playout/elo.c

   1 /* Playout player based on probability distribution generated over
   2  * the available moves. */
   3
   4 /* We use the ELO-based (Coulom, 2007) approach, where each board
   5  * feature (matched pattern, self-atari, capture, MC owner?, ...)
   6  * is pre-assigned "playing strength" (gamma).
   7  *
   8  * Then, the problem of choosing a move is basically a team
   9  * competition in ELO terms - each spot is represented by a team
  10  * of features appearing there; the team gamma is product of feature
  11  * gammas. The team gammas make for a probability distribution of
  12  * moves to be played.
  13  *
  14  * We use the general pattern classifier that will find the features
  15  * for us, and external datasets that can be harvested from a set
  16  * of game records (see the HACKING file for details): patterns.spat
  17  * as a dictionary of spatial stone configurations, and patterns.gamma
  18  * with strengths of particular features. */
  19
  20 /* TODO: Counter-captures are not appreciated properly. This seems to be
  21  * a major strength problem. */
  22
  23 #include <assert.h>
  24 #include <math.h>
  25 #include <stdio.h>
  26 #include <stdlib.h>
  27
  28 //#define DEBUG
  29 #include "board.h"
  30 #include "debug.h"
  31 #include "fixp.h"
  32 #include "pattern.h"
  33 #include "patternsp.h"
  34 #include "playout.h"
  35 #include "playout/elo.h"
  36 #include "random.h"
  37 #include "tactics.h"
  38 #include "uct/prior.h"
  39
  40 #define PLDEBUGL(n) DEBUGL_(p->debug_level, n)
  41
  42
  43 /* Note that the context can be shared by multiple threads! */
  44
  45 struct patternset {
  46         pattern_spec ps;
  47         struct pattern_config pc;
  48         struct features_gamma *fg;
  49 };
  50
  51 struct elo_policy {
  52         bool assess_fastpat;
  53         float selfatari;
  54         struct patternset choose, assess;
  55         playout_elo_callbackp callback; void *callback_data;
  56
  57         enum {
  58                 EAV_TOTAL,
  59                 EAV_BEST,
  60         } assess_eval;
  61         enum {
  62                 EAT_LINEAR,
  63                 EAT_ATAN,
  64                 EAT_SIGMOID,
  65         } assess_transform;
  66         double assess_sigmb;
  67 };
  68
  69
  70 /* This is the core of the policy - initializes and constructs the
  71  * probability distribution over the move candidates. */
  72
  73 int
  74 elo_get_probdist(struct playout_policy *p, struct patternset *ps, struct board *b, enum stone to_play, struct probdist *pd)
  75 {
  76         //struct elo_policy *pp = p->data;
  77         int moves = 0;
  78
  79         /* First, assign per-point probabilities. */
  80
  81         for (int f = 0; f < b->flen; f++) {
  82                 struct move m = { .coord = b->f[f], .color = to_play };
  83
  84                 /* Skip pass (for now)? */
  85                 if (is_pass(m.coord)) {
  86 skip_move:
  87                         probdist_set(pd, m.coord, 0);
  88                         continue;
  89                 }
  90                 if (PLDEBUGL(7))
  91                         fprintf(stderr, "<%d> %s\n", f, coord2sstr(m.coord, b));
  92
  93                 /* Skip invalid moves. */
  94                 if (!board_is_valid_move(b, &m))
  95                         goto skip_move;
  96
  97                 /* We shall never fill our own single-point eyes. */
  98                 /* XXX: In some rare situations, this prunes the best move:
  99                  * Bulk-five nakade with eye at 1-1 point. */
 100                 if (board_is_one_point_eye(b, m.coord, to_play)) {
 101                         goto skip_move;
 102                 }
 103
 104                 moves++;
 105                 /* Each valid move starts with gamma 1. */
 106                 double g = 1.f;
 107
 108                 /* Some easy features: */
 109                 /* XXX: We just disable them for now since we call the
 110                  * pattern matcher; you need the gammas file. */
 111 #if 0
 112                 if (is_bad_selfatari(b, to_play, m.coord))
 113                         g *= pp->selfatari;
 114 #endif
 115
 116                 /* Match pattern features: */
 117                 struct pattern pat;
 118                 pattern_match(&ps->pc, ps->ps, &pat, b, &m);
 119                 for (int i = 0; i < pat.n; i++) {
 120                         /* Multiply together gammas of all pattern features. */
 121                         double gamma = feature_gamma(ps->fg, &pat.f[i], NULL);
 122                         if (PLDEBUGL(7)) {
 123                                 char buf[256] = ""; feature2str(buf, &pat.f[i]);
 124                                 fprintf(stderr, "<%d> %s feat %s gamma %f\n", f, coord2sstr(m.coord, b), buf, gamma);
 125                         }
 126                         g *= gamma;
 127                 }
 128
 129                 probdist_set(pd, m.coord, double_to_fixp(g));
 130                 if (PLDEBUGL(7))
 131                         fprintf(stderr, "<%d> %s %f (E %f)\n", f, coord2sstr(m.coord, b), fixp_to_double(probdist_one(pd, m.coord)), g);
 132         }
 133
 134         return moves;
 135 }
 136
 137
 138 struct lprobdist {
 139         int n;
 140 #define LPD_MAX 8
 141         coord_t coords[LPD_MAX];
 142         fixp_t items[LPD_MAX];
 143         fixp_t total;
 144
 145         /* Backups of original totals for restoring. */
 146         fixp_t btotal;
 147         fixp_t browtotals_v[10];
 148         int browtotals_i[10];
 149         int browtotals_n;
 150 };
 151
 152 #if defined(BOARD_GAMMA) && defined(BOARD_TRAITS)
 153
 154 static void
 155 elo_check_probdist(struct playout_policy *p, struct board *b, enum stone to_play, struct probdist *pd, int *ignores, struct lprobdist *lpd, coord_t lc)
 156 {
 157 #if 0
 158 #define PROBDIST_EPSILON double_to_fixp(0.01)
 159         struct elo_policy *pp = p->data;
 160         if (pd->total == 0)
 161                 return;
 162
 163         /* Compare to the manually created distribution. */
 164         /* XXX: This is now broken if callback is used. */
 165
 166         probdist_alloca(pdx, b);
 167         elo_get_probdist(p, &pp->choose, b, to_play, &pdx);
 168         for (int i = 0; i < b->flen; i++) {
 169                 coord_t c = b->f[i];
 170                 if (is_pass(c)) continue;
 171                 if (c == b->ko.coord) continue;
 172                 fixp_t val = pd->items[c];
 173                 if (!is_pass(lc) && coord_is_8adjecent(lc, c, b))
 174                         for (int j = 0; j < lpd->n; j++)
 175                                 if (lpd->coords[j] == c) {
 176                                         val = lpd->items[j];
 177                                         probdist_mute(&pdx, c);
 178
 179                                 }
 180                 if (abs(pdx.items[c] - val) < PROBDIST_EPSILON)
 181                         continue;
 182                 printf("[%s %d] manual %f board %f (base %f) ", coord2sstr(c, b), b->pat3[c], fixp_to_double(pdx.items[c]), fixp_to_double(val), fixp_to_double(pd->items[c]));
 183                 board_gamma_update(b, c, to_play);
 184                 printf("plainboard %f\n", fixp_to_double(pd->items[c]));
 185                 assert(0);
 186         }
 187         for (int r = 0; r < board_size(b); r++) {
 188                 if (abs(pdx.rowtotals[r] - pd->rowtotals[r]) < PROBDIST_EPSILON)
 189                         continue;
 190                 fprintf(stderr, "row %d: manual %f board %f\n", r, fixp_to_double(pdx.rowtotals[r]), fixp_to_double(pd->rowtotals[r]));
 191                 assert(0);
 192         }
 193         assert(abs(pdx.total - pd->total) < PROBDIST_EPSILON);
 194 #undef PROBDIST_EPSILON
 195 #endif
 196 }
 197
 198 coord_t
 199 playout_elo_choose(struct playout_policy *p, struct board *b, enum stone to_play)
 200 {
 201         struct elo_policy *pp = p->data;
 202         /* The base board probdist. */
 203         struct probdist *pd = &b->prob[to_play - 1];
 204         /* The list of moves we do not consider in pd. */
 205         int ignores[10]; int ignores_n = 0;
 206         /* The list of local moves; we consider these separately. */
 207         struct lprobdist lpd = { .n = 0, .total = 0, .btotal = pd->total, .browtotals_n = 0 };
 208
 209         /* The engine might want to adjust our probdist. */
 210         if (pp->callback)
 211                 pp->callback(pp->callback_data, b, to_play, pd);
 212
 213         if (PLDEBUGL(5)) {
 214                 fprintf(stderr, "pd total pre %f lpd %f\n", fixp_to_double(pd->total), fixp_to_double(lpd.total));
 215         }
 216
 217 #define ignore_move(c_) do { \
 218         ignores[ignores_n++] = c_; \
 219         if (ignores_n > 1 && ignores[ignores_n - 1] < ignores[ignores_n - 2]) { \
 220                 /* Keep ignores[] sorted. We abuse the fact that we know \
 221                  * only one item can be out-of-order. */ \
 222                 coord_t cc = ignores[ignores_n - 2]; \
 223                 ignores[ignores_n - 2] = ignores[ignores_n - 1]; \
 224                 ignores[ignores_n - 1] = cc; \
 225         } \
 226         int rowi = coord_y(c_, pd->b); \
 227         lpd.browtotals_i[lpd.browtotals_n] = rowi; \
 228         lpd.browtotals_v[lpd.browtotals_n++] = pd->rowtotals[rowi]; \
 229         probdist_mute(pd, c_); \
 230         if (PLDEBUGL(6)) \
 231                 fprintf(stderr, "ignored move %s(%f) => tot pd %f lpd %f\n", coord2sstr(c_, pd->b), fixp_to_double(pd->items[c_]), fixp_to_double(pd->total), fixp_to_double(lpd.total)); \
 232 } while (0)
 233
 234         /* Make sure ko-prohibited move does not get picked. */
 235         if (!is_pass(b->ko.coord)) {
 236                 assert(b->ko.color == to_play);
 237                 ignore_move(b->ko.coord);
 238         }
 239
 240         /* Contiguity detection. */
 241         if (!is_pass(b->last_move.coord)) {
 242                 foreach_8neighbor(b, b->last_move.coord) {
 243                         if (c == b->ko.coord)
 244                                 continue; // already ignored
 245                         if (board_at(b, c) != S_NONE) {
 246                                 assert(probdist_one(pd, c) == 0);
 247                                 continue;
 248                         }
 249                         ignore_move(c);
 250
 251                         fixp_t val = double_to_fixp(fixp_to_double(probdist_one(pd, c)) * b->gamma->gamma[FEAT_CONTIGUITY][1]);
 252                         lpd.coords[lpd.n] = c;
 253                         lpd.items[lpd.n++] = val;
 254                         lpd.total += val;
 255                 } foreach_8neighbor_end;
 256         }
 257
 258         ignores[ignores_n] = pass;
 259         if (PLDEBUGL(5))
 260                 fprintf(stderr, "pd total post %f lpd %f\n", fixp_to_double(pd->total), fixp_to_double(lpd.total));
 261
 262         /* Verify sanity, possibly. */
 263         elo_check_probdist(p, b, to_play, pd, ignores, &lpd, b->last_move.coord);
 264
 265         /* Pick a move. */
 266         coord_t c = pass;
 267         fixp_t stab = fast_irandom(lpd.total + pd->total);
 268         if (PLDEBUGL(5))
 269                 fprintf(stderr, "stab %f / (%f + %f)\n", fixp_to_double(stab), fixp_to_double(lpd.total), fixp_to_double(pd->total));
 270         if (stab < lpd.total) {
 271                 /* Local probdist. */
 272                 if (PLDEBUGL(6)) {
 273                         /* Some debug prints. */
 274                         fixp_t tot = 0;
 275                         for (int i = 0; i < lpd.n; i++) {
 276                                 tot += lpd.items[i];
 277                                 struct pattern p;
 278                                 struct move m = { .color = to_play, .coord = lpd.coords[i] };
 279                                 if (board_at(b, m.coord) != S_NONE) {
 280                                         assert(lpd.items[i] == 0);
 281                                         continue;
 282                                 }
 283                                 pattern_match(&pp->choose.pc, pp->choose.ps, &p, b, &m);
 284                                 char s[256] = ""; pattern2str(s, &p);
 285                                 fprintf(stderr, "coord %s <%f> [tot %f] %s (p3:%d)\n",
 286                                         coord2sstr(lpd.coords[i], b), fixp_to_double(lpd.items[i]),
 287                                         fixp_to_double(tot), s,
 288                                         pattern3_by_spatial(pp->choose.pc.spat_dict, b->pat3[lpd.coords[i]]));
 289                         }
 290                 }
 291                 for (int i = 0; i < lpd.n; i++) {
 292                         if (stab <= lpd.items[i]) {
 293                                 c = lpd.coords[i];
 294                                 break;
 295                         }
 296                         stab -= lpd.items[i];
 297                 }
 298                 if (is_pass(c)) {
 299                         fprintf(stderr, "elo: local overstab [%f]\n", fixp_to_double(stab));
 300                         abort();
 301                 }
 302
 303         } else if (pd->total > 0) {
 304                 /* Global probdist. */
 305                 /* XXX: We re-stab inside. */
 306                 c = probdist_pick(pd, ignores);
 307
 308         } else {
 309                 if (PLDEBUGL(5))
 310                         fprintf(stderr, "ding!\n");
 311                 c = pass;
 312         }
 313
 314         /* Repair the damage. */
 315         if (pp->callback) {
 316                 /* XXX: Do something less horribly inefficient
 317                  * than just recomputing the whole pd. */
 318                 pd->total = 0;
 319                 for (int i = 0; i < board_size(pd->b); i++)
 320                         pd->rowtotals[i] = 0;
 321                 for (int i = 0; i < b->flen; i++) {
 322                         pd->items[b->f[i]] = 0;
 323                         board_gamma_update(b, b->f[i], to_play);
 324                 }
 325                 assert(pd->total == lpd.btotal);
 326
 327         } else {
 328                 pd->total = lpd.btotal;
 329                 /* If we touched a row multiple times (and we sure will),
 330                  * the latter value is obsolete; but since we go through
 331                  * the backups in reverse order, all is good. */
 332                 for (int j = lpd.browtotals_n - 1; j >= 0; j--)
 333                         pd->rowtotals[lpd.browtotals_i[j]] = lpd.browtotals_v[j];
 334         }
 335         return c;
 336 }
 337
 338 #else
 339
 340 coord_t
 341 playout_elo_choose(struct playout_policy *p, struct board *b, enum stone to_play)
 342 {
 343         struct elo_policy *pp = p->data;
 344         probdist_alloca(pd, b);
 345         elo_get_probdist(p, &pp->choose, b, to_play, &pd);
 346         if (pp->callback)
 347                 pp->callback(pp->callback_data, b, to_play, &pd);
 348         if (pd.total == 0)
 349                 return pass;
 350         int ignores[1] = { pass };
 351         coord_t c = probdist_pick(&pd, ignores);
 352         return c;
 353 }
 354
 355 #endif
 356
 357 void
 358 playout_elo_assess(struct playout_policy *p, struct prior_map *map, int games)
 359 {
 360         struct elo_policy *pp = p->data;
 361         probdist_alloca(pd, map->b);
 362
 363         int moves;
 364         moves = elo_get_probdist(p, &pp->assess, map->b, map->to_play, &pd);
 365
 366         /* It is a question how to transform the gamma to won games; we use
 367          * a naive approach currently, but not sure how well it works. */
 368         /* TODO: Try sqrt(p), atan(p)/pi*2. */
 369
 370         double pd_best = 0;
 371         if (pp->assess_eval == EAV_BEST) {
 372                 for (int f = 0; f < map->b->flen; f++) {
 373                         double pd_one = fixp_to_double(probdist_one(&pd, map->b->f[f]));
 374                         if (pd_one > pd_best)
 375                                 pd_best = pd_one;
 376                 }
 377         }
 378         double pd_total = fixp_to_double(probdist_total(&pd));
 379
 380         for (int f = 0; f < map->b->flen; f++) {
 381                 coord_t c = map->b->f[f];
 382                 if (!map->consider[c])
 383                         continue;
 384
 385                 double pd_one = fixp_to_double(probdist_one(&pd, c));
 386                 double val = 0;
 387                 switch (pp->assess_eval) {
 388                 case EAV_TOTAL:
 389                         val = pd_one / pd_total;
 390                         break;
 391                 case EAV_BEST:
 392                         val = pd_one / pd_best;
 393                         break;
 394                 default:
 395                         assert(0);
 396                 }
 397
 398                 switch (pp->assess_transform) {
 399                 case EAT_LINEAR:
 400                         val = val;
 401                         break;
 402                 case EAT_ATAN:
 403                         val = atan(val)/M_PI;
 404                         break;
 405                 case EAT_SIGMOID:
 406                         val = 1.0 / (1.0 + exp(-pp->assess_sigmb * (val - 0.5)));
 407                         break;
 408                 default:
 409                         assert(0);
 410                 }
 411
 412                 add_prior_value(map, c, val, games);
 413         }
 414 }
 415
 416 void
 417 playout_elo_done(struct playout_policy *p)
 418 {
 419         struct elo_policy *pp = p->data;
 420         features_gamma_done(pp->choose.fg);
 421         features_gamma_done(pp->assess.fg);
 422 }
 423
 424
 425 void
 426 playout_elo_callback(struct playout_policy *p, playout_elo_callbackp callback, void *data)
 427 {
 428         struct elo_policy *pp = p->data;
 429         pp->callback = callback;
 430         pp->callback_data = data;
 431 }
 432
 433 struct playout_policy *
 434 playout_elo_init(char *arg, struct board *b)
 435 {
 436         struct playout_policy *p = calloc2(1, sizeof(*p));
 437         struct elo_policy *pp = calloc2(1, sizeof(*pp));
 438         p->data = pp;
 439         p->choose = playout_elo_choose;
 440         p->assess = playout_elo_assess;
 441         p->done = playout_elo_done;
 442
 443         const char *gammafile = features_gamma_filename;
 444         pp->assess_sigmb = 10.0;
 445         /* Some defaults based on the table in Remi Coulom's paper. */
 446         pp->selfatari = 0.06;
 447
 448         struct pattern_config pc = DEFAULT_PATTERN_CONFIG;
 449         int xspat = -1;
 450         bool precise_selfatari = false;
 451
 452         if (arg) {
 453                 char *optspec, *next = arg;
 454                 while (*next) {
 455                         optspec = next;
 456                         next += strcspn(next, ":");
 457                         if (*next) { *next++ = 0; } else { *next = 0; }
 458
 459                         char *optname = optspec;
 460                         char *optval = strchr(optspec, '=');
 461                         if (optval) *optval++ = 0;
 462
 463                         if (!strcasecmp(optname, "selfatari") && optval) {
 464                                 pp->selfatari = atof(optval);
 465                         } else if (!strcasecmp(optname, "precisesa")) {
 466                                 /* Use precise self-atari detection within
 467                                  * fast patterns. */
 468                                 precise_selfatari = !optval || atoi(optval);
 469                         } else if (!strcasecmp(optname, "gammafile") && optval) {
 470                                 /* patterns.gamma by default. We use this,
 471                                  * and need also ${gammafile}f (e.g.
 472                                  * patterns.gammaf) for fast (MC) features. */
 473                                 gammafile = strdup(optval);
 474                         } else if (!strcasecmp(optname, "xspat") && optval) {
 475                                 /* xspat==0: don't match spatial features
 476                                  * xspat==1: match *only* spatial features */
 477                                 xspat = atoi(optval);
 478                         } else if (!strcasecmp(optname, "assess_fastpat")) {
 479                                 /* Use just fast pattern set even for the
 480                                  * node prior value assessment. */
 481                                 pp->assess_fastpat = !optval || atoi(optval);
 482                         } else if (!strcasecmp(optname, "assess_sigmb") && optval) {
 483                                 pp->assess_sigmb = atof(optval);
 484                         } else if (!strcasecmp(optname, "assess_eval") && optval) {
 485                                 /* Evaluation method for prior node value
 486                                  * assessment. */
 487                                 if (!strcasecmp(optval, "total")) {
 488                                         /* Proportion prob/totprob. */
 489                                         pp->assess_eval = EAV_TOTAL;
 490                                 } else if (!strcasecmp(optval, "best")) {
 491                                         /* Proportion prob/bestprob. */
 492                                         pp->assess_eval = EAV_BEST;
 493                                 } else {
 494                                         fprintf(stderr, "playout-elo: Invalid eval mode %s\n", optval);
 495                                         exit(1);
 496                                 }
 497                         } else if (!strcasecmp(optname, "assess_transform") && optval) {
 498                                 /* Transformation of evaluation for prior
 499                                  * node value assessment. */
 500                                 if (!strcasecmp(optval, "linear")) {
 501                                         /* No additional transformation. */
 502                                         pp->assess_transform = EAT_LINEAR;
 503                                 } else if (!strcasecmp(optval, "atan")) {
 504                                         /* atan-shape transformation;
 505                                          * pumps up low values. */
 506                                         pp->assess_transform = EAT_ATAN;
 507                                 } else if (!strcasecmp(optval, "sigmoid")) {
 508                                         /* Sigmoid transformation
 509                                          * according to assess_sigmb. */
 510                                         pp->assess_transform = EAT_SIGMOID;
 511                                 } else {
 512                                         fprintf(stderr, "playout-elo: Invalid eval mode %s\n", optval);
 513                                         exit(1);
 514                                 }
 515                         } else {
 516                                 fprintf(stderr, "playout-elo: Invalid policy argument %s or missing value\n", optname);
 517                                 exit(1);
 518                         }
 519                 }
 520         }
 521
 522         pc.spat_dict = spatial_dict_init(false);
 523
 524         /* In playouts, we need to operate with much smaller set of features
 525          * in order to keep reasonable speed. */
 526         /* TODO: Configurable. */ /* TODO: Tune. */
 527         pp->choose.pc = FAST_PATTERN_CONFIG;
 528         pp->choose.pc.spat_dict = pc.spat_dict;
 529         char cgammafile[256]; strcpy(stpcpy(cgammafile, gammafile), "f");
 530         pp->choose.fg = features_gamma_init(&pp->choose.pc, cgammafile);
 531         memcpy(pp->choose.ps, PATTERN_SPEC_MATCHFAST, sizeof(pattern_spec));
 532         for (int i = 0; i < FEAT_MAX; i++)
 533                 if ((xspat == 0 && i == FEAT_SPATIAL) || (xspat == 1 && i != FEAT_SPATIAL))
 534                         pp->choose.ps[i] = 0;
 535         if (precise_selfatari) {
 536                 pp->choose.ps[FEAT_SELFATARI] &= ~(1<<PF_SELFATARI_STUPID);
 537                 pp->choose.ps[FEAT_SELFATARI] |= (1<<PF_SELFATARI_SMART);
 538         }
 539         board_gamma_set(b, pp->choose.fg, precise_selfatari);
 540
 541         if (pp->assess_fastpat) {
 542                 pp->assess = pp->choose;
 543                 pp->assess.fg = features_gamma_init(&pp->assess.pc, cgammafile);
 544         } else {
 545                 /* More detailed set of features. */
 546                 pp->assess.pc = pc;
 547                 pp->assess.fg = features_gamma_init(&pp->assess.pc, gammafile);
 548                 memcpy(pp->assess.ps, PATTERN_SPEC_MATCHALL, sizeof(pattern_spec));
 549                 for (int i = 0; i < FEAT_MAX; i++)
 550                         if ((xspat == 0 && i == FEAT_SPATIAL) || (xspat == 1 && i != FEAT_SPATIAL))
 551                                 pp->assess.ps[i] = 0;
 552         }
 553
 554         return p;
 555 }