14 #include "uct/internal.h"
18 static struct tree_node
*
19 tree_init_node(struct tree
*t
, coord_t coord
, int depth
)
21 struct tree_node
*n
= calloc(1, sizeof(*n
));
24 if (depth
> t
->max_depth
)
30 tree_init(struct board
*board
, enum stone color
)
32 struct tree
*t
= calloc(1, sizeof(*t
));
34 /* The root PASS move is only virtual, we never play it. */
35 t
->root
= tree_init_node(t
, pass
, 0);
36 t
->root_symmetry
= board
->symmetry
;
42 tree_done_node(struct tree
*t
, struct tree_node
*n
)
44 struct tree_node
*ni
= n
->children
;
46 struct tree_node
*nj
= ni
->sibling
;
47 tree_done_node(t
, ni
);
54 tree_done(struct tree
*t
)
56 tree_done_node(t
, t
->root
);
62 tree_node_dump(struct tree
*tree
, struct tree_node
*node
, int l
, int thres
)
64 for (int i
= 0; i
< l
; i
++) fputc(' ', stderr
);
66 for (struct tree_node
*ni
= node
->children
; ni
; ni
= ni
->sibling
)
68 fprintf(stderr
, "[%s] %f (%d/%d playouts [prior %d/%d amaf %d/%d]; hints %x; %d children)\n", coord2sstr(node
->coord
, tree
->board
), node
->u
.value
, node
->u
.wins
, node
->u
.playouts
, node
->prior
.wins
, node
->prior
.playouts
, node
->amaf
.wins
, node
->amaf
.playouts
, node
->hints
, children
);
70 /* Print nodes sorted by #playouts. */
72 struct tree_node
*nbox
[1000]; int nboxl
= 0;
73 for (struct tree_node
*ni
= node
->children
; ni
; ni
= ni
->sibling
)
74 if (ni
->u
.playouts
> thres
)
79 for (int i
= 0; i
< nboxl
; i
++)
80 if (nbox
[i
] && (best
< 0 || nbox
[i
]->u
.playouts
> nbox
[best
]->u
.playouts
))
84 tree_node_dump(tree
, nbox
[best
], l
+ 1, /* node->u.value < 0.1 ? 0 : */ thres
);
90 tree_dump(struct tree
*tree
, int thres
)
92 if (thres
&& tree
->root
->u
.playouts
/ thres
> 100) {
93 /* Be a bit sensible about this; the opening book can create
94 * huge dumps at first. */
95 thres
= tree
->root
->u
.playouts
/ 100 * (thres
< 1000 ? 1 : thres
/ 1000);
97 tree_node_dump(tree
, tree
->root
, 0, thres
);
102 tree_book_name(struct board
*b
)
104 static char buf
[256];
105 if (b
->handicap
> 0) {
106 sprintf(buf
, "uctbook-%d-%02.01f-h%d.pachitree", b
->size
- 2, b
->komi
, b
->handicap
);
108 sprintf(buf
, "uctbook-%d-%02.01f.pachitree", b
->size
- 2, b
->komi
);
114 tree_node_save(FILE *f
, struct tree_node
*node
, int thres
)
117 fwrite(((void *) node
) + offsetof(struct tree_node
, depth
),
118 sizeof(struct tree_node
) - offsetof(struct tree_node
, depth
),
121 if (node
->u
.playouts
>= thres
)
122 for (struct tree_node
*ni
= node
->children
; ni
; ni
= ni
->sibling
)
123 tree_node_save(f
, ni
, thres
);
129 tree_save(struct tree
*tree
, struct board
*b
, int thres
)
131 char *filename
= tree_book_name(b
);
132 FILE *f
= fopen(filename
, "wb");
137 tree_node_save(f
, tree
->root
, thres
);
144 tree_node_load(FILE *f
, struct tree_node
*node
, int *num
, bool invert
)
148 fread(((void *) node
) + offsetof(struct tree_node
, depth
),
149 sizeof(struct tree_node
) - offsetof(struct tree_node
, depth
),
152 /* Keep values in sane scale, otherwise we start overflowing.
153 * We may go slow here but we must be careful about not getting
154 * too huge integers.*/
155 #define MAX_PLAYOUTS 10000000
156 if (node
->u
.playouts
> MAX_PLAYOUTS
) {
157 int over
= node
->u
.playouts
- MAX_PLAYOUTS
;
158 node
->u
.wins
-= ((double) node
->u
.wins
/ node
->u
.playouts
) * over
;
159 node
->u
.playouts
= MAX_PLAYOUTS
;
161 if (node
->amaf
.playouts
> MAX_PLAYOUTS
) {
162 int over
= node
->amaf
.playouts
- MAX_PLAYOUTS
;
163 node
->amaf
.wins
-= ((double) node
->amaf
.wins
/ node
->amaf
.playouts
) * over
;
164 node
->amaf
.playouts
= MAX_PLAYOUTS
;
168 node
->u
.wins
= node
->u
.playouts
- node
->u
.wins
;
169 node
->u
.value
= 1 - node
->u
.value
;
170 node
->amaf
.wins
= node
->amaf
.playouts
- node
->amaf
.wins
;
171 node
->amaf
.value
= 1 - node
->amaf
.value
;
172 node
->prior
.wins
= node
->prior
.playouts
- node
->prior
.wins
;
173 node
->prior
.value
= 1 - node
->prior
.value
;
176 struct tree_node
*ni
= NULL
, *ni_prev
= NULL
;
178 ni_prev
= ni
; ni
= calloc(1, sizeof(*ni
));
182 ni_prev
->sibling
= ni
;
184 tree_node_load(f
, ni
, num
, invert
);
189 tree_load(struct tree
*tree
, struct board
*b
, enum stone color
)
191 char *filename
= tree_book_name(b
);
192 FILE *f
= fopen(filename
, "rb");
196 fprintf(stderr
, "Loading opening book %s...\n", filename
);
200 tree_node_load(f
, tree
->root
, &num
, color
!= S_BLACK
);
201 fprintf(stderr
, "Loaded %d nodes.\n", num
);
207 static struct tree_node
*
208 tree_node_copy(struct tree_node
*node
)
210 struct tree_node
*n2
= malloc(sizeof(*n2
));
214 struct tree_node
*ni
= node
->children
;
215 struct tree_node
*ni2
= tree_node_copy(ni
);
216 n2
->children
= ni2
; ni2
->parent
= n2
;
217 while ((ni
= ni
->sibling
)) {
218 ni2
->sibling
= tree_node_copy(ni
);
219 ni2
= ni2
->sibling
; ni2
->parent
= n2
;
225 tree_copy(struct tree
*tree
)
227 struct tree
*t2
= malloc(sizeof(*t2
));
229 t2
->root
= tree_node_copy(tree
->root
);
235 tree_node_merge(struct tree_node
*dest
, struct tree_node
*src
)
237 dest
->hints
|= src
->hints
;
239 /* Merge the children, both are coord-sorted lists. */
240 struct tree_node
*di
= dest
->children
, *dip
= NULL
;
241 struct tree_node
*si
= src
->children
, *sip
= NULL
;
243 if (di
->coord
!= si
->coord
) {
244 /* src has some extra items or misses di */
245 struct tree_node
*si2
= si
->sibling
;
246 while (si2
&& di
->coord
!= si2
->coord
) {
250 goto next_di
; /* src misses di, move on */
251 /* chain the extra [si,si2) items before di */
256 while (si
->sibling
!= si2
) {
267 /* Matching nodes - recurse... */
268 tree_node_merge(di
, si
);
269 /* ...and move on. */
270 sip
= si
; si
= si
->sibling
;
272 dip
= di
; di
= di
->sibling
;
286 src
->children
= NULL
;
289 /* In case of prior playouts, we do not want to accumulate them
290 * over merges - they remain static after setup. However, different
291 * trees may have different priors non-deterministically. We just
292 * take the average. */
293 if (dest
->prior
.playouts
!= src
->prior
.playouts
294 || dest
->prior
.wins
!= src
->prior
.wins
) {
295 dest
->prior
.playouts
= (dest
->prior
.playouts
+ src
->prior
.playouts
) / 2;
296 dest
->prior
.wins
= (dest
->prior
.wins
+ src
->prior
.wins
) / 2;
297 if (dest
->prior
.playouts
)
298 dest
->prior
.value
= dest
->prior
.wins
/ dest
->prior
.playouts
;
301 dest
->amaf
.playouts
+= src
->amaf
.playouts
;
302 dest
->amaf
.wins
+= src
->amaf
.wins
;
303 if (dest
->amaf
.playouts
)
304 dest
->amaf
.value
= dest
->amaf
.wins
/ dest
->amaf
.playouts
;
306 dest
->u
.playouts
+= src
->u
.playouts
;
307 dest
->u
.wins
+= src
->u
.wins
;
308 if (dest
->prior
.playouts
+ dest
->amaf
.playouts
+ dest
->u
.playouts
)
309 tree_update_node_value(dest
);
312 /* Merge two trees built upon the same board. Note that the operation is
313 * destructive on src. */
315 tree_merge(struct tree
*dest
, struct tree
*src
)
317 if (src
->max_depth
> dest
->max_depth
)
318 dest
->max_depth
= src
->max_depth
;
319 tree_node_merge(dest
->root
, src
->root
);
323 /* Tree symmetry: When possible, we will localize the tree to a single part
324 * of the board in tree_expand_node() and possibly flip along symmetry axes
325 * to another part of the board in tree_promote_at(). We follow b->symmetry
326 * guidelines here. */
330 tree_expand_node(struct tree
*t
, struct tree_node
*node
, struct board
*b
, enum stone color
, int radar
, struct uct_policy
*policy
, int parity
)
332 struct tree_node
*ni
= tree_init_node(t
, pass
, node
->depth
+ 1);
333 ni
->parent
= node
; node
->children
= ni
;
335 policy
->prior(policy
, t
, ni
, b
, color
, parity
);
337 /* The loop considers only the symmetry playground. */
339 fprintf(stderr
, "expanding %s within [%d,%d],[%d,%d] %d-%d\n",
340 coord2sstr(node
->coord
, b
),
341 b
->symmetry
.x1
, b
->symmetry
.y1
,
342 b
->symmetry
.x2
, b
->symmetry
.y2
,
343 b
->symmetry
.type
, b
->symmetry
.d
);
345 for (int i
= b
->symmetry
.x1
; i
<= b
->symmetry
.x2
; i
++) {
346 for (int j
= b
->symmetry
.y1
; j
<= b
->symmetry
.y2
; j
++) {
348 int x
= b
->symmetry
.type
== SYM_DIAG_DOWN
? board_size(b
) - 1 - i
: i
;
351 fprintf(stderr
, "drop %d,%d\n", i
, j
);
356 coord_t c
= coord_xy_otf(i
, j
, t
->board
);
357 if (board_at(b
, c
) != S_NONE
)
359 /* This looks very useful on large boards - weeds out huge amount of crufty moves. */
360 if (b
->hash
/* not empty board */ && radar
&& !board_stone_radar(b
, c
, radar
))
363 struct tree_node
*nj
= tree_init_node(t
, c
, node
->depth
+ 1);
364 nj
->parent
= node
; ni
->sibling
= nj
; ni
= nj
;
367 policy
->prior(policy
, t
, ni
, b
, color
, parity
);
374 flip_coord(struct board
*b
, coord_t c
,
375 bool flip_horiz
, bool flip_vert
, int flip_diag
)
377 int x
= coord_x(c
, b
), y
= coord_y(c
, b
);
379 int z
= x
; x
= y
; y
= z
;
382 x
= board_size(b
) - 1 - x
;
385 y
= board_size(b
) - 1 - y
;
387 return coord_xy_otf(x
, y
, b
);
391 tree_fix_node_symmetry(struct board
*b
, struct tree_node
*node
,
392 bool flip_horiz
, bool flip_vert
, int flip_diag
)
394 node
->coord
= flip_coord(b
, node
->coord
, flip_horiz
, flip_vert
, flip_diag
);
396 for (struct tree_node
*ni
= node
->children
; ni
; ni
= ni
->sibling
)
397 tree_fix_node_symmetry(b
, ni
, flip_horiz
, flip_vert
, flip_diag
);
401 tree_fix_symmetry(struct tree
*tree
, struct board
*b
, coord_t c
)
403 struct board_symmetry
*s
= &tree
->root_symmetry
;
404 int cx
= coord_x(c
, b
), cy
= coord_y(c
, b
);
406 /* playground X->h->v->d normalization
412 bool flip_horiz
= cx
< s
->x1
|| cx
> s
->x2
;
413 bool flip_vert
= cy
< s
->y1
|| cy
> s
->y2
;
417 bool dir
= (s
->type
== SYM_DIAG_DOWN
);
418 int x
= dir
^ flip_horiz
^ flip_vert
? board_size(b
) - 1 - cx
: cx
;
419 if (flip_vert
? x
< cy
: x
> cy
) {
425 fprintf(stderr
, "%s will flip %d %d %d -> %s, sym %d (%d) -> %d (%d)\n",
426 coord2sstr(c
, b
), flip_horiz
, flip_vert
, flip_diag
,
427 coord2sstr(flip_coord(b
, c
, flip_horiz
, flip_vert
, flip_diag
), b
),
428 s
->type
, s
->d
, b
->symmetry
.type
, b
->symmetry
.d
);
430 tree_fix_node_symmetry(b
, tree
->root
, flip_horiz
, flip_vert
, flip_diag
);
435 tree_unlink_node(struct tree_node
*node
)
437 struct tree_node
*ni
= node
->parent
;
438 if (ni
->children
== node
) {
439 ni
->children
= node
->sibling
;
442 while (ni
->sibling
!= node
)
444 ni
->sibling
= node
->sibling
;
449 tree_delete_node(struct tree
*tree
, struct tree_node
*node
)
451 tree_unlink_node(node
);
452 tree_done_node(tree
, node
);
456 tree_promote_node(struct tree
*tree
, struct tree_node
*node
)
458 assert(node
->parent
== tree
->root
);
459 tree_unlink_node(node
);
460 tree_done_node(tree
, tree
->root
);
462 board_symmetry_update(tree
->board
, &tree
->root_symmetry
, node
->coord
);
467 tree_promote_at(struct tree
*tree
, struct board
*b
, coord_t c
)
469 tree_fix_symmetry(tree
, b
, c
);
471 for (struct tree_node
*ni
= tree
->root
->children
; ni
; ni
= ni
->sibling
) {
472 if (ni
->coord
== c
) {
473 tree_promote_node(tree
, ni
);
481 tree_leaf_node(struct tree_node
*node
)
483 return !(node
->children
);
487 tree_update_node_value(struct tree_node
*node
)
489 bool noamaf
= node
->hints
& NODE_HINT_NOAMAF
;
490 node
->u
.value
= (float)(node
->u
.wins
+ node
->prior
.wins
+ (!noamaf
? node
->amaf
.wins
: 0))
491 / (node
->u
.playouts
+ node
->prior
.playouts
+ (!noamaf
? node
->amaf
.playouts
: 0));
493 { struct board b2
; board_size(&b2
) = 9+2;
494 fprintf(stderr
, "%s->%s %d/%d %d/%d %f\n", node
->parent
? coord2sstr(node
->parent
->coord
, &b2
) : NULL
, coord2sstr(node
->coord
, &b2
), node
->u
.wins
, node
->u
.playouts
, node
->prior
.wins
, node
->prior
.playouts
, node
->u
.value
); }