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