17 #include "uct/dynkomi.h"
18 #include "uct/internal.h"
19 #include "uct/search.h"
24 #define DESCENT_DLEN 512
27 uct_progress_status(struct uct
*u
, struct tree
*t
, enum stone color
, int playouts
)
33 struct tree_node
*best
= u
->policy
->choose(u
->policy
, t
->root
, t
->board
, color
, resign
);
35 fprintf(stderr
, "... No moves left\n");
38 fprintf(stderr
, "[%d] size=%lu/%lu ", playouts
, t
->nodes_size
, t
->max_tree_size
);
39 fprintf(stderr
, "best %f ", tree_node_get_value(t
, 1, best
->u
.value
));
42 if (t
->use_extra_komi
)
43 fprintf(stderr
, "komi %.1f ", t
->extra_komi
);
46 fprintf(stderr
, "| seq ");
47 for (int depth
= 0; depth
< 4; depth
++) {
48 if (best
&& best
->u
.playouts
>= 25) {
49 fprintf(stderr
, "%3s ", coord2sstr(best
->coord
, t
->board
));
50 best
= u
->policy
->choose(u
->policy
, best
, t
->board
, color
, resign
);
57 fprintf(stderr
, "| can ");
59 struct tree_node
*can
[cans
];
60 memset(can
, 0, sizeof(can
));
61 best
= t
->root
->children
;
64 while ((!can
[c
] || best
->u
.playouts
> can
[c
]->u
.playouts
) && ++c
< cans
);
65 for (int d
= 0; d
< c
; d
++) can
[d
] = can
[d
+ 1];
66 if (c
> 0) can
[c
- 1] = best
;
71 fprintf(stderr
, "%3s(%.3f) ",
72 coord2sstr(can
[cans
]->coord
, t
->board
),
73 tree_node_get_value(t
, 1, can
[cans
]->u
.value
));
79 fprintf(stderr
, "\n");
84 record_amaf_move(struct playout_amafmap
*amaf
, coord_t coord
, enum stone color
)
86 if (amaf
->map
[coord
] == S_NONE
|| amaf
->map
[coord
] == color
) {
87 amaf
->map
[coord
] = color
;
88 } else { // XXX: Respect amaf->record_nakade
89 amaf_op(amaf
->map
[coord
], +);
91 amaf
->game
[amaf
->gamelen
].coord
= coord
;
92 amaf
->game
[amaf
->gamelen
].color
= color
;
94 assert(amaf
->gamelen
< sizeof(amaf
->game
) / sizeof(amaf
->game
[0]));
98 struct uct_playout_callback
{
101 struct tree_node
*lnode
;
106 uct_playout_hook(struct playout_policy
*playout
, struct playout_setup
*setup
, struct board
*b
, enum stone color
, int mode
)
108 /* XXX: This is used in some non-master branches. */
113 uct_playout_prepolicy(struct playout_policy
*playout
, struct playout_setup
*setup
, struct board
*b
, enum stone color
)
115 return uct_playout_hook(playout
, setup
, b
, color
, 0);
119 uct_playout_postpolicy(struct playout_policy
*playout
, struct playout_setup
*setup
, struct board
*b
, enum stone color
)
121 return uct_playout_hook(playout
, setup
, b
, color
, 1);
126 uct_leaf_node(struct uct
*u
, struct board
*b
, enum stone player_color
,
127 struct playout_amafmap
*amaf
,
128 struct uct_descent
*descent
, int *dlen
,
129 struct tree_node
*significant
[2],
130 struct tree
*t
, struct tree_node
*n
, enum stone node_color
,
133 enum stone next_color
= stone_other(node_color
);
134 int parity
= (next_color
== player_color
? 1 : -1);
136 /* Since filling all memory would cause the Monte-Carlo simulations to terminate, we try to
137 * avoid it by increasing the effective expand_p. */
138 unsigned long cur_size
= t
->nodes_size
; /* it is volatile and can change at any time, so we must cache its value */
139 if (cur_size
< u
->max_tree_size
&& (unsigned long) n
->u
.playouts
>= u
->max_tree_size
/ (u
->max_tree_size
- cur_size
) + 1) {
140 /* We need to make sure only one thread expands the node. If
141 * we are unlucky enough for two threads to meet in the same
142 * node, the latter one will simply do another simulation from
143 * the node itself, no big deal. t->nodes_size may exceed
144 * the maximum in multi-threaded case but not by much so it's ok.
145 * The size test must be before the test&set not after, to allow
146 * expansion of the node later if enough nodes have been freed. */
147 if (!__sync_lock_test_and_set(&n
->is_expanded
, 1)) {
148 tree_expand_node(t
, n
, b
, next_color
, u
, parity
);
152 fprintf(stderr
, "%s*-- UCT playout #%d start [%s] %f\n",
153 spaces
, n
->u
.playouts
, coord2sstr(n
->coord
, t
->board
),
154 tree_node_get_value(t
, parity
, n
->u
.value
));
156 struct uct_playout_callback upc
= {
159 /* TODO: Don't necessarily restart the sequence walk when
160 * entering playout. */
164 struct playout_setup ps
= {
165 .gamelen
= u
->gamelen
,
166 .mercymin
= u
->mercymin
,
167 .prepolicy_hook
= uct_playout_prepolicy
,
168 .postpolicy_hook
= uct_playout_postpolicy
,
171 int result
= play_random_game(&ps
, b
, next_color
,
172 u
->playout_amaf
? amaf
: NULL
,
173 &u
->ownermap
, u
->playout
);
174 if (next_color
== S_WHITE
) {
175 /* We need the result from black's perspective. */
179 fprintf(stderr
, "%s -- [%d..%d] %s random playout result %d\n",
180 spaces
, player_color
, next_color
, coord2sstr(n
->coord
, t
->board
), result
);
186 scale_value(struct uct
*u
, struct board
*b
, int result
)
188 floating_t rval
= result
> 0 ? 1.0 : result
< 0 ? 0.0 : 0.5;
189 if (u
->val_scale
&& result
!= 0) {
190 int vp
= u
->val_points
;
192 vp
= board_size(b
) - 1; vp
*= vp
; vp
*= 2;
195 floating_t sval
= (floating_t
) abs(result
) / vp
;
196 sval
= sval
> 1 ? 1 : sval
;
197 if (result
< 0) sval
= 1 - sval
;
199 rval
+= u
->val_scale
* sval
;
201 rval
= (1 - u
->val_scale
) * rval
+ u
->val_scale
* sval
;
202 // fprintf(stderr, "score %d => sval %f, rval %f\n", result, sval, rval);
208 record_local_sequence(struct uct
*u
, struct tree
*t
,
209 struct uct_descent
*descent
, int dlen
, int di
,
210 enum stone seq_color
, floating_t rval
)
212 #define LTREE_DEBUG if (UDEBUGL(6))
214 /* Ignore pass sequences. */
215 if (is_pass(descent
[di
].node
->coord
))
218 /* Transform the rval appropriately, based on the expected
219 * result at the root of the sequence. */
220 if (u
->local_tree_rootseqval
) {
221 float expval
= descent
[di
- 1].value
.value
;
222 rval
= stats_temper_value(rval
, expval
, u
->local_tree
);
225 LTREE_DEBUG
fprintf(stderr
, "recording result %f in local %s sequence: ",
226 rval
, stone2str(seq_color
));
228 /* Sequences starting deeper are less relevant in general. */
229 int pval
= LTREE_PLAYOUTS_MULTIPLIER
;
230 if (u
->local_tree
&& u
->local_tree_depth_decay
> 0)
231 pval
= ((floating_t
) pval
) / pow(u
->local_tree_depth_decay
, di
- 1);
233 LTREE_DEBUG
fprintf(stderr
, "too deep @%d\n", di
);
237 /* Pick the right local tree root... */
238 struct tree_node
*lnode
= seq_color
== S_BLACK
? t
->ltree_black
: t
->ltree_white
;
241 /* ...and record the sequence. */
243 while (di
< dlen
&& (di
== di0
|| descent
[di
].node
->d
< u
->tenuki_d
)) {
244 LTREE_DEBUG
fprintf(stderr
, "%s[%d] ",
245 coord2sstr(descent
[di
].node
->coord
, t
->board
),
246 descent
[di
].node
->d
);
247 lnode
= tree_get_node(t
, lnode
, descent
[di
++].node
->coord
, true);
249 stats_add_result(&lnode
->u
, rval
, pval
);
252 /* Add lnode for tenuki (pass) if we descended further. */
254 LTREE_DEBUG
fprintf(stderr
, "pass ");
255 lnode
= tree_get_node(t
, lnode
, pass
, true);
257 stats_add_result(&lnode
->u
, rval
, pval
);
260 LTREE_DEBUG
fprintf(stderr
, "\n");
265 uct_playout(struct uct
*u
, struct board
*b
, enum stone player_color
, struct tree
*t
)
270 struct playout_amafmap
*amaf
= NULL
;
271 if (u
->policy
->wants_amaf
) {
272 amaf
= calloc2(1, sizeof(*amaf
));
273 amaf
->map
= calloc2(board_size2(&b2
) + 1, sizeof(*amaf
->map
));
274 amaf
->map
++; // -1 is pass
277 /* Walk the tree until we find a leaf, then expand it and do
278 * a random playout. */
279 struct tree_node
*n
= t
->root
;
280 enum stone node_color
= stone_other(player_color
);
281 assert(node_color
== t
->root_color
);
283 /* Tree descent history. */
284 /* XXX: This is somewhat messy since @n and descent[dlen-1].node are
286 struct uct_descent descent
[DESCENT_DLEN
];
287 descent
[0].node
= n
; descent
[0].lnode
= NULL
;
289 /* Total value of the sequence. */
290 struct move_stats seq_value
= { .playouts
= 0 };
291 /* The last "significant" node along the descent (i.e. node
292 * with higher than configured number of playouts). For black
294 struct tree_node
*significant
[2] = { NULL
, NULL
};
295 if (n
->u
.playouts
>= u
->significant_threshold
)
296 significant
[node_color
- 1] = n
;
299 int pass_limit
= (board_size(&b2
) - 2) * (board_size(&b2
) - 2) / 2;
300 int passes
= is_pass(b
->last_move
.coord
) && b
->moves
> 0;
303 static char spaces
[DESCENT_DLEN
];
307 fprintf(stderr
, "--- UCT walk with color %d\n", player_color
);
309 while (!tree_leaf_node(n
) && passes
< 2) {
310 /*** Choose a node to descend to: */
312 /* Parity is chosen already according to the child color, since
313 * it is applied to children. */
314 node_color
= stone_other(node_color
);
315 int parity
= (node_color
== player_color
? 1 : -1);
317 assert(dlen
< DESCENT_DLEN
);
318 spaces
[dlen
- 1] = ' '; spaces
[dlen
] = 0;
319 descent
[dlen
] = descent
[dlen
- 1];
320 if (u
->local_tree
&& (!descent
[dlen
].lnode
|| descent
[dlen
].node
->d
>= u
->tenuki_d
)) {
321 /* Start new local sequence. */
322 /* Remember that node_color already holds color of the
323 * to-be-found child. */
324 descent
[dlen
].lnode
= node_color
== S_BLACK
? t
->ltree_black
: t
->ltree_white
;
327 if (!u
->random_policy_chance
|| fast_random(u
->random_policy_chance
))
328 u
->policy
->descend(u
->policy
, t
, &descent
[dlen
], parity
, b2
.moves
> pass_limit
);
330 u
->random_policy
->descend(u
->random_policy
, t
, &descent
[dlen
], parity
, b2
.moves
> pass_limit
);
333 /*** Perform the descent: */
335 if (descent
[dlen
].node
->u
.playouts
>= u
->significant_threshold
) {
336 significant
[node_color
- 1] = descent
[dlen
].node
;
339 seq_value
.playouts
+= descent
[dlen
].value
.playouts
;
340 seq_value
.value
+= descent
[dlen
].value
.value
* descent
[dlen
].value
.playouts
;
341 n
= descent
[dlen
++].node
;
342 assert(n
== t
->root
|| n
->parent
);
344 fprintf(stderr
, "%s+-- UCT sent us to [%s:%d] %d,%f\n",
345 spaces
, coord2sstr(n
->coord
, t
->board
),
346 n
->coord
, n
->u
.playouts
,
347 tree_node_get_value(t
, parity
, n
->u
.value
));
349 /* Add virtual loss if we need to; this is used to discourage
350 * other threads from visiting this node in case of multiple
351 * threads doing the tree search. */
353 stats_add_result(&n
->u
, node_color
== S_BLACK
? 0.0 : 1.0, u
->virtual_loss
);
355 assert(n
->coord
>= -1);
356 if (amaf
&& !is_pass(n
->coord
))
357 record_amaf_move(amaf
, n
->coord
, node_color
);
359 struct move m
= { n
->coord
, node_color
};
360 int res
= board_play(&b2
, &m
);
362 if (res
< 0 || (!is_pass(m
.coord
) && !group_at(&b2
, m
.coord
)) /* suicide */
363 || b2
.superko_violation
) {
365 for (struct tree_node
*ni
= n
; ni
; ni
= ni
->parent
)
366 fprintf(stderr
, "%s<%"PRIhash
"> ", coord2sstr(ni
->coord
, t
->board
), ni
->hash
);
367 fprintf(stderr
, "marking invalid %s node %d,%d res %d group %d spk %d\n",
368 stone2str(node_color
), coord_x(n
->coord
,b
), coord_y(n
->coord
,b
),
369 res
, group_at(&b2
, m
.coord
), b2
.superko_violation
);
371 n
->hints
|= TREE_HINT_INVALID
;
376 if (is_pass(n
->coord
))
383 amaf
->game_baselen
= amaf
->gamelen
;
384 amaf
->record_nakade
= u
->playout_amaf_nakade
;
387 if (t
->use_extra_komi
&& u
->dynkomi
->persim
) {
388 b2
.komi
+= round(u
->dynkomi
->persim(u
->dynkomi
, &b2
, t
, n
));
392 /* XXX: No dead groups support. */
393 floating_t score
= board_official_score(&b2
, NULL
);
394 /* Result from black's perspective (no matter who
395 * the player; black's perspective is always
396 * what the tree stores. */
397 result
= - (score
* 2);
400 fprintf(stderr
, "[%d..%d] %s p-p scoring playout result %d (W %f)\n",
401 player_color
, node_color
, coord2sstr(n
->coord
, t
->board
), result
, score
);
403 board_print(&b2
, stderr
);
405 board_ownermap_fill(&u
->ownermap
, &b2
);
407 } else { // assert(tree_leaf_node(n));
408 /* In case of parallel tree search, the assertion might
409 * not hold if two threads chew on the same node. */
410 result
= uct_leaf_node(u
, &b2
, player_color
, amaf
, descent
, &dlen
, significant
, t
, n
, node_color
, spaces
);
413 if (amaf
&& u
->playout_amaf_cutoff
) {
414 unsigned int cutoff
= amaf
->game_baselen
;
415 cutoff
+= (amaf
->gamelen
- amaf
->game_baselen
) * u
->playout_amaf_cutoff
/ 100;
416 /* Now, reconstruct the amaf map. */
417 memset(amaf
->map
, 0, board_size2(&b2
) * sizeof(*amaf
->map
));
418 for (unsigned int i
= 0; i
< cutoff
; i
++) {
419 coord_t coord
= amaf
->game
[i
].coord
;
420 enum stone color
= amaf
->game
[i
].color
;
421 if (amaf
->map
[coord
] == S_NONE
|| amaf
->map
[coord
] == color
) {
422 amaf
->map
[coord
] = color
;
423 /* Nakade always recorded for in-tree part */
424 } else if (amaf
->record_nakade
|| i
<= amaf
->game_baselen
) {
425 amaf_op(amaf
->map
[n
->coord
], +);
430 /* Record the result. */
432 assert(n
== t
->root
|| n
->parent
);
433 floating_t rval
= scale_value(u
, b
, result
);
434 u
->policy
->update(u
->policy
, t
, n
, node_color
, player_color
, amaf
, &b2
, rval
);
436 if (t
->use_extra_komi
) {
437 stats_add_result(&u
->dynkomi
->score
, (floating_t
) 0.5 * result
, 1);
438 stats_add_result(&u
->dynkomi
->value
, rval
, 1);
441 if (u
->local_tree
&& n
->parent
&& !is_pass(n
->coord
) && dlen
> 0) {
442 /* Possibly transform the rval appropriately. */
443 if (!u
->local_tree_rootseqval
) {
444 floating_t expval
= seq_value
.value
/ seq_value
.playouts
;
445 rval
= stats_temper_value(rval
, expval
, u
->local_tree
);
448 /* Get the local sequences and record them in ltree. */
449 /* We will look for sequence starts in our descent
450 * history, then run record_local_sequence() for each
451 * found sequence start; record_local_sequence() may
452 * pick longer sequences from descent history then,
453 * which is expected as it will create new lnodes. */
454 enum stone seq_color
= player_color
;
455 /* First move always starts a sequence. */
456 record_local_sequence(u
, t
, descent
, dlen
, 1, seq_color
, rval
);
457 seq_color
= stone_other(seq_color
);
458 for (int dseqi
= 2; dseqi
< dlen
; dseqi
++, seq_color
= stone_other(seq_color
)) {
459 if (u
->local_tree_allseq
) {
460 /* We are configured to record all subsequences. */
461 record_local_sequence(u
, t
, descent
, dlen
, dseqi
, seq_color
, rval
);
464 if (descent
[dseqi
].node
->d
>= u
->tenuki_d
) {
465 /* Tenuki! Record the fresh sequence. */
466 record_local_sequence(u
, t
, descent
, dlen
, dseqi
, seq_color
, rval
);
469 if (descent
[dseqi
].lnode
&& !descent
[dseqi
].lnode
) {
470 /* Record result for in-descent picked sequence. */
471 record_local_sequence(u
, t
, descent
, dlen
, dseqi
, seq_color
, rval
);
478 /* We need to undo the virtual loss we added during descend. */
479 if (u
->virtual_loss
) {
480 floating_t loss
= node_color
== S_BLACK
? 0.0 : 1.0;
481 for (; n
->parent
; n
= n
->parent
) {
482 stats_rm_result(&n
->u
, loss
, u
->virtual_loss
);
491 board_done_noalloc(&b2
);
496 uct_playouts(struct uct
*u
, struct board
*b
, enum stone color
, struct tree
*t
, struct time_info
*ti
)
499 if (ti
&& ti
->dim
== TD_GAMES
) {
500 /* We must halt if uct_search_check_stop() decides that it is time to stop. For
501 * example, if memory gets full, we may well have to stop early; in any case since
502 * uct_search_progress() is no longer called to update the dynkomi information and
503 * show progress information, continuing might well give strange results. Normally
504 * uct_search_check_stop() will stop at s->stop.worst.playouts, which is equal to
505 * ti->len.games according to timeinfo.c:time_stop_conditions(). */
506 for (i
= 0; !uct_halt
&& t
->root
->u
.playouts
<= ti
->len
.games
; i
++)
507 uct_playout(u
, b
, color
, t
);
509 for (i
= 0; !uct_halt
; i
++)
510 uct_playout(u
, b
, color
, t
);