17 #include "uct/internal.h"
23 uct_get_extra_komi(struct uct
*u
, struct board
*b
)
25 float extra_komi
= board_effective_handicap(b
) * (u
->dynkomi
- b
->moves
) / u
->dynkomi
;
30 uct_progress_status(struct uct
*u
, struct tree
*t
, enum stone color
, int playouts
)
36 struct tree_node
*best
= u
->policy
->choose(u
->policy
, t
->root
, t
->board
, color
);
38 fprintf(stderr
, "... No moves left\n");
41 fprintf(stderr
, "[%d] ", playouts
);
42 fprintf(stderr
, "best %f ", tree_node_get_value(t
, 1, best
->u
.value
));
45 fprintf(stderr
, "deepest % 2d ", t
->max_depth
- t
->root
->depth
);
48 fprintf(stderr
, "| seq ");
49 for (int depth
= 0; depth
< 6; depth
++) {
50 if (best
&& best
->u
.playouts
>= 25) {
51 fprintf(stderr
, "%3s ", coord2sstr(best
->coord
, t
->board
));
52 best
= u
->policy
->choose(u
->policy
, best
, t
->board
, color
);
59 fprintf(stderr
, "| can ");
61 struct tree_node
*can
[cans
];
62 memset(can
, 0, sizeof(can
));
63 best
= t
->root
->children
;
66 while ((!can
[c
] || best
->u
.playouts
> can
[c
]->u
.playouts
) && ++c
< cans
);
67 for (int d
= 0; d
< c
; d
++) can
[d
] = can
[d
+ 1];
68 if (c
> 0) can
[c
- 1] = best
;
73 fprintf(stderr
, "%3s(%.3f) ",
74 coord2sstr(can
[cans
]->coord
, t
->board
),
75 tree_node_get_value(t
, 1, can
[cans
]->u
.value
));
81 fprintf(stderr
, "\n");
86 uct_leaf_node(struct uct
*u
, struct board
*b
, enum stone player_color
,
87 struct playout_amafmap
*amaf
,
88 struct tree
*t
, struct tree_node
*n
, enum stone node_color
,
91 enum stone next_color
= stone_other(node_color
);
92 int parity
= (next_color
== player_color
? 1 : -1);
94 /* If we don't anticipate well the opponent move during pondering
95 * (the played move has few playouts) we still need more memory
96 * during genmove to explore the tree actually played. */
97 unsigned long max_tree_size
= u
->max_tree_size
;
99 max_tree_size
= (max_tree_size
* (100 - MIN_FREE_MEM_PERCENT
)) / 100;
101 /* We need to make sure only one thread expands the node. If
102 * we are unlucky enough for two threads to meet in the same
103 * node, the latter one will simply do another simulation from
104 * the node itself, no big deal. t->nodes_size may exceed
105 * the maximum in multi-threaded case but not by much so it's ok.
106 * The size test must be before the test&set not after, to allow
107 * expansion of the node later if enough nodes have been freed. */
108 if (n
->u
.playouts
>= u
->expand_p
&& t
->nodes_size
< max_tree_size
109 && !__sync_lock_test_and_set(&n
->is_expanded
, 1)) {
110 tree_expand_node(t
, n
, b
, next_color
, u
, parity
);
113 fprintf(stderr
, "%s*-- UCT playout #%d start [%s] %f\n",
114 spaces
, n
->u
.playouts
, coord2sstr(n
->coord
, t
->board
),
115 tree_node_get_value(t
, parity
, n
->u
.value
));
117 struct playout_setup ps
= { .gamelen
= u
->gamelen
, .mercymin
= u
->mercymin
};
118 int result
= play_random_game(&ps
, b
, next_color
,
119 u
->playout_amaf
? amaf
: NULL
,
120 &u
->ownermap
, u
->playout
);
121 if (next_color
== S_WHITE
) {
122 /* We need the result from black's perspective. */
126 fprintf(stderr
, "%s -- [%d..%d] %s random playout result %d\n",
127 spaces
, player_color
, next_color
, coord2sstr(n
->coord
, t
->board
), result
);
133 scale_value(struct uct
*u
, struct board
*b
, int result
)
135 float rval
= result
> 0;
137 int vp
= u
->val_points
;
139 vp
= board_size(b
) - 1; vp
*= vp
; vp
*= 2;
142 float sval
= (float) abs(result
) / vp
;
143 sval
= sval
> 1 ? 1 : sval
;
144 if (result
< 0) sval
= 1 - sval
;
146 rval
+= u
->val_scale
* sval
;
148 rval
= (1 - u
->val_scale
) * rval
+ u
->val_scale
* sval
;
149 // fprintf(stderr, "score %d => sval %f, rval %f\n", result, sval, rval);
156 uct_playout(struct uct
*u
, struct board
*b
, enum stone player_color
, struct tree
*t
)
161 struct playout_amafmap
*amaf
= NULL
;
162 if (u
->policy
->wants_amaf
) {
163 amaf
= calloc(1, sizeof(*amaf
));
164 amaf
->map
= calloc(board_size2(&b2
) + 1, sizeof(*amaf
->map
));
165 amaf
->map
++; // -1 is pass
168 /* Walk the tree until we find a leaf, then expand it and do
169 * a random playout. */
170 struct tree_node
*n
= t
->root
;
171 enum stone node_color
= stone_other(player_color
);
172 assert(node_color
== t
->root_color
);
174 void *dstate
= NULL
, *dstater
= NULL
;
177 int pass_limit
= (board_size(&b2
) - 2) * (board_size(&b2
) - 2) / 2;
178 int passes
= is_pass(b
->last_move
.coord
) && b
->moves
> 0;
182 static char spaces
[] = "\0 ";
185 fprintf(stderr
, "--- UCT walk with color %d\n", player_color
);
187 while (!tree_leaf_node(n
) && passes
< 2) {
188 spaces
[depth
++] = ' '; spaces
[depth
] = 0;
190 /* Parity is chosen already according to the child color, since
191 * it is applied to children. */
192 node_color
= stone_other(node_color
);
193 int parity
= (node_color
== player_color
? 1 : -1);
194 n
= (!u
->random_policy_chance
|| fast_random(u
->random_policy_chance
))
195 ? u
->policy
->descend(u
->policy
, &dstate
, t
, n
, parity
, b2
.moves
> pass_limit
)
196 : u
->random_policy
->descend(u
->random_policy
, &dstater
, t
, n
, parity
, b2
.moves
> pass_limit
);
198 assert(n
== t
->root
|| n
->parent
);
200 fprintf(stderr
, "%s+-- UCT sent us to [%s:%d] %f\n",
201 spaces
, coord2sstr(n
->coord
, t
->board
), n
->coord
,
202 tree_node_get_value(t
, parity
, n
->u
.value
));
204 /* Add virtual loss if we need to; this is used to discourage
205 * other threads from visiting this node in case of multiple
206 * threads doing the tree search. */
208 stats_add_result(&n
->u
, tree_parity(t
, parity
) > 0 ? 0 : 1, 1);
210 assert(n
->coord
>= -1);
211 if (amaf
&& !is_pass(n
->coord
)) {
212 if (amaf
->map
[n
->coord
] == S_NONE
|| amaf
->map
[n
->coord
] == node_color
) {
213 amaf
->map
[n
->coord
] = node_color
;
214 } else { // XXX: Respect amaf->record_nakade
215 amaf_op(amaf
->map
[n
->coord
], +);
217 amaf
->game
[amaf
->gamelen
].coord
= n
->coord
;
218 amaf
->game
[amaf
->gamelen
].color
= node_color
;
220 assert(amaf
->gamelen
< sizeof(amaf
->game
) / sizeof(amaf
->game
[0]));
223 struct move m
= { n
->coord
, node_color
};
224 int res
= board_play(&b2
, &m
);
226 if (res
< 0 || (!is_pass(m
.coord
) && !group_at(&b2
, m
.coord
)) /* suicide */
227 || b2
.superko_violation
) {
229 for (struct tree_node
*ni
= n
; ni
; ni
= ni
->parent
)
230 fprintf(stderr
, "%s<%"PRIhash
"> ", coord2sstr(ni
->coord
, t
->board
), ni
->hash
);
231 fprintf(stderr
, "marking invalid %s node %d,%d res %d group %d spk %d\n",
232 stone2str(node_color
), coord_x(n
->coord
,b
), coord_y(n
->coord
,b
),
233 res
, group_at(&b2
, m
.coord
), b2
.superko_violation
);
235 n
->hints
|= TREE_HINT_INVALID
;
240 if (is_pass(n
->coord
))
247 amaf
->game_baselen
= amaf
->gamelen
;
248 amaf
->record_nakade
= u
->playout_amaf_nakade
;
251 if (u
->dynkomi
> b2
.moves
&& (player_color
& u
->dynkomi_mask
))
252 b2
.komi
+= uct_get_extra_komi(u
, &b2
);
255 /* XXX: No dead groups support. */
256 float score
= board_official_score(&b2
, NULL
);
257 /* Result from black's perspective (no matter who
258 * the player; black's perspective is always
259 * what the tree stores. */
260 result
= - (score
* 2);
263 fprintf(stderr
, "[%d..%d] %s p-p scoring playout result %d (W %f)\n",
264 player_color
, node_color
, coord2sstr(n
->coord
, t
->board
), result
, score
);
266 board_print(&b2
, stderr
);
268 board_ownermap_fill(&u
->ownermap
, &b2
);
270 } else { assert(u
->parallel_tree
|| tree_leaf_node(n
));
271 /* In case of parallel tree search, the assertion might
272 * not hold if two threads chew on the same node. */
273 result
= uct_leaf_node(u
, &b2
, player_color
, amaf
, t
, n
, node_color
, spaces
);
276 if (amaf
&& u
->playout_amaf_cutoff
) {
277 int cutoff
= amaf
->game_baselen
;
278 cutoff
+= (amaf
->gamelen
- amaf
->game_baselen
) * u
->playout_amaf_cutoff
/ 100;
279 /* Now, reconstruct the amaf map. */
280 memset(amaf
->map
, 0, board_size2(&b2
) * sizeof(*amaf
->map
));
281 for (int i
= 0; i
< cutoff
; i
++) {
282 coord_t coord
= amaf
->game
[i
].coord
;
283 enum stone color
= amaf
->game
[i
].color
;
284 if (amaf
->map
[coord
] == S_NONE
|| amaf
->map
[coord
] == color
) {
285 amaf
->map
[coord
] = color
;
286 /* Nakade always recorded for in-tree part */
287 } else if (amaf
->record_nakade
|| i
<= amaf
->game_baselen
) {
288 amaf_op(amaf
->map
[n
->coord
], +);
293 assert(n
== t
->root
|| n
->parent
);
295 float rval
= scale_value(u
, b
, result
);
297 u
->policy
->update(u
->policy
, t
, n
, node_color
, player_color
, amaf
, rval
);
299 if (u
->root_heuristic
&& n
->parent
) {
301 t
->chvals
= calloc(board_size2(b
), sizeof(t
->chvals
[0]));
302 t
->chchvals
= calloc(board_size2(b
), sizeof(t
->chchvals
[0]));
305 /* Possibly transform the rval appropriately. */
306 rval
= stats_temper_value(rval
, n
->parent
->u
.value
, u
->root_heuristic
);
308 struct tree_node
*ni
= n
;
309 while (ni
->parent
->parent
&& ni
->parent
->parent
->parent
)
311 if (ni
->parent
->parent
) {
312 if (likely(!is_pass(ni
->coord
)))
313 stats_add_result(&t
->chchvals
[ni
->coord
], rval
, 1);
316 assert(ni
->parent
&& !ni
->parent
->parent
);
318 if (likely(!is_pass(ni
->coord
)))
319 stats_add_result(&t
->chvals
[ni
->coord
], rval
, 1);
324 /* We need to undo the virtual loss we added during descend. */
325 if (u
->virtual_loss
) {
326 int parity
= (node_color
== player_color
? 1 : -1);
327 for (; n
->parent
; n
= n
->parent
) {
328 stats_rm_result(&n
->u
, tree_parity(t
, parity
) > 0 ? 0 : 1, 1);
333 if (dstater
) free(dstater
);
334 if (dstate
) free(dstate
);
339 board_done_noalloc(&b2
);
344 uct_playouts(struct uct
*u
, struct board
*b
, enum stone color
, struct tree
*t
)
347 for (i
= 0; !uct_halt
; i
++)
348 uct_playout(u
, b
, color
, t
);