UCT jsonbig: Include information about the board coloring
[pachi/t.git] / uct / walk.c
blob61fd0584d286e0e6816be636d66f91cf2a081beb
1 #include <assert.h>
2 #include <math.h>
3 #include <pthread.h>
4 #include <signal.h>
5 #include <stdio.h>
6 #include <stdlib.h>
7 #include <string.h>
9 #define DEBUG
11 #include "debug.h"
12 #include "board.h"
13 #include "move.h"
14 #include "playout.h"
15 #include "probdist.h"
16 #include "random.h"
17 #include "tactics/util.h"
18 #include "uct/dynkomi.h"
19 #include "uct/internal.h"
20 #include "uct/search.h"
21 #include "uct/tree.h"
22 #include "uct/uct.h"
23 #include "uct/walk.h"
25 #define DESCENT_DLEN 512
28 void
29 uct_progress_text(struct uct *u, struct tree *t, enum stone color, int playouts, bool final)
31 if (!UDEBUGL(0))
32 return;
34 /* Best move */
35 struct tree_node *best = u->policy->choose(u->policy, t->root, t->board, color, resign);
36 if (!best) {
37 fprintf(stderr, "... No moves left\n");
38 return;
40 fprintf(stderr, "[%d] ", playouts);
41 fprintf(stderr, "best %f ", tree_node_get_value(t, 1, best->u.value));
43 /* Dynamic komi */
44 if (t->use_extra_komi)
45 fprintf(stderr, "komi %.1f ", t->extra_komi);
47 /* Best sequence */
48 fprintf(stderr, "| seq ");
49 for (int depth = 0; depth < 4; depth++) {
50 if (best && best->u.playouts >= 25) {
51 fprintf(stderr, "%3s ", coord2sstr(node_coord(best), t->board));
52 best = u->policy->choose(u->policy, best, t->board, color, resign);
53 } else {
54 fprintf(stderr, " ");
58 /* Best candidates */
59 fprintf(stderr, "| can ");
60 int cans = 4;
61 struct tree_node *can[cans];
62 memset(can, 0, sizeof(can));
63 best = t->root->children;
64 while (best) {
65 int c = 0;
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;
69 best = best->sibling;
71 while (--cans >= 0) {
72 if (can[cans]) {
73 fprintf(stderr, "%3s(%.3f) ",
74 coord2sstr(node_coord(can[cans]), t->board),
75 tree_node_get_value(t, 1, can[cans]->u.value));
76 } else {
77 fprintf(stderr, " ");
81 fprintf(stderr, "\n");
84 void
85 uct_progress_json(struct uct *u, struct tree *t, enum stone color, int playouts, bool final, bool big)
87 /* Prefix indicating JSON line. */
88 fprintf(stderr, "{\"%s\": {", final ? "move" : "frame");
90 /* Plaout count */
91 fprintf(stderr, "\"playouts\": %d", playouts);
93 /* Dynamic komi */
94 if (t->use_extra_komi)
95 fprintf(stderr, ", \"extrakomi\": %.1f", t->extra_komi);
97 struct tree_node *best = u->policy->choose(u->policy, t->root, t->board, color, resign);
98 if (best) {
99 /* Best move */
100 fprintf(stderr, ", \"best\": {\"%s\": %f}",
101 coord2sstr(best->coord, t->board),
102 tree_node_get_value(t, 1, best->u.value));
104 /* Best sequence */
105 fprintf(stderr, ", \"seq\": [");
106 for (int depth = 0; depth < 4; depth++) {
107 if (!best || best->u.playouts < 25) break;
108 fprintf(stderr, "%s\"%s\"", depth > 0 ? "," : "",
109 coord2sstr(best->coord, t->board));
110 best = u->policy->choose(u->policy, best, t->board, color, resign);
112 fprintf(stderr, "]");
115 /* Best candidates */
116 int cans = 4;
117 struct tree_node *can[cans];
118 memset(can, 0, sizeof(can));
119 best = t->root->children;
120 while (best) {
121 int c = 0;
122 while ((!can[c] || best->u.playouts > can[c]->u.playouts) && ++c < cans);
123 for (int d = 0; d < c; d++) can[d] = can[d + 1];
124 if (c > 0) can[c - 1] = best;
125 best = best->sibling;
127 fprintf(stderr, ", \"can\": [");
128 while (--cans >= 0) {
129 if (!can[cans]) break;
130 fprintf(stderr, "%s{\"%s\":%.3f}",
131 cans < 3 ? "," : "",
132 coord2sstr(can[cans]->coord, t->board),
133 tree_node_get_value(t, 1, can[cans]->u.value));
135 fprintf(stderr, "]");
137 if (big) {
138 /* Average score. */
139 if (t->avg_score.playouts > 0)
140 fprintf(stderr, ", \"avg\": {\"score\": %.3f}", t->avg_score.value);
141 /* Per-intersection information. */
142 fprintf(stderr, ", \"boards\": {");
143 /* Position coloring information. */
144 fprintf(stderr, "\"colors\": [");
145 int f = 0;
146 foreach_point(t->board) {
147 if (board_at(t->board, c) == S_OFFBOARD) continue;
148 fprintf(stderr, "%s%d", f++ > 0 ? "," : "", board_at(t->board, c));
149 } foreach_point_end;
150 fprintf(stderr, "]");
151 /* Ownership statistics. Value (0..1000) for each possible
152 * point describes likelihood of this point becoming black.
153 * Normally, white rate is 1000-value; exception are possible
154 * seki points, but these should be rare. */
155 fprintf(stderr, ", \"territory\": [");
156 f = 0;
157 foreach_point(t->board) {
158 if (board_at(t->board, c) == S_OFFBOARD) continue;
159 int rate = u->ownermap.map[c][S_BLACK] * 1000 / u->ownermap.playouts;
160 fprintf(stderr, "%s%d", f++ > 0 ? "," : "", rate);
161 } foreach_point_end;
162 fprintf(stderr, "]");
163 fprintf(stderr, "}");
166 fprintf(stderr, "}}\n");
169 void
170 uct_progress_status(struct uct *u, struct tree *t, enum stone color, int playouts, bool final)
172 switch (u->reporting) {
173 case UR_TEXT:
174 uct_progress_text(u, t, color, playouts, final);
175 break;
176 case UR_JSON:
177 case UR_JSON_BIG:
178 uct_progress_json(u, t, color, playouts, final,
179 u->reporting == UR_JSON_BIG);
180 break;
181 default: assert(0);
186 static inline void
187 record_amaf_move(struct playout_amafmap *amaf, coord_t coord)
189 assert(amaf->gamelen < MAX_GAMELEN);
190 amaf->game[amaf->gamelen++] = coord;
194 struct uct_playout_callback {
195 struct uct *uct;
196 struct tree *tree;
197 struct tree_node *lnode;
201 static coord_t
202 uct_playout_hook(struct playout_policy *playout, struct playout_setup *setup, struct board *b, enum stone color, int mode)
204 /* XXX: This is used in some non-master branches. */
205 return pass;
208 static coord_t
209 uct_playout_prepolicy(struct playout_policy *playout, struct playout_setup *setup, struct board *b, enum stone color)
211 return uct_playout_hook(playout, setup, b, color, 0);
214 static coord_t
215 uct_playout_postpolicy(struct playout_policy *playout, struct playout_setup *setup, struct board *b, enum stone color)
217 return uct_playout_hook(playout, setup, b, color, 1);
221 static int
222 uct_leaf_node(struct uct *u, struct board *b, enum stone player_color,
223 struct playout_amafmap *amaf,
224 struct uct_descent *descent, int *dlen,
225 struct tree_node *significant[2],
226 struct tree *t, struct tree_node *n, enum stone node_color,
227 char *spaces)
229 enum stone next_color = stone_other(node_color);
230 int parity = (next_color == player_color ? 1 : -1);
232 if (UDEBUGL(7))
233 fprintf(stderr, "%s*-- UCT playout #%d start [%s] %f\n",
234 spaces, n->u.playouts, coord2sstr(node_coord(n), t->board),
235 tree_node_get_value(t, parity, n->u.value));
237 struct uct_playout_callback upc = {
238 .uct = u,
239 .tree = t,
240 /* TODO: Don't necessarily restart the sequence walk when
241 * entering playout. */
242 .lnode = NULL,
245 struct playout_setup ps = {
246 .gamelen = u->gamelen,
247 .mercymin = u->mercymin,
248 .prepolicy_hook = uct_playout_prepolicy,
249 .postpolicy_hook = uct_playout_postpolicy,
250 .hook_data = &upc,
252 int result = play_random_game(&ps, b, next_color,
253 u->playout_amaf ? amaf : NULL,
254 &u->ownermap, u->playout);
255 if (next_color == S_WHITE) {
256 /* We need the result from black's perspective. */
257 result = - result;
259 if (UDEBUGL(7))
260 fprintf(stderr, "%s -- [%d..%d] %s random playout result %d\n",
261 spaces, player_color, next_color, coord2sstr(node_coord(n), t->board), result);
263 return result;
266 static floating_t
267 scale_value(struct uct *u, struct board *b, int result)
269 floating_t rval = result > 0 ? 1.0 : result < 0 ? 0.0 : 0.5;
270 if (u->val_scale && result != 0) {
271 int vp = u->val_points;
272 if (!vp) {
273 vp = board_size(b) - 1; vp *= vp; vp *= 2;
276 floating_t sval = (floating_t) abs(result) / vp;
277 sval = sval > 1 ? 1 : sval;
278 if (result < 0) sval = 1 - sval;
279 if (u->val_extra)
280 rval += u->val_scale * sval;
281 else
282 rval = (1 - u->val_scale) * rval + u->val_scale * sval;
283 // fprintf(stderr, "score %d => sval %f, rval %f\n", result, sval, rval);
285 return rval;
288 static double
289 local_value(struct uct *u, struct board *b, coord_t coord, enum stone color)
291 /* Tactical evaluation of move @coord by color @color, given
292 * simulation end position @b. I.e., a move is tactically good
293 * if the resulting group stays on board until the game end. */
294 /* We can also take into account surrounding stones, e.g. to
295 * encourage taking off external liberties during a semeai. */
296 double val = board_local_value(u->local_tree_neival, b, coord, color);
297 return (color == S_WHITE) ? 1.f - val : val;
300 static void
301 record_local_sequence(struct uct *u, struct tree *t, struct board *endb,
302 struct uct_descent *descent, int dlen, int di,
303 enum stone seq_color)
305 #define LTREE_DEBUG if (UDEBUGL(6))
307 /* Ignore pass sequences. */
308 if (is_pass(node_coord(descent[di].node)))
309 return;
311 LTREE_DEBUG board_print(endb, stderr);
312 LTREE_DEBUG fprintf(stderr, "recording local %s sequence: ",
313 stone2str(seq_color));
315 /* Sequences starting deeper are less relevant in general. */
316 int pval = LTREE_PLAYOUTS_MULTIPLIER;
317 if (u->local_tree && u->local_tree_depth_decay > 0)
318 pval = ((floating_t) pval) / pow(u->local_tree_depth_decay, di - 1);
319 if (!pval) {
320 LTREE_DEBUG fprintf(stderr, "too deep @%d\n", di);
321 return;
324 /* Pick the right local tree root... */
325 struct tree_node *lnode = seq_color == S_BLACK ? t->ltree_black : t->ltree_white;
326 lnode->u.playouts++;
328 /* ...determine the sequence value... */
329 double sval = 0.5;
330 if (u->local_tree_eval != LTE_EACH) {
331 sval = local_value(u, endb, node_coord(descent[di].node), seq_color);
332 LTREE_DEBUG fprintf(stderr, "(goal %s[%s %1.3f][%d]) ",
333 coord2sstr(node_coord(descent[di].node), t->board),
334 stone2str(seq_color), sval, descent[di].node->d);
336 if (u->local_tree_eval == LTE_TOTAL) {
337 int di0 = di;
338 while (di < dlen && (di == di0 || descent[di].node->d < u->tenuki_d)) {
339 enum stone color = (di - di0) % 2 ? stone_other(seq_color) : seq_color;
340 double rval = local_value(u, endb, node_coord(descent[di].node), color);
341 if ((di - di0) % 2)
342 rval = 1 - rval;
343 sval += rval;
344 di++;
346 sval /= (di - di0 + 1);
347 di = di0;
351 /* ...and record the sequence. */
352 int di0 = di;
353 while (di < dlen && !is_pass(node_coord(descent[di].node))
354 && (di == di0 || descent[di].node->d < u->tenuki_d)) {
355 enum stone color = (di - di0) % 2 ? stone_other(seq_color) : seq_color;
356 double rval;
357 if (u->local_tree_eval != LTE_EACH)
358 rval = sval;
359 else
360 rval = local_value(u, endb, node_coord(descent[di].node), color);
361 LTREE_DEBUG fprintf(stderr, "%s[%s %1.3f][%d] ",
362 coord2sstr(node_coord(descent[di].node), t->board),
363 stone2str(color), rval, descent[di].node->d);
364 lnode = tree_get_node(t, lnode, node_coord(descent[di++].node), true);
365 assert(lnode);
366 stats_add_result(&lnode->u, rval, pval);
369 /* Add lnode for tenuki (pass) if we descended further. */
370 if (di < dlen) {
371 double rval = u->local_tree_eval != LTE_EACH ? sval : 0.5;
372 LTREE_DEBUG fprintf(stderr, "pass ");
373 lnode = tree_get_node(t, lnode, pass, true);
374 assert(lnode);
375 stats_add_result(&lnode->u, rval, pval);
378 LTREE_DEBUG fprintf(stderr, "\n");
383 uct_playout(struct uct *u, struct board *b, enum stone player_color, struct tree *t)
385 struct board b2;
386 board_copy(&b2, b);
388 struct playout_amafmap amaf;
389 amaf.gamelen = amaf.game_baselen = 0;
391 /* Walk the tree until we find a leaf, then expand it and do
392 * a random playout. */
393 struct tree_node *n = t->root;
394 enum stone node_color = stone_other(player_color);
395 assert(node_color == t->root_color);
397 /* Make sure the root node is expanded. */
398 if (tree_leaf_node(n) && !__sync_lock_test_and_set(&n->is_expanded, 1))
399 tree_expand_node(t, n, &b2, player_color, u, 1);
401 /* Tree descent history. */
402 /* XXX: This is somewhat messy since @n and descent[dlen-1].node are
403 * redundant. */
404 struct uct_descent descent[DESCENT_DLEN];
405 descent[0].node = n; descent[0].lnode = NULL;
406 int dlen = 1;
407 /* Total value of the sequence. */
408 struct move_stats seq_value = { .playouts = 0 };
409 /* The last "significant" node along the descent (i.e. node
410 * with higher than configured number of playouts). For black
411 * and white. */
412 struct tree_node *significant[2] = { NULL, NULL };
413 if (n->u.playouts >= u->significant_threshold)
414 significant[node_color - 1] = n;
416 int result;
417 int pass_limit = (board_size(&b2) - 2) * (board_size(&b2) - 2) / 2;
418 int passes = is_pass(b->last_move.coord) && b->moves > 0;
420 /* debug */
421 static char spaces[] = "\0 ";
422 /* /debug */
423 if (UDEBUGL(8))
424 fprintf(stderr, "--- UCT walk with color %d\n", player_color);
426 while (!tree_leaf_node(n) && passes < 2) {
427 spaces[dlen - 1] = ' '; spaces[dlen] = 0;
430 /*** Choose a node to descend to: */
432 /* Parity is chosen already according to the child color, since
433 * it is applied to children. */
434 node_color = stone_other(node_color);
435 int parity = (node_color == player_color ? 1 : -1);
437 assert(dlen < DESCENT_DLEN);
438 descent[dlen] = descent[dlen - 1];
439 if (u->local_tree && (!descent[dlen].lnode || descent[dlen].node->d >= u->tenuki_d)) {
440 /* Start new local sequence. */
441 /* Remember that node_color already holds color of the
442 * to-be-found child. */
443 descent[dlen].lnode = node_color == S_BLACK ? t->ltree_black : t->ltree_white;
446 if (!u->random_policy_chance || fast_random(u->random_policy_chance))
447 u->policy->descend(u->policy, t, &descent[dlen], parity, b2.moves > pass_limit);
448 else
449 u->random_policy->descend(u->random_policy, t, &descent[dlen], parity, b2.moves > pass_limit);
452 /*** Perform the descent: */
454 if (descent[dlen].node->u.playouts >= u->significant_threshold) {
455 significant[node_color - 1] = descent[dlen].node;
458 seq_value.playouts += descent[dlen].value.playouts;
459 seq_value.value += descent[dlen].value.value * descent[dlen].value.playouts;
460 n = descent[dlen++].node;
461 assert(n == t->root || n->parent);
462 if (UDEBUGL(7))
463 fprintf(stderr, "%s+-- UCT sent us to [%s:%d] %d,%f\n",
464 spaces, coord2sstr(node_coord(n), t->board),
465 node_coord(n), n->u.playouts,
466 tree_node_get_value(t, parity, n->u.value));
468 /* Add virtual loss if we need to; this is used to discourage
469 * other threads from visiting this node in case of multiple
470 * threads doing the tree search. */
471 if (u->virtual_loss)
472 stats_add_result(&n->u, node_color == S_BLACK ? 0.0 : 1.0, u->virtual_loss);
474 assert(node_coord(n) >= -1);
475 record_amaf_move(&amaf, node_coord(n));
477 struct move m = { node_coord(n), node_color };
478 int res = board_play(&b2, &m);
480 if (res < 0 || (!is_pass(m.coord) && !group_at(&b2, m.coord)) /* suicide */
481 || b2.superko_violation) {
482 if (UDEBUGL(4)) {
483 for (struct tree_node *ni = n; ni; ni = ni->parent)
484 fprintf(stderr, "%s<%"PRIhash"> ", coord2sstr(node_coord(ni), t->board), ni->hash);
485 fprintf(stderr, "marking invalid %s node %d,%d res %d group %d spk %d\n",
486 stone2str(node_color), coord_x(node_coord(n),b), coord_y(node_coord(n),b),
487 res, group_at(&b2, m.coord), b2.superko_violation);
489 n->hints |= TREE_HINT_INVALID;
490 result = 0;
491 goto end;
494 if (is_pass(node_coord(n)))
495 passes++;
496 else
497 passes = 0;
499 enum stone next_color = stone_other(node_color);
500 /* We need to make sure only one thread expands the node. If
501 * we are unlucky enough for two threads to meet in the same
502 * node, the latter one will simply do another simulation from
503 * the node itself, no big deal. t->nodes_size may exceed
504 * the maximum in multi-threaded case but not by much so it's ok.
505 * The size test must be before the test&set not after, to allow
506 * expansion of the node later if enough nodes have been freed. */
507 if (tree_leaf_node(n)
508 && n->u.playouts - u->virtual_loss >= u->expand_p && t->nodes_size < u->max_tree_size
509 && !__sync_lock_test_and_set(&n->is_expanded, 1))
510 tree_expand_node(t, n, &b2, next_color, u, -parity);
513 amaf.game_baselen = amaf.gamelen;
515 if (t->use_extra_komi && u->dynkomi->persim) {
516 b2.komi += round(u->dynkomi->persim(u->dynkomi, &b2, t, n));
519 if (passes >= 2) {
520 /* XXX: No dead groups support. */
521 floating_t score = board_official_score(&b2, NULL);
522 /* Result from black's perspective (no matter who
523 * the player; black's perspective is always
524 * what the tree stores. */
525 result = - (score * 2);
527 if (UDEBUGL(5))
528 fprintf(stderr, "[%d..%d] %s p-p scoring playout result %d (W %f)\n",
529 player_color, node_color, coord2sstr(node_coord(n), t->board), result, score);
530 if (UDEBUGL(6))
531 board_print(&b2, stderr);
533 board_ownermap_fill(&u->ownermap, &b2);
535 } else { // assert(tree_leaf_node(n));
536 /* In case of parallel tree search, the assertion might
537 * not hold if two threads chew on the same node. */
538 result = uct_leaf_node(u, &b2, player_color, &amaf, descent, &dlen, significant, t, n, node_color, spaces);
541 if (u->policy->wants_amaf && u->playout_amaf_cutoff) {
542 unsigned int cutoff = amaf.game_baselen;
543 cutoff += (amaf.gamelen - amaf.game_baselen) * u->playout_amaf_cutoff / 100;
544 amaf.gamelen = cutoff;
547 /* Record the result. */
549 assert(n == t->root || n->parent);
550 floating_t rval = scale_value(u, b, result);
551 u->policy->update(u->policy, t, n, node_color, player_color, &amaf, &b2, rval);
553 stats_add_result(&t->avg_score, result / 2, 1);
554 if (t->use_extra_komi) {
555 stats_add_result(&u->dynkomi->score, result / 2, 1);
556 stats_add_result(&u->dynkomi->value, rval, 1);
559 if (u->local_tree && n->parent && !is_pass(node_coord(n)) && dlen > 0) {
560 /* Get the local sequences and record them in ltree. */
561 /* We will look for sequence starts in our descent
562 * history, then run record_local_sequence() for each
563 * found sequence start; record_local_sequence() may
564 * pick longer sequences from descent history then,
565 * which is expected as it will create new lnodes. */
566 enum stone seq_color = player_color;
567 /* First move always starts a sequence. */
568 record_local_sequence(u, t, &b2, descent, dlen, 1, seq_color);
569 seq_color = stone_other(seq_color);
570 for (int dseqi = 2; dseqi < dlen; dseqi++, seq_color = stone_other(seq_color)) {
571 if (u->local_tree_allseq) {
572 /* We are configured to record all subsequences. */
573 record_local_sequence(u, t, &b2, descent, dlen, dseqi, seq_color);
574 continue;
576 if (descent[dseqi].node->d >= u->tenuki_d) {
577 /* Tenuki! Record the fresh sequence. */
578 record_local_sequence(u, t, &b2, descent, dlen, dseqi, seq_color);
579 continue;
581 if (descent[dseqi].lnode && !descent[dseqi].lnode) {
582 /* Record result for in-descent picked sequence. */
583 record_local_sequence(u, t, &b2, descent, dlen, dseqi, seq_color);
584 continue;
589 end:
590 /* We need to undo the virtual loss we added during descend. */
591 if (u->virtual_loss) {
592 floating_t loss = node_color == S_BLACK ? 0.0 : 1.0;
593 for (; n->parent; n = n->parent) {
594 stats_rm_result(&n->u, loss, u->virtual_loss);
595 loss = 1.0 - loss;
599 board_done_noalloc(&b2);
600 return result;
604 uct_playouts(struct uct *u, struct board *b, enum stone color, struct tree *t, struct time_info *ti)
606 int i;
607 if (ti && ti->dim == TD_GAMES) {
608 for (i = 0; t->root->u.playouts <= ti->len.games && !uct_halt; i++)
609 uct_playout(u, b, color, t);
610 } else {
611 for (i = 0; !uct_halt; i++)
612 uct_playout(u, b, color, t);
614 return i;