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