16 #include "joseki/base.h"
18 #include "playout/moggy.h"
19 #include "playout/light.h"
20 #include "tactics/goals.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_stop(struct engine
*e
)
297 /* This is called on game over notification. However, an undo
298 * and game resume can follow, so don't panic yet and just
299 * relax and stop thinking so that we don't waste CPU. */
300 struct uct
*u
= e
->data
;
301 uct_pondering_stop(u
);
305 uct_done(struct engine
*e
)
307 /* This is called on engine reset, especially when clear_board
308 * is received and new game should begin. */
309 struct uct
*u
= e
->data
;
310 uct_pondering_stop(u
);
311 if (u
->t
) reset_state(u
);
312 free(u
->ownermap
.map
);
315 free(u
->random_policy
);
316 playout_policy_done(u
->playout
);
317 uct_prior_done(u
->prior
);
318 joseki_done(u
->jdict
);
319 pluginset_done(u
->plugins
);
324 /* Run time-limited MCTS search on foreground. */
326 uct_search(struct uct
*u
, struct board
*b
, struct time_info
*ti
, enum stone color
, struct tree
*t
, bool print_progress
)
328 struct uct_search_state s
;
329 uct_search_start(u
, b
, color
, t
, ti
, &s
);
330 if (UDEBUGL(2) && s
.base_playouts
> 0)
331 fprintf(stderr
, "<pre-simulated %d games>\n", s
.base_playouts
);
333 /* The search tree is ctx->t. This is currently == . It is important
334 * to reference ctx->t directly since the
335 * thread manager will swap the tree pointer asynchronously. */
337 /* Now, just periodically poll the search tree. */
338 /* Note that in case of TD_GAMES, threads will not wait for
339 * the uct_search_check_stop() signalization. */
341 time_sleep(TREE_BUSYWAIT_INTERVAL
);
342 /* TREE_BUSYWAIT_INTERVAL should never be less than desired time, or the
343 * time control is broken. But if it happens to be less, we still search
344 * at least 100ms otherwise the move is completely random. */
346 int i
= uct_search_games(&s
);
347 /* Print notifications etc. */
348 uct_search_progress(u
, b
, color
, t
, ti
, &s
, i
);
349 /* Check if we should stop the search. */
350 if (uct_search_check_stop(u
, b
, color
, t
, ti
, &s
, i
))
354 struct uct_thread_ctx
*ctx
= uct_search_stop();
355 if (UDEBUGL(2)) tree_dump(t
, u
->dumpthres
);
357 fprintf(stderr
, "(avg score %f/%d; dynkomi's %f/%d value %f/%d)\n",
358 t
->avg_score
.value
, t
->avg_score
.playouts
,
359 u
->dynkomi
->score
.value
, u
->dynkomi
->score
.playouts
,
360 u
->dynkomi
->value
.value
, u
->dynkomi
->value
.playouts
);
362 uct_progress_status(u
, t
, color
, ctx
->games
, NULL
);
364 u
->played_own
+= ctx
->games
;
368 /* Start pondering background with @color to play. */
370 uct_pondering_start(struct uct
*u
, struct board
*b0
, struct tree
*t
, enum stone color
)
373 fprintf(stderr
, "Starting to ponder with color %s\n", stone2str(stone_other(color
)));
376 /* We need a local board copy to ponder upon. */
377 struct board
*b
= malloc2(sizeof(*b
)); board_copy(b
, b0
);
379 /* *b0 did not have the genmove'd move played yet. */
380 struct move m
= { node_coord(t
->root
), t
->root_color
};
381 int res
= board_play(b
, &m
);
383 setup_dynkomi(u
, b
, stone_other(m
.color
));
385 /* Start MCTS manager thread "headless". */
386 static struct uct_search_state s
;
387 uct_search_start(u
, b
, color
, t
, NULL
, &s
);
390 /* uct_search_stop() frontend for the pondering (non-genmove) mode, and
391 * to stop the background search for a slave in the distributed engine. */
393 uct_pondering_stop(struct uct
*u
)
395 if (!thread_manager_running
)
398 /* Stop the thread manager. */
399 struct uct_thread_ctx
*ctx
= uct_search_stop();
401 if (u
->pondering
) fprintf(stderr
, "(pondering) ");
402 uct_progress_status(u
, ctx
->t
, ctx
->color
, ctx
->games
, NULL
);
406 u
->pondering
= false;
412 uct_genmove_setup(struct uct
*u
, struct board
*b
, enum stone color
)
414 if (b
->superko_violation
) {
415 fprintf(stderr
, "!!! WARNING: SUPERKO VIOLATION OCCURED BEFORE THIS MOVE\n");
416 fprintf(stderr
, "Maybe you play with situational instead of positional superko?\n");
417 fprintf(stderr
, "I'm going to ignore the violation, but note that I may miss\n");
418 fprintf(stderr
, "some moves valid under this ruleset because of this.\n");
419 b
->superko_violation
= false;
422 uct_prepare_move(u
, b
, color
);
427 /* How to decide whether to use dynkomi in this game? Since we use
428 * pondering, it's not simple "who-to-play" matter. Decide based on
429 * the last genmove issued. */
430 u
->t
->use_extra_komi
= !!(u
->dynkomi_mask
& color
);
431 setup_dynkomi(u
, b
, color
);
433 if (b
->rules
== RULES_JAPANESE
)
434 u
->territory_scoring
= true;
436 /* Make pessimistic assumption about komi for Japanese rules to
437 * avoid losing by 0.5 when winning by 0.5 with Chinese rules.
438 * The rules usually give the same winner if the integer part of komi
439 * is odd so we adjust the komi only if it is even (for a board of
440 * odd size). We are not trying to get an exact evaluation for rare
441 * cases of seki. For details see http://home.snafu.de/jasiek/parity.html */
442 if (u
->territory_scoring
&& (((int)floor(b
->komi
) + board_size(b
)) & 1)) {
443 b
->komi
+= (color
== S_BLACK
? 1.0 : -1.0);
445 fprintf(stderr
, "Setting komi to %.1f assuming Japanese rules\n",
451 uct_genmove(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
453 double start_time
= time_now();
454 struct uct
*u
= e
->data
;
455 u
->pass_all_alive
|= pass_all_alive
;
456 uct_pondering_stop(u
);
457 uct_genmove_setup(u
, b
, color
);
459 /* Start the Monte Carlo Tree Search! */
460 int base_playouts
= u
->t
->root
->u
.playouts
;
461 int played_games
= uct_search(u
, b
, ti
, color
, u
->t
, false);
464 struct tree_node
*best
;
465 best
= uct_search_result(u
, b
, color
, u
->pass_all_alive
, played_games
, base_playouts
, &best_coord
);
468 double time
= time_now() - start_time
+ 0.000001; /* avoid divide by zero */
469 fprintf(stderr
, "genmove in %0.2fs (%d games/s, %d games/s/thread)\n",
470 time
, (int)(played_games
/time
), (int)(played_games
/time
/u
->threads
));
473 uct_progress_status(u
, u
->t
, color
, played_games
, &best_coord
);
476 /* Pass or resign. */
477 if (is_pass(best_coord
))
478 u
->initial_extra_komi
= u
->t
->extra_komi
;
480 return coord_copy(best_coord
);
482 tree_promote_node(u
->t
, &best
);
484 /* After a pass, pondering is harmful for two reasons:
485 * (i) We might keep pondering even when the game is over.
486 * Of course this is the case for opponent resign as well.
487 * (ii) More importantly, the ownermap will get skewed since
488 * the UCT will start cutting off any playouts. */
489 if (u
->pondering_opt
&& !is_pass(node_coord(best
))) {
490 uct_pondering_start(u
, b
, u
->t
, stone_other(color
));
492 return coord_copy(best_coord
);
497 uct_gentbook(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
)
499 struct uct
*u
= e
->data
;
500 if (!u
->t
) uct_prepare_move(u
, b
, color
);
503 if (ti
->dim
== TD_GAMES
) {
504 /* Don't count in games that already went into the tbook. */
505 ti
->len
.games
+= u
->t
->root
->u
.playouts
;
507 uct_search(u
, b
, ti
, color
, u
->t
, true);
509 assert(ti
->dim
== TD_GAMES
);
510 tree_save(u
->t
, b
, ti
->len
.games
/ 100);
516 uct_dumptbook(struct engine
*e
, struct board
*b
, enum stone color
)
518 struct uct
*u
= e
->data
;
519 struct tree
*t
= tree_init(b
, color
, u
->fast_alloc
? u
->max_tree_size
: 0,
520 u
->max_pruned_size
, u
->pruning_threshold
, u
->local_tree_aging
, 0);
528 uct_evaluate_one(struct engine
*e
, struct board
*b
, struct time_info
*ti
, coord_t c
, enum stone color
)
530 struct uct
*u
= e
->data
;
534 struct move m
= { c
, color
};
535 int res
= board_play(&b2
, &m
);
538 color
= stone_other(color
);
540 if (u
->t
) reset_state(u
);
541 uct_prepare_move(u
, &b2
, color
);
545 uct_search(u
, &b2
, ti
, color
, u
->t
, true);
546 struct tree_node
*best
= u
->policy
->choose(u
->policy
, u
->t
->root
, &b2
, color
, resign
);
548 bestval
= NAN
; // the opponent has no reply!
550 bestval
= tree_node_get_value(u
->t
, 1, best
->u
.value
);
553 reset_state(u
); // clean our junk
555 return isnan(bestval
) ? NAN
: 1.0f
- bestval
;
559 uct_evaluate(struct engine
*e
, struct board
*b
, struct time_info
*ti
, floating_t
*vals
, enum stone color
)
561 for (int i
= 0; i
< b
->flen
; i
++) {
562 if (is_pass(b
->f
[i
]))
565 vals
[i
] = uct_evaluate_one(e
, b
, ti
, b
->f
[i
], color
);
571 uct_state_init(char *arg
, struct board
*b
)
573 struct uct
*u
= calloc2(1, sizeof(struct uct
));
574 bool pat_setup
= false;
576 u
->debug_level
= debug_level
;
577 u
->reportfreq
= 10000;
578 u
->gamelen
= MC_GAMELEN
;
579 u
->resign_threshold
= 0.2;
580 u
->sure_win_threshold
= 0.95;
582 u
->significant_threshold
= 50;
585 u
->playout_amaf
= true;
586 u
->amaf_prior
= false;
587 u
->max_tree_size
= 1408ULL * 1048576;
588 u
->fast_alloc
= true;
589 u
->pruning_threshold
= 0;
592 u
->thread_model
= TM_TREEVL
;
595 u
->pondering_opt
= true;
597 u
->fuseki_end
= 20; // max time at 361*20% = 72 moves (our 36th move, still 99 to play)
598 u
->yose_start
= 40; // (100-40-25)*361/100/2 = 63 moves still to play by us then
599 u
->bestr_ratio
= 0.02;
600 // 2.5 is clearly too much, but seems to compensate well for overly stern time allocations.
601 // TODO: Further tuning and experiments with better time allocation schemes.
602 u
->best2_ratio
= 2.5;
603 // Higher values of max_maintime_ratio sometimes cause severe time trouble in tournaments
604 // It might be necessary to reduce it to 1.5 on large board, but more tuning is needed.
605 u
->max_maintime_ratio
= 2.0;
607 u
->val_scale
= 0; u
->val_points
= 40;
608 u
->dynkomi_interval
= 1000;
609 u
->dynkomi_mask
= S_BLACK
| S_WHITE
;
612 u
->local_tree_aging
= 80;
613 u
->local_tree_depth_decay
= 1.5;
614 u
->local_tree_eval
= LTE_ROOT
;
615 u
->local_tree_neival
= true;
619 u
->stats_delay
= 0.01; // 10 ms
620 u
->shared_levels
= 1;
622 u
->plugins
= pluginset_init(b
);
624 u
->jdict
= joseki_load(b
->size
);
627 char *optspec
, *next
= arg
;
630 next
+= strcspn(next
, ",");
631 if (*next
) { *next
++ = 0; } else { *next
= 0; }
633 char *optname
= optspec
;
634 char *optval
= strchr(optspec
, '=');
635 if (optval
) *optval
++ = 0;
639 if (!strcasecmp(optname
, "debug")) {
641 u
->debug_level
= atoi(optval
);
644 } else if (!strcasecmp(optname
, "reporting") && optval
) {
645 /* The format of output for detailed progress
646 * information (such as current best move and
647 * its value, etc.). */
648 if (!strcasecmp(optval
, "text")) {
649 /* Plaintext traditional output. */
650 u
->reporting
= UR_TEXT
;
651 } else if (!strcasecmp(optval
, "json")) {
652 /* JSON output. Implies debug=0. */
653 u
->reporting
= UR_JSON
;
655 } else if (!strcasecmp(optval
, "jsonbig")) {
656 /* JSON output, but much more detailed.
657 * Implies debug=0. */
658 u
->reporting
= UR_JSON_BIG
;
661 fprintf(stderr
, "UCT: Invalid reporting format %s\n", optval
);
664 } else if (!strcasecmp(optname
, "reportfreq") && optval
) {
665 /* The progress information line will be shown
666 * every <reportfreq> simulations. */
667 u
->reportfreq
= atoi(optval
);
668 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
669 /* When dumping the UCT tree on output, include
670 * nodes with at least this many playouts.
671 * (This value is re-scaled "intelligently"
672 * in case of very large trees.) */
673 u
->dumpthres
= atoi(optval
);
674 } else if (!strcasecmp(optname
, "resign_threshold") && optval
) {
675 /* Resign when this ratio of games is lost
676 * after GJ_MINGAMES sample is taken. */
677 u
->resign_threshold
= atof(optval
);
678 } else if (!strcasecmp(optname
, "sure_win_threshold") && optval
) {
679 /* Stop reading when this ratio of games is won
680 * after PLAYOUT_EARLY_BREAK_MIN sample is
681 * taken. (Prevents stupid time losses,
682 * friendly to human opponents.) */
683 u
->sure_win_threshold
= atof(optval
);
684 } else if (!strcasecmp(optname
, "force_seed") && optval
) {
685 /* Set RNG seed at the tree setup. */
686 u
->force_seed
= atoi(optval
);
687 } else if (!strcasecmp(optname
, "no_tbook")) {
688 /* Disable UCT opening tbook. */
690 } else if (!strcasecmp(optname
, "pass_all_alive")) {
691 /* Whether to consider passing only after all
692 * dead groups were removed from the board;
693 * this is like all genmoves are in fact
694 * kgs-genmove_cleanup. */
695 u
->pass_all_alive
= !optval
|| atoi(optval
);
696 } else if (!strcasecmp(optname
, "allow_losing_pass")) {
697 /* Whether to consider passing in a clear
698 * but losing situation, to be scored as a loss
700 u
->allow_losing_pass
= !optval
|| atoi(optval
);
701 } else if (!strcasecmp(optname
, "territory_scoring")) {
702 /* Use territory scoring (default is area scoring).
703 * An explicit kgs-rules command overrides this. */
704 u
->territory_scoring
= !optval
|| atoi(optval
);
705 } else if (!strcasecmp(optname
, "stones_only")) {
706 /* Do not count eyes. Nice to teach go to kids.
707 * http://strasbourg.jeudego.org/regle_strasbourgeoise.htm */
708 b
->rules
= RULES_STONES_ONLY
;
709 u
->pass_all_alive
= true;
710 } else if (!strcasecmp(optname
, "banner") && optval
) {
711 /* Additional banner string. This must come as the
712 * last engine parameter. You can use '+' instead
713 * of ' ' if you are wrestling with kgsGtp. */
714 if (*next
) *--next
= ',';
715 u
->banner
= strdup(optval
);
716 for (char *b
= u
->banner
; *b
; b
++) {
717 if (*b
== '+') *b
= ' ';
720 } else if (!strcasecmp(optname
, "plugin") && optval
) {
721 /* Load an external plugin; filename goes before the colon,
722 * extra arguments after the colon. */
723 char *pluginarg
= strchr(optval
, ':');
726 plugin_load(u
->plugins
, optval
, pluginarg
);
728 /** UCT behavior and policies */
730 } else if ((!strcasecmp(optname
, "policy")
731 /* Node selection policy. ucb1amaf is the
732 * default policy implementing RAVE, while
733 * ucb1 is the simple exploration/exploitation
734 * policy. Policies can take further extra
736 || !strcasecmp(optname
, "random_policy")) && optval
) {
737 /* A policy to be used randomly with small
738 * chance instead of the default policy. */
739 char *policyarg
= strchr(optval
, ':');
740 struct uct_policy
**p
= !strcasecmp(optname
, "policy") ? &u
->policy
: &u
->random_policy
;
743 if (!strcasecmp(optval
, "ucb1")) {
744 *p
= policy_ucb1_init(u
, policyarg
);
745 } else if (!strcasecmp(optval
, "ucb1amaf")) {
746 *p
= policy_ucb1amaf_init(u
, policyarg
, b
);
748 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
751 } else if (!strcasecmp(optname
, "playout") && optval
) {
752 /* Random simulation (playout) policy.
753 * moggy is the default policy with large
754 * amount of domain-specific knowledge and
755 * heuristics. light is a simple uniformly
756 * random move selection policy. */
757 char *playoutarg
= strchr(optval
, ':');
760 if (!strcasecmp(optval
, "moggy")) {
761 u
->playout
= playout_moggy_init(playoutarg
, b
, u
->jdict
);
762 } else if (!strcasecmp(optval
, "light")) {
763 u
->playout
= playout_light_init(playoutarg
, b
);
765 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
768 } else if (!strcasecmp(optname
, "prior") && optval
) {
769 /* Node priors policy. When expanding a node,
770 * it will seed node values heuristically
771 * (most importantly, based on playout policy
772 * opinion, but also with regard to other
773 * things). See uct/prior.c for details.
774 * Use prior=eqex=0 to disable priors. */
775 u
->prior
= uct_prior_init(optval
, b
, u
);
776 } else if (!strcasecmp(optname
, "mercy") && optval
) {
777 /* Minimal difference of black/white captures
778 * to stop playout - "Mercy Rule". Speeds up
779 * hopeless playouts at the expense of some
781 u
->mercymin
= atoi(optval
);
782 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
783 /* Maximum length of single simulation
785 u
->gamelen
= atoi(optval
);
786 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
787 /* Expand UCT nodes after it has been
788 * visited this many times. */
789 u
->expand_p
= atoi(optval
);
790 } else if (!strcasecmp(optname
, "random_policy_chance") && optval
) {
791 /* If specified (N), with probability 1/N, random_policy policy
792 * descend is used instead of main policy descend; useful
793 * if specified policy (e.g. UCB1AMAF) can make unduly biased
794 * choices sometimes, you can fall back to e.g.
795 * random_policy=UCB1. */
796 u
->random_policy_chance
= atoi(optval
);
798 /** General AMAF behavior */
799 /* (Only relevant if the policy supports AMAF.
800 * More variables can be tuned as policy
803 } else if (!strcasecmp(optname
, "playout_amaf")) {
804 /* Whether to include random playout moves in
805 * AMAF as well. (Otherwise, only tree moves
806 * are included in AMAF. Of course makes sense
807 * only in connection with an AMAF policy.) */
808 /* with-without: 55.5% (+-4.1) */
809 if (optval
&& *optval
== '0')
810 u
->playout_amaf
= false;
812 u
->playout_amaf
= true;
813 } else if (!strcasecmp(optname
, "playout_amaf_cutoff") && optval
) {
814 /* Keep only first N% of playout stage AMAF
816 u
->playout_amaf_cutoff
= atoi(optval
);
817 } else if (!strcasecmp(optname
, "amaf_prior") && optval
) {
818 /* In node policy, consider prior values
819 * part of the real result term or part
820 * of the AMAF term? */
821 u
->amaf_prior
= atoi(optval
);
823 /** Performance and memory management */
825 } else if (!strcasecmp(optname
, "threads") && optval
) {
826 /* By default, Pachi will run with only single
827 * tree search thread! */
828 u
->threads
= atoi(optval
);
829 } else if (!strcasecmp(optname
, "thread_model") && optval
) {
830 if (!strcasecmp(optval
, "tree")) {
831 /* Tree parallelization - all threads
832 * grind on the same tree. */
833 u
->thread_model
= TM_TREE
;
835 } else if (!strcasecmp(optval
, "treevl")) {
836 /* Tree parallelization, but also
837 * with virtual losses - this discou-
838 * rages most threads choosing the
839 * same tree branches to read. */
840 u
->thread_model
= TM_TREEVL
;
842 fprintf(stderr
, "UCT: Invalid thread model %s\n", optval
);
845 } else if (!strcasecmp(optname
, "virtual_loss") && optval
) {
846 /* Number of virtual losses added before evaluating a node. */
847 u
->virtual_loss
= atoi(optval
);
848 } else if (!strcasecmp(optname
, "pondering")) {
849 /* Keep searching even during opponent's turn. */
850 u
->pondering_opt
= !optval
|| atoi(optval
);
851 } else if (!strcasecmp(optname
, "max_tree_size") && optval
) {
852 /* Maximum amount of memory [MiB] consumed by the move tree.
853 * For fast_alloc it includes the temp tree used for pruning.
854 * Default is 3072 (3 GiB). */
855 u
->max_tree_size
= atol(optval
) * 1048576;
856 } else if (!strcasecmp(optname
, "fast_alloc")) {
857 u
->fast_alloc
= !optval
|| atoi(optval
);
858 } else if (!strcasecmp(optname
, "pruning_threshold") && optval
) {
859 /* Force pruning at beginning of a move if the tree consumes
860 * more than this [MiB]. Default is 10% of max_tree_size.
861 * Increase to reduce pruning time overhead if memory is plentiful.
862 * This option is meaningful only for fast_alloc. */
863 u
->pruning_threshold
= atol(optval
) * 1048576;
867 } else if (!strcasecmp(optname
, "best2_ratio") && optval
) {
868 /* If set, prolong simulating while
869 * first_best/second_best playouts ratio
870 * is less than best2_ratio. */
871 u
->best2_ratio
= atof(optval
);
872 } else if (!strcasecmp(optname
, "bestr_ratio") && optval
) {
873 /* If set, prolong simulating while
874 * best,best_best_child values delta
875 * is more than bestr_ratio. */
876 u
->bestr_ratio
= atof(optval
);
877 } else if (!strcasecmp(optname
, "max_maintime_ratio") && optval
) {
878 /* If set and while not in byoyomi, prolong simulating no more than
879 * max_maintime_ratio times the normal desired thinking time. */
880 u
->max_maintime_ratio
= atof(optval
);
881 } else if (!strcasecmp(optname
, "fuseki_end") && optval
) {
882 /* At the very beginning it's not worth thinking
883 * too long because the playout evaluations are
884 * very noisy. So gradually increase the thinking
885 * time up to maximum when fuseki_end percent
886 * of the board has been played.
887 * This only applies if we are not in byoyomi. */
888 u
->fuseki_end
= atoi(optval
);
889 } else if (!strcasecmp(optname
, "yose_start") && optval
) {
890 /* When yose_start percent of the board has been
891 * played, or if we are in byoyomi, stop spending
892 * more time and spread the remaining time
894 * Between fuseki_end and yose_start, we spend
895 * a constant proportion of the remaining time
896 * on each move. (yose_start should actually
897 * be much earlier than when real yose start,
898 * but "yose" is a good short name to convey
900 u
->yose_start
= atoi(optval
);
904 } else if (!strcasecmp(optname
, "dynkomi") && optval
) {
905 /* Dynamic komi approach; there are multiple
906 * ways to adjust komi dynamically throughout
907 * play. We currently support two: */
908 char *dynkomiarg
= strchr(optval
, ':');
911 if (!strcasecmp(optval
, "none")) {
912 u
->dynkomi
= uct_dynkomi_init_none(u
, dynkomiarg
, b
);
913 } else if (!strcasecmp(optval
, "linear")) {
914 /* You should set dynkomi_mask=1 or a very low
915 * handicap_value for white. */
916 u
->dynkomi
= uct_dynkomi_init_linear(u
, dynkomiarg
, b
);
917 } else if (!strcasecmp(optval
, "adaptive")) {
918 /* There are many more knobs to
919 * crank - see uct/dynkomi.c. */
920 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, dynkomiarg
, b
);
922 fprintf(stderr
, "UCT: Invalid dynkomi mode %s\n", optval
);
925 } else if (!strcasecmp(optname
, "dynkomi_mask") && optval
) {
926 /* Bitmask of colors the player must be
927 * for dynkomi be applied; the default dynkomi_mask=3 allows
928 * dynkomi even in games where Pachi is white. */
929 u
->dynkomi_mask
= atoi(optval
);
930 } else if (!strcasecmp(optname
, "dynkomi_interval") && optval
) {
931 /* If non-zero, re-adjust dynamic komi
932 * throughout a single genmove reading,
933 * roughly every N simulations. */
934 /* XXX: Does not work with tree
935 * parallelization. */
936 u
->dynkomi_interval
= atoi(optval
);
937 } else if (!strcasecmp(optname
, "extra_komi") && optval
) {
938 /* Initial dynamic komi settings. This
939 * is useful for the adaptive dynkomi
940 * policy as the value to start with
941 * (this is NOT kept fixed) in case
942 * there is not enough time in the search
943 * to adjust the value properly (e.g. the
944 * game was interrupted). */
945 u
->initial_extra_komi
= atof(optval
);
947 /** Node value result scaling */
949 } else if (!strcasecmp(optname
, "val_scale") && optval
) {
950 /* How much of the game result value should be
951 * influenced by win size. Zero means it isn't. */
952 u
->val_scale
= atof(optval
);
953 } else if (!strcasecmp(optname
, "val_points") && optval
) {
954 /* Maximum size of win to be scaled into game
955 * result value. Zero means boardsize^2. */
956 u
->val_points
= atoi(optval
) * 2; // result values are doubled
957 } else if (!strcasecmp(optname
, "val_extra")) {
958 /* If false, the score coefficient will be simply
959 * added to the value, instead of scaling the result
960 * coefficient because of it. */
961 u
->val_extra
= !optval
|| atoi(optval
);
962 } else if (!strcasecmp(optname
, "val_byavg")) {
963 /* If true, the score included in the value will
964 * be relative to average score in the current
965 * search episode inst. of jigo. */
966 u
->val_byavg
= !optval
|| atoi(optval
);
967 } else if (!strcasecmp(optname
, "val_bytemp")) {
968 /* If true, the value scaling coefficient
969 * is different based on value extremity
970 * (dist. from 0.5), linear between
971 * val_bytemp_min, val_scale. */
972 u
->val_bytemp
= !optval
|| atoi(optval
);
973 } else if (!strcasecmp(optname
, "val_bytemp_min") && optval
) {
974 /* Minimum val_scale in case of val_bytemp. */
975 u
->val_bytemp_min
= atof(optval
);
978 /* (Purely experimental. Does not work - yet!) */
980 } else if (!strcasecmp(optname
, "local_tree")) {
981 /* Whether to bias exploration by local tree values. */
982 u
->local_tree
= !optval
|| atoi(optval
);
983 } else if (!strcasecmp(optname
, "tenuki_d") && optval
) {
984 /* Tenuki distance at which to break the local tree. */
985 u
->tenuki_d
= atoi(optval
);
986 if (u
->tenuki_d
> TREE_NODE_D_MAX
+ 1) {
987 fprintf(stderr
, "uct: tenuki_d must not be larger than TREE_NODE_D_MAX+1 %d\n", TREE_NODE_D_MAX
+ 1);
990 } else if (!strcasecmp(optname
, "local_tree_aging") && optval
) {
991 /* How much to reduce local tree values between moves. */
992 u
->local_tree_aging
= atof(optval
);
993 } else if (!strcasecmp(optname
, "local_tree_depth_decay") && optval
) {
994 /* With value x>0, during the descent the node
995 * contributes 1/x^depth playouts in
996 * the local tree. I.e., with x>1, nodes more
997 * distant from local situation contribute more
998 * than nodes near the root. */
999 u
->local_tree_depth_decay
= atof(optval
);
1000 } else if (!strcasecmp(optname
, "local_tree_allseq")) {
1001 /* If disabled, only complete sequences are stored
1002 * in the local tree. If this is on, also
1003 * subsequences starting at each move are stored. */
1004 u
->local_tree_allseq
= !optval
|| atoi(optval
);
1005 } else if (!strcasecmp(optname
, "local_tree_neival")) {
1006 /* If disabled, local node value is not
1007 * computed just based on terminal status
1008 * of the coordinate, but also its neighbors. */
1009 u
->local_tree_neival
= !optval
|| atoi(optval
);
1010 } else if (!strcasecmp(optname
, "local_tree_eval")) {
1011 /* How is the value inserted in the local tree
1013 if (!strcasecmp(optval
, "root"))
1014 /* All moves within a tree branch are
1015 * considered wrt. their merit
1016 * reaching tachtical goal of making
1017 * the first move in the branch
1019 u
->local_tree_eval
= LTE_ROOT
;
1020 else if (!strcasecmp(optval
, "each"))
1021 /* Each move is considered wrt.
1022 * its own survival. */
1023 u
->local_tree_eval
= LTE_EACH
;
1024 else if (!strcasecmp(optval
, "total"))
1025 /* The tactical goal is the survival
1026 * of all the moves of my color and
1027 * non-survival of all the opponent
1028 * moves. Local values (and their
1029 * inverses) are averaged. */
1030 u
->local_tree_eval
= LTE_TOTAL
;
1032 fprintf(stderr
, "uct: unknown local_tree_eval %s\n", optval
);
1035 } else if (!strcasecmp(optname
, "local_tree_rootchoose")) {
1036 /* If disabled, only moves within the local
1037 * tree branch are considered; the values
1038 * of the branch roots (i.e. root children)
1039 * are ignored. This may make sense together
1040 * with eval!=each, we consider only moves
1041 * that influence the goal, not the "rating"
1042 * of the goal itself. (The real solution
1043 * will be probably using criticality to pick
1044 * local tree branches.) */
1045 u
->local_tree_rootchoose
= !optval
|| atoi(optval
);
1047 /** Other heuristics */
1048 } else if (!strcasecmp(optname
, "patterns")) {
1049 /* Load pattern database. Various modules
1050 * (priors, policies etc.) may make use
1051 * of this database. They will request
1052 * it automatically in that case, but you
1053 * can use this option to tweak the pattern
1055 patterns_init(&u
->pat
, optval
, false, true);
1056 u
->want_pat
= pat_setup
= true;
1057 } else if (!strcasecmp(optname
, "significant_threshold") && optval
) {
1058 /* Some heuristics (XXX: none in mainline) rely
1059 * on the knowledge of the last "significant"
1060 * node in the descent. Such a node is
1061 * considered reasonably trustworthy to carry
1062 * some meaningful information in the values
1063 * of the node and its children. */
1064 u
->significant_threshold
= atoi(optval
);
1065 } else if (!strcasecmp(optname
, "libmap")) {
1066 /* Online learning of move tactical ratings by
1068 libmap_setup(optval
);
1071 /** Distributed engine slaves setup */
1073 } else if (!strcasecmp(optname
, "slave")) {
1074 /* Act as slave for the distributed engine. */
1075 u
->slave
= !optval
|| atoi(optval
);
1076 } else if (!strcasecmp(optname
, "slave_index") && optval
) {
1077 /* Optional index if per-slave behavior is desired.
1078 * Must be given as index/max */
1079 u
->slave_index
= atoi(optval
);
1080 char *p
= strchr(optval
, '/');
1081 if (p
) u
->max_slaves
= atoi(++p
);
1082 } else if (!strcasecmp(optname
, "shared_nodes") && optval
) {
1083 /* Share at most shared_nodes between master and slave at each genmoves.
1084 * Must use the same value in master and slaves. */
1085 u
->shared_nodes
= atoi(optval
);
1086 } else if (!strcasecmp(optname
, "shared_levels") && optval
) {
1087 /* Share only nodes of level <= shared_levels. */
1088 u
->shared_levels
= atoi(optval
);
1089 } else if (!strcasecmp(optname
, "stats_hbits") && optval
) {
1090 /* Set hash table size to 2^stats_hbits for the shared stats. */
1091 u
->stats_hbits
= atoi(optval
);
1092 } else if (!strcasecmp(optname
, "stats_delay") && optval
) {
1093 /* How long to wait in slave for initial stats to build up before
1094 * replying to the genmoves command (in ms) */
1095 u
->stats_delay
= 0.001 * atof(optval
);
1099 } else if (!strcasecmp(optname
, "maximize_score")) {
1100 /* A combination of settings that will make
1101 * Pachi try to maximize his points (instead
1102 * of playing slack yose) or minimize his loss
1103 * (and proceed to counting even when losing). */
1104 /* Please note that this preset might be
1105 * somewhat weaker than normal Pachi, and the
1106 * score maximization is approximate; point size
1107 * of win/loss still should not be used to judge
1108 * strength of Pachi or the opponent. */
1109 /* See README for some further notes. */
1110 if (!optval
|| atoi(optval
)) {
1111 /* Allow scoring a lost game. */
1112 u
->allow_losing_pass
= true;
1113 /* Make Pachi keep his calm when losing
1114 * and/or maintain winning marging. */
1115 /* Do not play games that are losing
1117 /* XXX: komi_ratchet_age=40000 is necessary
1118 * with losing_komi_ratchet, but 40000
1119 * is somewhat arbitrary value. */
1120 char dynkomi_args
[] = "losing_komi_ratchet:komi_ratchet_age=60000:no_komi_at_game_end=0:max_losing_komi=30";
1121 u
->dynkomi
= uct_dynkomi_init_adaptive(u
, dynkomi_args
, b
);
1122 /* XXX: Values arbitrary so far. */
1123 /* XXX: Also, is bytemp sensible when
1124 * combined with dynamic komi?! */
1125 u
->val_scale
= 0.01;
1126 u
->val_bytemp
= true;
1127 u
->val_bytemp_min
= 0.001;
1128 u
->val_byavg
= true;
1132 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
1139 u
->policy
= policy_ucb1amaf_init(u
, NULL
, b
);
1141 if (!!u
->random_policy_chance
^ !!u
->random_policy
) {
1142 fprintf(stderr
, "uct: Only one of random_policy and random_policy_chance is set\n");
1146 if (!u
->local_tree
) {
1147 /* No ltree aging. */
1148 u
->local_tree_aging
= 1.0f
;
1151 if (u
->fast_alloc
) {
1152 if (u
->pruning_threshold
< u
->max_tree_size
/ 10)
1153 u
->pruning_threshold
= u
->max_tree_size
/ 10;
1154 if (u
->pruning_threshold
> u
->max_tree_size
/ 2)
1155 u
->pruning_threshold
= u
->max_tree_size
/ 2;
1157 /* Limit pruning temp space to 20% of memory. Beyond this we discard
1158 * the nodes and recompute them at the next move if necessary. */
1159 u
->max_pruned_size
= u
->max_tree_size
/ 5;
1160 u
->max_tree_size
-= u
->max_pruned_size
;
1162 /* Reserve 5% memory in case the background free() are slower
1163 * than the concurrent allocations. */
1164 u
->max_tree_size
-= u
->max_tree_size
/ 20;
1168 u
->prior
= uct_prior_init(NULL
, b
, u
);
1171 u
->playout
= playout_moggy_init(NULL
, b
, u
->jdict
);
1172 if (!u
->playout
->debug_level
)
1173 u
->playout
->debug_level
= u
->debug_level
;
1175 if (u
->want_pat
&& !pat_setup
)
1176 patterns_init(&u
->pat
, NULL
, false, true);
1178 u
->ownermap
.map
= malloc2(board_size2(b
) * sizeof(u
->ownermap
.map
[0]));
1181 if (!u
->stats_hbits
) u
->stats_hbits
= DEFAULT_STATS_HBITS
;
1182 if (!u
->shared_nodes
) u
->shared_nodes
= DEFAULT_SHARED_NODES
;
1183 assert(u
->shared_levels
* board_bits2(b
) <= 8 * (int)sizeof(path_t
));
1187 u
->dynkomi
= board_small(b
) ? uct_dynkomi_init_none(u
, NULL
, b
)
1188 : uct_dynkomi_init_linear(u
, NULL
, b
);
1190 /* Some things remain uninitialized for now - the opening tbook
1191 * is not loaded and the tree not set up. */
1192 /* This will be initialized in setup_state() at the first move
1193 * received/requested. This is because right now we are not aware
1194 * about any komi or handicap setup and such. */
1200 engine_uct_init(char *arg
, struct board
*b
)
1202 struct uct
*u
= uct_state_init(arg
, b
);
1203 struct engine
*e
= calloc2(1, sizeof(struct engine
));
1205 e
->printhook
= uct_printhook_ownermap
;
1206 e
->notify_play
= uct_notify_play
;
1209 e
->result
= uct_result
;
1210 e
->genmove
= uct_genmove
;
1211 e
->genmoves
= uct_genmoves
;
1212 e
->evaluate
= uct_evaluate
;
1213 e
->dead_group_list
= uct_dead_group_list
;
1218 e
->notify
= uct_notify
;
1220 const char banner
[] = "If you believe you have won but I am still playing, "
1221 "please help me understand by capturing all dead stones. "
1222 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
1223 if (!u
->banner
) u
->banner
= "";
1224 e
->comment
= malloc2(sizeof(banner
) + strlen(u
->banner
) + 1);
1225 sprintf(e
->comment
, "%s %s", banner
, u
->banner
);