UCT avg_score: Track avg score of last move search explicitly in the tree
[pachi/t.git] / uct / walk.c
blobb48cd5b4b829cf9177d2d9e038f80467223789f8
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 /* Ownership statistics. Value (0..1000) for each possible
142 * point describes likelihood of this point becoming black.
143 * Normally, white rate is 1000-value; exception are possible
144 * seki points, but these should be rare. */
145 fprintf(stderr, ", \"boards\": {\"territory\": [");
146 int f = 0;
147 foreach_point(t->board) {
148 if (board_at(t->board, c) == S_OFFBOARD) continue;
149 int rate = u->ownermap.map[c][S_BLACK] * 1000 / u->ownermap.playouts;
150 fprintf(stderr, "%s%d", f++ > 0 ? "," : "", rate);
151 } foreach_point_end;
152 fprintf(stderr, "]}");
156 fprintf(stderr, "}}\n");
159 void
160 uct_progress_status(struct uct *u, struct tree *t, enum stone color, int playouts, bool final)
162 switch (u->reporting) {
163 case UR_TEXT:
164 uct_progress_text(u, t, color, playouts, final);
165 break;
166 case UR_JSON:
167 case UR_JSON_BIG:
168 uct_progress_json(u, t, color, playouts, final,
169 u->reporting == UR_JSON_BIG);
170 break;
171 default: assert(0);
176 static inline void
177 record_amaf_move(struct playout_amafmap *amaf, coord_t coord)
179 assert(amaf->gamelen < MAX_GAMELEN);
180 amaf->game[amaf->gamelen++] = coord;
184 struct uct_playout_callback {
185 struct uct *uct;
186 struct tree *tree;
187 struct tree_node *lnode;
191 static coord_t
192 uct_playout_hook(struct playout_policy *playout, struct playout_setup *setup, struct board *b, enum stone color, int mode)
194 /* XXX: This is used in some non-master branches. */
195 return pass;
198 static coord_t
199 uct_playout_prepolicy(struct playout_policy *playout, struct playout_setup *setup, struct board *b, enum stone color)
201 return uct_playout_hook(playout, setup, b, color, 0);
204 static coord_t
205 uct_playout_postpolicy(struct playout_policy *playout, struct playout_setup *setup, struct board *b, enum stone color)
207 return uct_playout_hook(playout, setup, b, color, 1);
211 static int
212 uct_leaf_node(struct uct *u, struct board *b, enum stone player_color,
213 struct playout_amafmap *amaf,
214 struct uct_descent *descent, int *dlen,
215 struct tree_node *significant[2],
216 struct tree *t, struct tree_node *n, enum stone node_color,
217 char *spaces)
219 enum stone next_color = stone_other(node_color);
220 int parity = (next_color == player_color ? 1 : -1);
222 if (UDEBUGL(7))
223 fprintf(stderr, "%s*-- UCT playout #%d start [%s] %f\n",
224 spaces, n->u.playouts, coord2sstr(node_coord(n), t->board),
225 tree_node_get_value(t, parity, n->u.value));
227 struct uct_playout_callback upc = {
228 .uct = u,
229 .tree = t,
230 /* TODO: Don't necessarily restart the sequence walk when
231 * entering playout. */
232 .lnode = NULL,
235 struct playout_setup ps = {
236 .gamelen = u->gamelen,
237 .mercymin = u->mercymin,
238 .prepolicy_hook = uct_playout_prepolicy,
239 .postpolicy_hook = uct_playout_postpolicy,
240 .hook_data = &upc,
242 int result = play_random_game(&ps, b, next_color,
243 u->playout_amaf ? amaf : NULL,
244 &u->ownermap, u->playout);
245 if (next_color == S_WHITE) {
246 /* We need the result from black's perspective. */
247 result = - result;
249 if (UDEBUGL(7))
250 fprintf(stderr, "%s -- [%d..%d] %s random playout result %d\n",
251 spaces, player_color, next_color, coord2sstr(node_coord(n), t->board), result);
253 return result;
256 static floating_t
257 scale_value(struct uct *u, struct board *b, int result)
259 floating_t rval = result > 0 ? 1.0 : result < 0 ? 0.0 : 0.5;
260 if (u->val_scale && result != 0) {
261 int vp = u->val_points;
262 if (!vp) {
263 vp = board_size(b) - 1; vp *= vp; vp *= 2;
266 floating_t sval = (floating_t) abs(result) / vp;
267 sval = sval > 1 ? 1 : sval;
268 if (result < 0) sval = 1 - sval;
269 if (u->val_extra)
270 rval += u->val_scale * sval;
271 else
272 rval = (1 - u->val_scale) * rval + u->val_scale * sval;
273 // fprintf(stderr, "score %d => sval %f, rval %f\n", result, sval, rval);
275 return rval;
278 static double
279 local_value(struct uct *u, struct board *b, coord_t coord, enum stone color)
281 /* Tactical evaluation of move @coord by color @color, given
282 * simulation end position @b. I.e., a move is tactically good
283 * if the resulting group stays on board until the game end. */
284 /* We can also take into account surrounding stones, e.g. to
285 * encourage taking off external liberties during a semeai. */
286 double val = board_local_value(u->local_tree_neival, b, coord, color);
287 return (color == S_WHITE) ? 1.f - val : val;
290 static void
291 record_local_sequence(struct uct *u, struct tree *t, struct board *endb,
292 struct uct_descent *descent, int dlen, int di,
293 enum stone seq_color)
295 #define LTREE_DEBUG if (UDEBUGL(6))
297 /* Ignore pass sequences. */
298 if (is_pass(node_coord(descent[di].node)))
299 return;
301 LTREE_DEBUG board_print(endb, stderr);
302 LTREE_DEBUG fprintf(stderr, "recording local %s sequence: ",
303 stone2str(seq_color));
305 /* Sequences starting deeper are less relevant in general. */
306 int pval = LTREE_PLAYOUTS_MULTIPLIER;
307 if (u->local_tree && u->local_tree_depth_decay > 0)
308 pval = ((floating_t) pval) / pow(u->local_tree_depth_decay, di - 1);
309 if (!pval) {
310 LTREE_DEBUG fprintf(stderr, "too deep @%d\n", di);
311 return;
314 /* Pick the right local tree root... */
315 struct tree_node *lnode = seq_color == S_BLACK ? t->ltree_black : t->ltree_white;
316 lnode->u.playouts++;
318 /* ...determine the sequence value... */
319 double sval = 0.5;
320 if (u->local_tree_eval != LTE_EACH) {
321 sval = local_value(u, endb, node_coord(descent[di].node), seq_color);
322 LTREE_DEBUG fprintf(stderr, "(goal %s[%s %1.3f][%d]) ",
323 coord2sstr(node_coord(descent[di].node), t->board),
324 stone2str(seq_color), sval, descent[di].node->d);
326 if (u->local_tree_eval == LTE_TOTAL) {
327 int di0 = di;
328 while (di < dlen && (di == di0 || descent[di].node->d < u->tenuki_d)) {
329 enum stone color = (di - di0) % 2 ? stone_other(seq_color) : seq_color;
330 double rval = local_value(u, endb, node_coord(descent[di].node), color);
331 if ((di - di0) % 2)
332 rval = 1 - rval;
333 sval += rval;
334 di++;
336 sval /= (di - di0 + 1);
337 di = di0;
341 /* ...and record the sequence. */
342 int di0 = di;
343 while (di < dlen && !is_pass(node_coord(descent[di].node))
344 && (di == di0 || descent[di].node->d < u->tenuki_d)) {
345 enum stone color = (di - di0) % 2 ? stone_other(seq_color) : seq_color;
346 double rval;
347 if (u->local_tree_eval != LTE_EACH)
348 rval = sval;
349 else
350 rval = local_value(u, endb, node_coord(descent[di].node), color);
351 LTREE_DEBUG fprintf(stderr, "%s[%s %1.3f][%d] ",
352 coord2sstr(node_coord(descent[di].node), t->board),
353 stone2str(color), rval, descent[di].node->d);
354 lnode = tree_get_node(t, lnode, node_coord(descent[di++].node), true);
355 assert(lnode);
356 stats_add_result(&lnode->u, rval, pval);
359 /* Add lnode for tenuki (pass) if we descended further. */
360 if (di < dlen) {
361 double rval = u->local_tree_eval != LTE_EACH ? sval : 0.5;
362 LTREE_DEBUG fprintf(stderr, "pass ");
363 lnode = tree_get_node(t, lnode, pass, true);
364 assert(lnode);
365 stats_add_result(&lnode->u, rval, pval);
368 LTREE_DEBUG fprintf(stderr, "\n");
373 uct_playout(struct uct *u, struct board *b, enum stone player_color, struct tree *t)
375 struct board b2;
376 board_copy(&b2, b);
378 struct playout_amafmap amaf;
379 amaf.gamelen = amaf.game_baselen = 0;
381 /* Walk the tree until we find a leaf, then expand it and do
382 * a random playout. */
383 struct tree_node *n = t->root;
384 enum stone node_color = stone_other(player_color);
385 assert(node_color == t->root_color);
387 /* Make sure the root node is expanded. */
388 if (tree_leaf_node(n) && !__sync_lock_test_and_set(&n->is_expanded, 1))
389 tree_expand_node(t, n, &b2, player_color, u, 1);
391 /* Tree descent history. */
392 /* XXX: This is somewhat messy since @n and descent[dlen-1].node are
393 * redundant. */
394 struct uct_descent descent[DESCENT_DLEN];
395 descent[0].node = n; descent[0].lnode = NULL;
396 int dlen = 1;
397 /* Total value of the sequence. */
398 struct move_stats seq_value = { .playouts = 0 };
399 /* The last "significant" node along the descent (i.e. node
400 * with higher than configured number of playouts). For black
401 * and white. */
402 struct tree_node *significant[2] = { NULL, NULL };
403 if (n->u.playouts >= u->significant_threshold)
404 significant[node_color - 1] = n;
406 int result;
407 int pass_limit = (board_size(&b2) - 2) * (board_size(&b2) - 2) / 2;
408 int passes = is_pass(b->last_move.coord) && b->moves > 0;
410 /* debug */
411 static char spaces[] = "\0 ";
412 /* /debug */
413 if (UDEBUGL(8))
414 fprintf(stderr, "--- UCT walk with color %d\n", player_color);
416 while (!tree_leaf_node(n) && passes < 2) {
417 spaces[dlen - 1] = ' '; spaces[dlen] = 0;
420 /*** Choose a node to descend to: */
422 /* Parity is chosen already according to the child color, since
423 * it is applied to children. */
424 node_color = stone_other(node_color);
425 int parity = (node_color == player_color ? 1 : -1);
427 assert(dlen < DESCENT_DLEN);
428 descent[dlen] = descent[dlen - 1];
429 if (u->local_tree && (!descent[dlen].lnode || descent[dlen].node->d >= u->tenuki_d)) {
430 /* Start new local sequence. */
431 /* Remember that node_color already holds color of the
432 * to-be-found child. */
433 descent[dlen].lnode = node_color == S_BLACK ? t->ltree_black : t->ltree_white;
436 if (!u->random_policy_chance || fast_random(u->random_policy_chance))
437 u->policy->descend(u->policy, t, &descent[dlen], parity, b2.moves > pass_limit);
438 else
439 u->random_policy->descend(u->random_policy, t, &descent[dlen], parity, b2.moves > pass_limit);
442 /*** Perform the descent: */
444 if (descent[dlen].node->u.playouts >= u->significant_threshold) {
445 significant[node_color - 1] = descent[dlen].node;
448 seq_value.playouts += descent[dlen].value.playouts;
449 seq_value.value += descent[dlen].value.value * descent[dlen].value.playouts;
450 n = descent[dlen++].node;
451 assert(n == t->root || n->parent);
452 if (UDEBUGL(7))
453 fprintf(stderr, "%s+-- UCT sent us to [%s:%d] %d,%f\n",
454 spaces, coord2sstr(node_coord(n), t->board),
455 node_coord(n), n->u.playouts,
456 tree_node_get_value(t, parity, n->u.value));
458 /* Add virtual loss if we need to; this is used to discourage
459 * other threads from visiting this node in case of multiple
460 * threads doing the tree search. */
461 if (u->virtual_loss)
462 stats_add_result(&n->u, node_color == S_BLACK ? 0.0 : 1.0, u->virtual_loss);
464 assert(node_coord(n) >= -1);
465 record_amaf_move(&amaf, node_coord(n));
467 struct move m = { node_coord(n), node_color };
468 int res = board_play(&b2, &m);
470 if (res < 0 || (!is_pass(m.coord) && !group_at(&b2, m.coord)) /* suicide */
471 || b2.superko_violation) {
472 if (UDEBUGL(4)) {
473 for (struct tree_node *ni = n; ni; ni = ni->parent)
474 fprintf(stderr, "%s<%"PRIhash"> ", coord2sstr(node_coord(ni), t->board), ni->hash);
475 fprintf(stderr, "marking invalid %s node %d,%d res %d group %d spk %d\n",
476 stone2str(node_color), coord_x(node_coord(n),b), coord_y(node_coord(n),b),
477 res, group_at(&b2, m.coord), b2.superko_violation);
479 n->hints |= TREE_HINT_INVALID;
480 result = 0;
481 goto end;
484 if (is_pass(node_coord(n)))
485 passes++;
486 else
487 passes = 0;
489 enum stone next_color = stone_other(node_color);
490 /* We need to make sure only one thread expands the node. If
491 * we are unlucky enough for two threads to meet in the same
492 * node, the latter one will simply do another simulation from
493 * the node itself, no big deal. t->nodes_size may exceed
494 * the maximum in multi-threaded case but not by much so it's ok.
495 * The size test must be before the test&set not after, to allow
496 * expansion of the node later if enough nodes have been freed. */
497 if (tree_leaf_node(n)
498 && n->u.playouts - u->virtual_loss >= u->expand_p && t->nodes_size < u->max_tree_size
499 && !__sync_lock_test_and_set(&n->is_expanded, 1))
500 tree_expand_node(t, n, &b2, next_color, u, -parity);
503 amaf.game_baselen = amaf.gamelen;
505 if (t->use_extra_komi && u->dynkomi->persim) {
506 b2.komi += round(u->dynkomi->persim(u->dynkomi, &b2, t, n));
509 if (passes >= 2) {
510 /* XXX: No dead groups support. */
511 floating_t score = board_official_score(&b2, NULL);
512 /* Result from black's perspective (no matter who
513 * the player; black's perspective is always
514 * what the tree stores. */
515 result = - (score * 2);
517 if (UDEBUGL(5))
518 fprintf(stderr, "[%d..%d] %s p-p scoring playout result %d (W %f)\n",
519 player_color, node_color, coord2sstr(node_coord(n), t->board), result, score);
520 if (UDEBUGL(6))
521 board_print(&b2, stderr);
523 board_ownermap_fill(&u->ownermap, &b2);
525 } else { // assert(tree_leaf_node(n));
526 /* In case of parallel tree search, the assertion might
527 * not hold if two threads chew on the same node. */
528 result = uct_leaf_node(u, &b2, player_color, &amaf, descent, &dlen, significant, t, n, node_color, spaces);
531 if (u->policy->wants_amaf && u->playout_amaf_cutoff) {
532 unsigned int cutoff = amaf.game_baselen;
533 cutoff += (amaf.gamelen - amaf.game_baselen) * u->playout_amaf_cutoff / 100;
534 amaf.gamelen = cutoff;
537 /* Record the result. */
539 assert(n == t->root || n->parent);
540 floating_t rval = scale_value(u, b, result);
541 u->policy->update(u->policy, t, n, node_color, player_color, &amaf, &b2, rval);
543 stats_add_result(&t->avg_score, result / 2, 1);
544 if (t->use_extra_komi) {
545 stats_add_result(&u->dynkomi->score, result / 2, 1);
546 stats_add_result(&u->dynkomi->value, rval, 1);
549 if (u->local_tree && n->parent && !is_pass(node_coord(n)) && dlen > 0) {
550 /* Get the local sequences and record them in ltree. */
551 /* We will look for sequence starts in our descent
552 * history, then run record_local_sequence() for each
553 * found sequence start; record_local_sequence() may
554 * pick longer sequences from descent history then,
555 * which is expected as it will create new lnodes. */
556 enum stone seq_color = player_color;
557 /* First move always starts a sequence. */
558 record_local_sequence(u, t, &b2, descent, dlen, 1, seq_color);
559 seq_color = stone_other(seq_color);
560 for (int dseqi = 2; dseqi < dlen; dseqi++, seq_color = stone_other(seq_color)) {
561 if (u->local_tree_allseq) {
562 /* We are configured to record all subsequences. */
563 record_local_sequence(u, t, &b2, descent, dlen, dseqi, seq_color);
564 continue;
566 if (descent[dseqi].node->d >= u->tenuki_d) {
567 /* Tenuki! Record the fresh sequence. */
568 record_local_sequence(u, t, &b2, descent, dlen, dseqi, seq_color);
569 continue;
571 if (descent[dseqi].lnode && !descent[dseqi].lnode) {
572 /* Record result for in-descent picked sequence. */
573 record_local_sequence(u, t, &b2, descent, dlen, dseqi, seq_color);
574 continue;
579 end:
580 /* We need to undo the virtual loss we added during descend. */
581 if (u->virtual_loss) {
582 floating_t loss = node_color == S_BLACK ? 0.0 : 1.0;
583 for (; n->parent; n = n->parent) {
584 stats_rm_result(&n->u, loss, u->virtual_loss);
585 loss = 1.0 - loss;
589 board_done_noalloc(&b2);
590 return result;
594 uct_playouts(struct uct *u, struct board *b, enum stone color, struct tree *t, struct time_info *ti)
596 int i;
597 if (ti && ti->dim == TD_GAMES) {
598 for (i = 0; t->root->u.playouts <= ti->len.games && !uct_halt; i++)
599 uct_playout(u, b, color, t);
600 } else {
601 for (i = 0; !uct_halt; i++)
602 uct_playout(u, b, color, t);
604 return i;