept fixed) in case
928 * there is not enough time in the search
929 * to adjust the value properly (e.g. the
930 * game was interrupted). */
931 u
->initial_extra_komi
= atof(optval
);
933 /** Node value result scaling */
935 } else if (!strcasecmp(optname
, "val_scale") && optval
) {
936 /* How much of the game result value should be
937 * influenced by win size. Zero means it isn't. */
938 u
->val_scale
= atof(optval
);
939 } else if (!strcasecmp(optname
, "val_points") && optval
) {
940 /* Maximum size of win to be scaled into game
941 * result value. Zero means boardsize^2. */
942 u
->val_points
= atoi(optval
) * 2; // result values are doubled
943 } else if (!strcasecmp(optname
, "val_extra")) {
944 /* If false, the score coefficient will be simply
945 * added to the value, instead of scaling the result
946 * coefficient because of it. */
947 u
->val_extra
= !optval
|| atoi(optval
);
948 } else if (!strcasecmp(optname
, "val_byavg")) {
949 /* If true, the score included in the value will
950 * be relative to average score in the current
951 * search episode inst. of jigo. */
952 u
->val_byavg
= !optval
|| atoi(optval
);
953 } else if (!strcasecmp(optname
, "val_bytemp")) {
954 /* If true, the value scaling coefficient
955 * is different based on value extremity
956 * (dist. from 0.5), linear between
957 * val_bytemp_min, val_scale. */
958 u
->val_bytemp
= !optval
|| atoi(optval
);
959 } else if (!strcasecmp(optname
, "val_bytemp_min") && optval
) {
960 /* Minimum val_scale in case of val_bytemp. */
961 u
->val_bytemp_min
= atof(optval
);
964 /* (Purely experimental. Does not work - yet!) */
966 } else if (!strcasecmp(optname
, "local_tree")) {
967 /* Whether to bias exploration by local tree values. */
968 u
->local_tree
= !optval
|| atoi(optval
);
969 } else if (!strcasecmp(optname
, "tenuki_d") && optval
) {
970 /* Tenuki distance at which to break the local tree. */
971 u
->tenuki_d
= atoi(optval
);
972 if (u
->tenuki_d
> TREE_NODE_D_MAX
+ 1) {
973 fprintf(stderr
, "uct: tenuki_d must not be larger than TREE_NODE_D_MAX+1 %d\n", TREE_NODE_D_MAX
+ 1);
976 } else if (!strcasecmp(optname
, "local_tree_aging") && optval
) {
977 /* How much to reduce local tree values between moves. */
978 u
->local_tree_aging
= atof(optval
);
979 } else if (!strcasecmp(optname
, "local_tree_depth_decay") && optval
) {
980 /* With value x>0, during the descent the node
981 * contributes 1/x^depth playouts in
982 * the local tree. I.e., with x>1, nodes more
983 * distant from local situation contribute more
984 * than nodes near the root. */
985 u
->local_tree_depth_decay
= atof(optval
);
986 } else if (!strcasecmp(optname
, "local_tree_allseq")) {
987 /* If disabled, only complete sequences are stored
988 * in the local tree. If this is on, also
989 * subsequences starting at each move are stored. */
990 u
->local_tree_allseq
= !optval
|| atoi(optval
);
991 } else if (!strcasecmp(optname
, "local_tree_neival")) {
992 /* If disabled, local node value is not
993 * computed just based on terminal status
994 * of the coordinate, but also its neighbors. */
995 u
->local_tree_neival
= !optval
|| atoi(optval
);
996 } else if (!strcasecmp(optname
, "local_tree_eval")) {
997 /* How is the value inserted in the local tree
999 if (!strcasecmp(optval
, "root"))
1000 /* All moves within a tree branch are
1001 * considered wrt. their merit
1002 * reaching tachtical goal of making
1003 * the first move in the branch
1005 u
->local_tree_eval
= LTE_ROOT
;
1006 else if (!strcasecmp(optval
, "each"))
1007 /* Each move is considered wrt.
1008 * its own survival. */
1009 u
->local_tree_eval
= LTE_EACH
;
1010 else if (!strcasecmp(optval
, "total"))
1011 /* The tactical goal is the survival
1012 * of all the moves of my color and
1013 * non-survival of all the opponent
1014 * moves. Local values (and their
1015 * inverses) are averaged. */
1016 u
->local_tree_eval
= LTE_TOTAL
;
1018 fprintf(stderr
, "uct: unknown local_tree_eval %s\n", optval
);
1021 } else if (!strcasecmp(optname
, "local_tree_rootchoose")) {
1022 /* If disabled, only moves within the local
1023 * tree branch are considered; the values
1024 * of the branch roots (i.e. root children)
1025 * are ignored. This may make sense together
1026 * with eval!=each, we consider only moves
1027 * that influence the goal, not the "rating"
1028 * of the goal itself. (The real solution
1029 * will be probably using criticality to pick
1030 * local tree branches.) */
1031 u
->local_tree_rootchoose
= !optval
|| atoi(optval
);
1033 /** Other heuristics */
1034 } else if (!strcasecmp(optname
, "patterns")) {
1035 /* Load pattern database. Various modules
1036 * (priors, policies etc.) may make use
1037 * of this database. They will request
1038 * it automatically in that case, but you
1039 * can use this option to tweak the pattern
1041 patterns_init(&u
->pat
, optval
, false, true);
1042 u
->want_pat
= pat_setup
= true;
1043 } else if (!strcasecmp(optname
, "significant_threshold") && optval
) {
1044 /* Some heuristics (XXX: none in mainline) rely
1045 * on the knowledge of the last "significant"
1046 * node in the descent. Such a node is
1047 * considered reasonably trustworthy to carry
1048 * some meaningful information in the values
1049 * of the node and its children. */
1050 u
->significant_threshold
= atoi(optval
);
1051 } else if (!strcasecmp(optname
, "libmap")) {
1052 /* Online learning of move tactical ratings by
1054 libmap_setup(optval
);
1057 /** Distributed engine slaves setup */
1059 } else if (!strcasecmp(optname
, "slave")) {
1060 /* Act as slave for the distributed engine. */
1061 u
->slave
= !optval
|| atoi(optval
);
1062 } else if (!strcasecmp(optname
, "slave_index") && optval
) {
1063 /* Optional index if per-slave behavior is desired.
1064 * Must be given as index/max */
1065 u
->slave_index
= atoi(optval
);
1066 char *p
= strchr(optval
, '/');
1067 if (p
) u
->max_slaves
= atoi(++p
);
1068 } else if (!strcasecmp(optname
, "shared_nodes") && optval
) {
1069 /* Share at most shared_nodes between master and slave at each genmoves.
1070 * Must use the same value in master and slaves. */
1071 u
->shared_nodes
= atoi(optval
);
1072 } else if (!strcasecmp(optname
, "shared_levels") && optval
) {
1073 /* Share only nodes of level <= shared_levels. */
1074 u
->shared_levels
= atoi(optval
);
1075 } else if (!strcasecmp(optname
, "stats_hbits") && optval
) {
1076 /* Set hash table size to 2^stats_hbits for the shared stats. */
1077 u
->stats_hbits
= atoi(optval
);
1078 } else if (!strcasecmp(optname
, "stats_delay") && optval
) {
1079 /* How long to wait in slave for initial stats to build up before
1080 * replying to the genmoves command (in ms) */
1081 u
->stats_delay
= 0.001 * atof(optval
);
1085 } else if (!strcasecmp(optname
, "maximize_score")) {
1086 /* A combination of settings that will make
1087 * Pachi try to maximize his points (instead
1088 * of playing slack yose) or minimize his loss
1089 * (and proceed to counting even when losing). */
1090 /* Please note that this preset might be
1091 * somewhat weaker than normal Pachi, and the
1092 * score maximization is approximate; point size
1093 * of win/loss still should not be used to judge
1094 * strength of Pachi or the opponent. */
1095 /* See README for some further notes. */
1096 if (!optval
|| atoi(optval
)) {
1097 /* Allow scoring a lost game. */
1098 u
->allow_losing_pass
= true;
1099 /* Make Pachi keep his calm when losing
1100 * and/or maintain winning marging. */
1101 /* Do not play games that are losing
1103 /* XXX: komi_ratchet_age=40000 is necessary
1104 * with losing_komi_ratchet, but 40000
1105 * is somewhat arbitrary value. */
1106 char dynkomi_args
[] = "losing_komi_ratchet:komi_ratchet_age=60000:no_komi_at_game_end=0:max_losing_komi=30";
1107 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, dynkomi_args
, b
);
1108 /* XXX: Values arbitrary so far. */
1109 /* XXX: Also, is bytemp sensible when
1110 * combined with dynamic komi?! */
1111 u
->val_scale
= 0.01;
1112 u
->val_bytemp
= true;
1113 u
->val_bytemp_min
= 0.001;
1114 u
->val_byavg
= true;
1118 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
1125 u
->policy
= policy_ucb1amaf_init(u
, NULL
, b
);
1127 if (!!u
->random_policy_chance
^ !!u
->random_policy
) {
1128 fprintf(stderr
, "uct: Only one of random_policy and random_policy_chance is set\n");
1132 if (!u
->local_tree
) {
1133 /* No ltree aging. */
1134 u
->local_tree_aging
= 1.0f
;
1137 if (u
->fast_alloc
) {
1138 if (u
->pruning_threshold
< u
->max_tree_size
/ 10)
1139 u
->pruning_threshold
= u
->max_tree_size
/ 10;
1140 if (u
->pruning_threshold
> u
->max_tree_size
/ 2)
1141 u
->pruning_threshold
= u
->max_tree_size
/ 2;
1143 /* Limit pruning temp space to 20% of memory. Beyond this we discard
1144 * the nodes and recompute them at the next move if necessary. */
1145 u
->max_pruned_size
= u
->max_tree_size
/ 5;
1146 u
->max_tree_size
-= u
->max_pruned_size
;
1148 /* Reserve 5% memory in case the background free() are slower
1149 * than the concurrent allocations. */
1150 u
->max_tree_size
-= u
->max_tree_size
/ 20;
1154 u
->prior
= uct_prior_init(NULL
, b
, u
);
1157 u
->playout
= playout_moggy_init(NULL
, b
, u
->jdict
);
1158 if (!u
->playout
->debug_level
)
1159 u
->playout
->debug_level
= u
->debug_level
;
1161 if (u
->want_pat
&& !pat_setup
)
1162 patterns_init(&u
->pat
, NULL
, false, true);
1164 u
->ownermap
.map
= malloc2(board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
1167 if (!u
->stats_hbits
) u
->stats_hbits
= DEFAULT_STATS_HBITS
;
1168 if (!u
->shared_nodes
) u
->shared_nodes
= DEFAULT_SHARED_NODES
;
1169 assert(u
->shared_levels
* board_bits2(b
) <= 8 * (int)sizeof(path_t
));
1173 u
->dynkomi
= board_small(b
) ? uct_dynkomi_init_none(u
, NULL
, b
)
1174 : uct_dynkomi_init_linear(u
, NULL
, b
);
1176 /* Some things remain uninitialized for now - the opening tbook
1177 * is not loaded and the tree not set up. */
1178 /* This will be initialized in setup_state() at the first move
1179 * received/requested. This is because right now we are not aware
1180 * about any komi or handicap setup and such. */
1186 engine_uct_init(char *arg
, struct board
*b
)
1188 struct uct
*u
= uct_state_init(arg
, b
);
1189 struct engine
*e
= calloc2(1, sizeof(struct engine
));
1191 e
->printhook
= uct_printhook_ownermap
;
1192 e
->notify_play
= uct_notify_play
;
1195 e
->result
= uct_result
;
1196 e
->genmove
= uct_genmove
;
1197 e
->genmoves
= uct_genmoves
;
1198 e
->evaluate
= uct_evaluate
;
1199 e
->dead_group_list
= uct_dead_group_list
;
1203 e
->notify
= uct_notify
;
1205 const char banner
[] = "If you believe you have won but I am still playing, "
1206 "please help me understand by capturing all dead stones. "
1207 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
1208 if (!u
->banner
) u
->banner
= "";
1209 e
->comment
= malloc2(sizeof(banner
) + strlen(u
->banner
) + 1);
1210 sprintf(e
->comment
, "%s %s", banner
, u
->banner
);