13 #include "distributed/distributed.h"
17 #include "uct/dynkomi.h"
18 #include "uct/internal.h"
19 #include "uct/search.h"
25 /* Default number of simulations to perform per move.
26 * Note that this is now in total over all threads!. */
27 #define MC_GAMES 80000
28 static const struct time_info default_ti
= {
31 .len
= { .games
= MC_GAMES
},
34 /* Once per how many simulations (per thread) to show a progress report line. */
35 #define TREE_SIMPROGRESS_INTERVAL 2500
37 /* When terminating UCT search early, the safety margin to add to the
38 * remaining playout number estimate when deciding whether the result can
40 #define PLAYOUT_DELTA_SAFEMARGIN 1000
42 /* Minimal number of simulations to consider early break. */
43 #define PLAYOUT_EARLY_BREAK_MIN 5000
46 /* Pachi threading structure:
49 * | main(), GTP communication, ...
50 * | starts and stops the search managed by thread_manager
53 * | spawns and collects worker threads
59 * uct_playouts() loop, doing descend-playout until uct_halt
61 * Another way to look at it is by functions (lines denote thread boundaries):
64 * | uct_search() (uct_search_start() .. uct_search_stop())
65 * | -----------------------
66 * | spawn_thread_manager()
67 * | -----------------------
71 /* Set in thread manager in case the workers should stop. */
72 volatile sig_atomic_t uct_halt
= 0;
73 /* ID of the thread manager. */
74 static pthread_t thread_manager
;
75 bool thread_manager_running
;
77 static pthread_mutex_t finish_mutex
= PTHREAD_MUTEX_INITIALIZER
;
78 static pthread_cond_t finish_cond
= PTHREAD_COND_INITIALIZER
;
79 static volatile int finish_thread
;
80 static pthread_mutex_t finish_serializer
= PTHREAD_MUTEX_INITIALIZER
;
83 spawn_worker(void *ctx_
)
85 struct uct_thread_ctx
*ctx
= ctx_
;
87 fast_srandom(ctx
->seed
);
89 ctx
->games
= uct_playouts(ctx
->u
, ctx
->b
, ctx
->color
, ctx
->t
, ctx
->ti
);
91 pthread_mutex_lock(&finish_serializer
);
92 pthread_mutex_lock(&finish_mutex
);
93 finish_thread
= ctx
->tid
;
94 pthread_cond_signal(&finish_cond
);
95 pthread_mutex_unlock(&finish_mutex
);
99 /* Thread manager, controlling worker threads. It must be called with
100 * finish_mutex lock held, but it will unlock it itself before exiting;
101 * this is necessary to be completely deadlock-free. */
102 /* The finish_cond can be signalled for it to stop; in that case,
103 * the caller should set finish_thread = -1. */
104 /* After it is started, it will update mctx->t to point at some tree
105 * used for the actual search, on return
106 * it will set mctx->games to the number of performed simulations. */
108 spawn_thread_manager(void *ctx_
)
110 /* In thread_manager, we use only some of the ctx fields. */
111 struct uct_thread_ctx
*mctx
= ctx_
;
112 struct uct
*u
= mctx
->u
;
113 struct tree
*t
= mctx
->t
;
114 fast_srandom(mctx
->seed
);
116 int played_games
= 0;
117 pthread_t threads
[u
->threads
];
122 /* Garbage collect the tree by preference when pondering. */
123 if (u
->pondering
&& t
->nodes
&& t
->nodes_size
>= t
->pruning_threshold
) {
124 t
->root
= tree_garbage_collect(t
, t
->root
);
127 /* Spawn threads... */
128 for (int ti
= 0; ti
< u
->threads
; ti
++) {
129 struct uct_thread_ctx
*ctx
= malloc2(sizeof(*ctx
));
130 ctx
->u
= u
; ctx
->b
= mctx
->b
; ctx
->color
= mctx
->color
;
131 mctx
->t
= ctx
->t
= t
;
132 ctx
->tid
= ti
; ctx
->seed
= fast_random(65536) + ti
;
134 pthread_create(&threads
[ti
], NULL
, spawn_worker
, ctx
);
136 fprintf(stderr
, "Spawned worker %d\n", ti
);
139 /* ...and collect them back: */
140 while (joined
< u
->threads
) {
141 /* Wait for some thread to finish... */
142 pthread_cond_wait(&finish_cond
, &finish_mutex
);
143 if (finish_thread
< 0) {
144 /* Stop-by-caller. Tell the workers to wrap up
145 * and unblock them from terminating. */
147 /* We need to make sure the workers do not complete
148 * the termination sequence before we get officially
149 * stopped - their wake and the stop wake could get
151 pthread_mutex_unlock(&finish_serializer
);
154 /* ...and gather its remnants. */
155 struct uct_thread_ctx
*ctx
;
156 pthread_join(threads
[finish_thread
], (void **) &ctx
);
157 played_games
+= ctx
->games
;
161 fprintf(stderr
, "Joined worker %d\n", finish_thread
);
162 pthread_mutex_unlock(&finish_serializer
);
165 pthread_mutex_unlock(&finish_mutex
);
167 mctx
->games
= played_games
;
172 /*** THREAD MANAGER end */
174 /*** Search infrastructure: */
178 uct_search_games(struct uct_search_state
*s
)
180 return s
->ctx
->t
->root
->u
.playouts
;
184 uct_search_start(struct uct
*u
, struct board
*b
, enum stone color
,
185 struct tree
*t
, struct time_info
*ti
,
186 struct uct_search_state
*s
)
188 /* Set up search state. */
189 s
->base_playouts
= s
->last_dynkomi
= s
->last_print
= t
->root
->u
.playouts
;
190 s
->print_interval
= TREE_SIMPROGRESS_INTERVAL
* u
->threads
;
194 if (ti
->period
== TT_NULL
) *ti
= default_ti
;
195 time_stop_conditions(ti
, b
, u
->fuseki_end
, u
->yose_start
, u
->max_maintime_ratio
, &s
->stop
);
198 /* Fire up the tree search thread manager, which will in turn
199 * spawn the searching threads. */
200 assert(u
->threads
> 0);
201 assert(!thread_manager_running
);
202 static struct uct_thread_ctx mctx
;
203 mctx
= (struct uct_thread_ctx
) { .u
= u
, .b
= b
, .color
= color
, .t
= t
, .seed
= fast_random(65536), .ti
= ti
};
205 pthread_mutex_lock(&finish_serializer
);
206 pthread_mutex_lock(&finish_mutex
);
207 pthread_create(&thread_manager
, NULL
, spawn_thread_manager
, s
->ctx
);
208 thread_manager_running
= true;
211 struct uct_thread_ctx
*
212 uct_search_stop(void)
214 assert(thread_manager_running
);
216 /* Signal thread manager to stop the workers. */
217 pthread_mutex_lock(&finish_mutex
);
219 pthread_cond_signal(&finish_cond
);
220 pthread_mutex_unlock(&finish_mutex
);
222 /* Collect the thread manager. */
223 struct uct_thread_ctx
*pctx
;
224 thread_manager_running
= false;
225 pthread_join(thread_manager
, (void **) &pctx
);
231 uct_search_progress(struct uct
*u
, struct board
*b
, enum stone color
,
232 struct tree
*t
, struct time_info
*ti
,
233 struct uct_search_state
*s
, int i
)
235 struct uct_thread_ctx
*ctx
= s
->ctx
;
237 /* Adjust dynkomi? */
238 int di
= u
->dynkomi_interval
* u
->threads
;
239 if (ctx
->t
->use_extra_komi
&& u
->dynkomi
->permove
240 && !u
->pondering
&& di
241 && i
> s
->last_dynkomi
+ di
) {
242 s
->last_dynkomi
+= di
;
243 floating_t old_dynkomi
= ctx
->t
->extra_komi
;
244 ctx
->t
->extra_komi
= u
->dynkomi
->permove(u
->dynkomi
, b
, ctx
->t
);
245 if (UDEBUGL(3) && old_dynkomi
!= ctx
->t
->extra_komi
)
246 fprintf(stderr
, "dynkomi adjusted (%f -> %f)\n",
247 old_dynkomi
, ctx
->t
->extra_komi
);
250 /* Print progress? */
251 if (i
- s
->last_print
> s
->print_interval
) {
252 s
->last_print
+= s
->print_interval
; // keep the numbers tidy
253 uct_progress_status(u
, ctx
->t
, color
, s
->last_print
);
256 if (!s
->fullmem
&& ctx
->t
->nodes_size
> u
->max_tree_size
) {
258 fprintf(stderr
, "memory limit hit (%lu > %lu)\n",
259 ctx
->t
->nodes_size
, u
->max_tree_size
);
265 /* Determine whether we should terminate the search early. */
267 uct_search_stop_early(struct uct
*u
, struct tree
*t
, struct board
*b
,
268 struct time_info
*ti
, struct time_stop
*stop
,
269 struct tree_node
*best
, struct tree_node
*best2
,
270 int played
, bool fullmem
)
272 /* If the memory is full, stop immediately. Since the tree
273 * cannot grow anymore, some non-well-expanded nodes will
274 * quickly take over with extremely high ratio since the
275 * counters are not properly simulated (just as if we use
276 * non-UCT MonteCarlo). */
277 /* (XXX: A proper solution would be to prune the tree
282 /* Think at least 100ms to avoid a random move. This is particularly
283 * important in distributed mode, where this function is called frequently. */
284 double elapsed
= 0.0;
285 if (ti
->dim
== TD_WALLTIME
) {
286 elapsed
= time_now() - ti
->len
.t
.timer_start
;
287 if (elapsed
< TREE_BUSYWAIT_INTERVAL
) return false;
290 /* Break early if we estimate the second-best move cannot
291 * catch up in assigned time anymore. We use all our time
292 * if we are in byoyomi with single stone remaining in our
293 * period, however - it's better to pre-ponder. */
294 bool time_indulgent
= (!ti
->len
.t
.main_time
&& ti
->len
.t
.byoyomi_stones
== 1);
295 if (best2
&& ti
->dim
== TD_WALLTIME
296 && played
>= PLAYOUT_EARLY_BREAK_MIN
&& !time_indulgent
) {
297 double remaining
= stop
->worst
.time
- elapsed
;
298 double pps
= ((double)played
) / elapsed
;
299 double estplayouts
= remaining
* pps
+ PLAYOUT_DELTA_SAFEMARGIN
;
300 if (best
->u
.playouts
> best2
->u
.playouts
+ estplayouts
) {
302 fprintf(stderr
, "Early stop, result cannot change: "
303 "best %d, best2 %d, estimated %f simulations to go (%d/%f=%f pps)\n",
304 best
->u
.playouts
, best2
->u
.playouts
, estplayouts
, played
, elapsed
, pps
);
310 /* Early break in won situation. */
311 /* NOTE: disabled in order to allow dynkomi to catch up */
312 if (ti
->dim
!= TD_GAMES
&& best
->u
.playouts
>= PLAYOUT_EARLY_BREAK_MIN
313 && tree_node_get_value(t
, 1, best
->u
.value
) >= u
->sure_win_threshold
) {
321 /* Determine whether we should terminate the search later than expected. */
323 uct_search_keep_looking(struct uct
*u
, struct tree
*t
, struct board
*b
,
324 struct time_info
*ti
, struct time_stop
*stop
,
325 struct tree_node
*best
, struct tree_node
*best2
,
326 struct tree_node
*bestr
, struct tree_node
*winner
, int i
)
330 fprintf(stderr
, "Did not find best move, still trying...\n");
334 /* Do not waste time if we are winning. Spend up to worst time if
335 * we are unsure, but only desired time if we are sure of winning. */
336 floating_t beta
= 2 * (tree_node_get_value(t
, 1, best
->u
.value
) - 0.5);
337 if (ti
->dim
== TD_WALLTIME
&& beta
> 0) {
338 double good_enough
= stop
->desired
.time
* beta
+ stop
->worst
.time
* (1 - beta
);
339 double elapsed
= time_now() - ti
->len
.t
.timer_start
;
340 if (elapsed
> good_enough
) return false;
343 if (u
->best2_ratio
> 0) {
344 /* Check best/best2 simulations ratio. If the
345 * two best moves give very similar results,
346 * keep simulating. */
347 if (best2
&& best2
->u
.playouts
348 && (double)best
->u
.playouts
/ best2
->u
.playouts
< u
->best2_ratio
) {
350 fprintf(stderr
, "Best2 ratio %f < threshold %f\n",
351 (double)best
->u
.playouts
/ best2
->u
.playouts
,
357 if (u
->bestr_ratio
> 0) {
358 /* Check best, best_best value difference. If the best move
359 * and its best child do not give similar enough results,
360 * keep simulating. */
361 if (bestr
&& bestr
->u
.playouts
362 && fabs((double)best
->u
.value
- bestr
->u
.value
) > u
->bestr_ratio
) {
364 fprintf(stderr
, "Bestr delta %f > threshold %f\n",
365 fabs((double)best
->u
.value
- bestr
->u
.value
),
371 if (winner
&& winner
!= best
) {
372 /* Keep simulating if best explored
373 * does not have also highest value. */
375 fprintf(stderr
, "[%d] best %3s [%d] %f != winner %3s [%d] %f\n", i
,
376 coord2sstr(best
->coord
, t
->board
),
377 best
->u
.playouts
, tree_node_get_value(t
, 1, best
->u
.value
),
378 coord2sstr(winner
->coord
, t
->board
),
379 winner
->u
.playouts
, tree_node_get_value(t
, 1, winner
->u
.value
));
383 /* No reason to keep simulating, bye. */
388 uct_search_check_stop(struct uct
*u
, struct board
*b
, enum stone color
,
389 struct tree
*t
, struct time_info
*ti
,
390 struct uct_search_state
*s
, int i
)
392 struct uct_thread_ctx
*ctx
= s
->ctx
;
394 /* Never consider stopping if we played too few simulations.
395 * Maybe we risk losing on time when playing in super-extreme
396 * time pressure but the tree is going to be just too messed
397 * up otherwise - we might even play invalid suicides or pass
398 * when we mustn't. */
399 assert(!(ti
->dim
== TD_GAMES
&& ti
->len
.games
< GJ_MINGAMES
));
403 struct tree_node
*best
= NULL
;
404 struct tree_node
*best2
= NULL
; // Second-best move.
405 struct tree_node
*bestr
= NULL
; // best's best child.
406 struct tree_node
*winner
= NULL
;
408 best
= u
->policy
->choose(u
->policy
, ctx
->t
->root
, b
, color
, resign
);
409 if (best
) best2
= u
->policy
->choose(u
->policy
, ctx
->t
->root
, b
, color
, best
->coord
);
411 /* Possibly stop search early if it's no use to try on. */
412 int played
= u
->played_all
+ i
- s
->base_playouts
;
413 if (best
&& uct_search_stop_early(u
, ctx
->t
, b
, ti
, &s
->stop
, best
, best2
, played
, s
->fullmem
))
416 /* Check against time settings. */
418 if (ti
->dim
== TD_WALLTIME
) {
419 double elapsed
= time_now() - ti
->len
.t
.timer_start
;
420 if (elapsed
> s
->stop
.worst
.time
) return true;
421 desired_done
= elapsed
> s
->stop
.desired
.time
;
424 assert(ti
->dim
== TD_GAMES
);
425 if (i
> s
->stop
.worst
.playouts
) return true;
426 desired_done
= i
> s
->stop
.desired
.playouts
;
429 /* We want to stop simulating, but are willing to keep trying
430 * if we aren't completely sure about the winner yet. */
432 if (u
->policy
->winner
&& u
->policy
->evaluate
) {
433 struct uct_descent descent
= { .node
= ctx
->t
->root
};
434 u
->policy
->winner(u
->policy
, ctx
->t
, &descent
);
435 winner
= descent
.node
;
438 bestr
= u
->policy
->choose(u
->policy
, best
, b
, stone_other(color
), resign
);
439 if (!uct_search_keep_looking(u
, ctx
->t
, b
, ti
, &s
->stop
, best
, best2
, bestr
, winner
, i
))
443 /* TODO: Early break if best->variance goes under threshold
444 * and we already have enough playouts (possibly thanks to tbook
445 * or to pondering)? */
451 uct_search_result(struct uct
*u
, struct board
*b
, enum stone color
,
452 bool pass_all_alive
, int played_games
, int base_playouts
,
455 /* Choose the best move from the tree. */
456 struct tree_node
*best
= u
->policy
->choose(u
->policy
, u
->t
->root
, b
, color
, resign
);
461 *best_coord
= best
->coord
;
463 fprintf(stderr
, "*** WINNER is %s (%d,%d) with score %1.4f (%d/%d:%d/%d games), extra komi %f\n",
464 coord2sstr(best
->coord
, b
), coord_x(best
->coord
, b
), coord_y(best
->coord
, b
),
465 tree_node_get_value(u
->t
, 1, best
->u
.value
), best
->u
.playouts
,
466 u
->t
->root
->u
.playouts
, u
->t
->root
->u
.playouts
- base_playouts
, played_games
,
469 /* Do not resign if we're so short of time that evaluation of best
470 * move is completely unreliable, we might be winning actually.
471 * In this case best is almost random but still better than resign.
472 * Also do not resign if we are getting bad results while actually
473 * giving away extra komi points (dynkomi). */
474 /* NOTE: When dynkomi worsens quickly from a winning position, the root value can be rather
475 * biased. We should actually look at the most losing komi used during this move. Since
476 * komi is unlimited, we simply disable resigning for now. */
478 if (tree_node_get_value(u
->t
, 1, best
->u
.value
) < u
->resign_threshold
479 && !is_pass(best
->coord
) && best
->u
.playouts
> GJ_MINGAMES
480 && (!u
->t
->use_extra_komi
|| komi_by_color(u
->t
->extra_komi
, color
) < 0.5)) {
481 *best_coord
= resign
;
486 /* If the opponent just passed and we win counting, always
488 if (b
->moves
> 1 && is_pass(b
->last_move
.coord
)) {
489 /* Make sure enough playouts are simulated. */
490 while (u
->ownermap
.playouts
< GJ_MINGAMES
)
491 uct_playout(u
, b
, color
, u
->t
);
492 if (uct_pass_is_safe(u
, b
, color
, u
->pass_all_alive
|| pass_all_alive
)) {
494 fprintf(stderr
, "<Will rather pass, looks safe enough; score %f>\n",
495 board_official_score(b
, NULL
) / 2);
497 best
= u
->t
->root
->children
; // pass is the first child
498 assert(is_pass(best
->coord
));