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 "pattern.h"
  29 #include "patternsp.h"
  30 #include "playout.h"
  31 #include "playout/elo.h"
  32 #include "random.h"
  33 #include "tactics.h"
  34 #include "uct/prior.h"
  35
  36 #define PLDEBUGL(n) DEBUGL_(p->debug_level, n)
  37
  38
  39 /* Note that the context can be shared by multiple threads! */
  40
  41 struct patternset {
  42         pattern_spec ps;
  43         struct pattern_config pc;
  44         struct features_gamma *fg;
  45 };
  46
  47 struct elo_policy {
  48         float selfatari;
  49         struct patternset choose, assess;
  50         playout_elo_callbackp callback; void *callback_data;
  51 };
  52
  53
  54 /* This is the core of the policy - initializes and constructs the
  55  * probability distribution over the move candidates. */
  56
  57 int
  58 elo_get_probdist(struct playout_policy *p, struct patternset *ps, struct board *b, enum stone to_play, struct probdist *pd)
  59 {
  60         //struct elo_policy *pp = p->data;
  61         int moves = 0;
  62
  63         /* First, assign per-point probabilities. */
  64
  65         for (int f = 0; f < b->flen; f++) {
  66                 struct move m = { .coord = b->f[f], .color = to_play };
  67
  68                 /* Skip pass (for now)? */
  69                 if (is_pass(m.coord)) {
  70 skip_move:
  71                         probdist_set(pd, m.coord, 0);
  72                         continue;
  73                 }
  74                 //fprintf(stderr, "<%d> %s\n", f, coord2sstr(m.coord, b));
  75
  76                 /* Skip invalid moves. */
  77                 if (!board_is_valid_move(b, &m))
  78                         goto skip_move;
  79
  80                 /* We shall never fill our own single-point eyes. */
  81                 /* XXX: In some rare situations, this prunes the best move:
  82                  * Bulk-five nakade with eye at 1-1 point. */
  83                 if (board_is_one_point_eye(b, m.coord, to_play)) {
  84                         goto skip_move;
  85                 }
  86
  87                 moves++;
  88                 /* Each valid move starts with gamma 1. */
  89                 double g = 1.f;
  90
  91                 /* Some easy features: */
  92                 /* XXX: We just disable them for now since we call the
  93                  * pattern matcher; you need the gammas file. */
  94 #if 0
  95                 if (is_bad_selfatari(b, to_play, m.coord))
  96                         g *= pp->selfatari;
  97 #endif
  98
  99                 /* Match pattern features: */
 100                 struct pattern p;
 101                 pattern_match(&ps->pc, ps->ps, &p, b, &m);
 102                 for (int i = 0; i < p.n; i++) {
 103                         /* Multiply together gammas of all pattern features. */
 104                         double gamma = feature_gamma(ps->fg, &p.f[i], NULL);
 105                         //char buf[256] = ""; feature2str(buf, &p.f[i]);
 106                         //fprintf(stderr, "<%d> %s feat %s gamma %f\n", f, coord2sstr(m.coord, b), buf, gamma);
 107                         g *= gamma;
 108                 }
 109
 110                 probdist_set(pd, m.coord, g);
 111                 //fprintf(stderr, "<%d> %s %f (E %f)\n", f, coord2sstr(m.coord, b), probdist_one(pd, m.coord), pd->items[f]);
 112         }
 113
 114         return moves;
 115 }
 116
 117
 118 struct lprobdist {
 119         int n;
 120 #define LPD_MAX 8
 121         coord_t coords[LPD_MAX];
 122         double items[LPD_MAX];
 123         double total;
 124
 125         /* Backups of original totals for restoring. */
 126         double btotal;
 127         double browtotals_v[10];
 128         int browtotals_i[10];
 129         int browtotals_n;
 130 };
 131
 132 #ifdef BOARD_GAMMA
 133
 134 static void
 135 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)
 136 {
 137 #if 0
 138         struct elo_policy *pp = p->data;
 139         if (pd->total < PROBDIST_EPSILON)
 140                 return;
 141
 142         /* Compare to the manually created distribution. */
 143         /* XXX: This is now broken if callback is used. */
 144
 145         double pdi[board_size2(b)]; memset(pdi, 0, sizeof(pdi));
 146         double pdr[board_size(b)]; memset(pdr, 0, sizeof(pdr));
 147         struct probdist pdx = { .n = board_size2(b), .n1 = board_size(b), .items = pdi, .rowtotals = pdr, .total = 0 };
 148         elo_get_probdist(p, &pp->choose, b, to_play, &pdx);
 149         for (int i = 0; i < b->flen; i++) {
 150                 coord_t c = b->f[i];
 151                 if (is_pass(c)) continue;
 152                 // XXX: Hardcoded ignores[] structure
 153                 if (ignores[0] == c) continue;
 154                 double val = pd->items[c];
 155                 if (!is_pass(lc) && coord_is_8adjecent(lc, c, b))
 156                         for (int j = 0; j < lpd->n; j++)
 157                                 if (lpd->coords[j] == c)
 158                                         val = lpd->items[j];
 159                 if (fabs(pdx.items[c] - val) < PROBDIST_EPSILON)
 160                         continue;
 161                 printf("[%s %d] manual %f board %f ", coord2sstr(c, b), b->pat3[c], pdx.items[c], pd->items[c]);
 162                 board_gamma_update(b, c, to_play);
 163                 printf("plainboard %f\n", pd->items[c]);
 164                 assert(0);
 165         }
 166 #endif
 167 }
 168
 169 coord_t
 170 playout_elo_choose(struct playout_policy *p, struct board *b, enum stone to_play)
 171 {
 172         struct elo_policy *pp = p->data;
 173         /* The base board probdist. */
 174         struct probdist *pd = &b->prob[to_play - 1];
 175         /* The list of moves we do not consider in pd. */
 176         int ignores[10]; int ignores_n = 0;
 177         /* The list of local moves; we consider these separately. */
 178         struct lprobdist lpd = { .n = 0, .total = 0, .btotal = pd->total, .browtotals_n = 0 };
 179
 180         /* The engine might want to adjust our probdist. */
 181         if (pp->callback)
 182                 pp->callback(pp->callback_data, b, to_play, pd);
 183
 184 #define ignore_move(c_) do { \
 185         ignores[ignores_n++] = c_; \
 186         int rowi = c_ / pd->n1; \
 187         lpd.browtotals_i[lpd.browtotals_n] = rowi; \
 188         lpd.browtotals_v[lpd.browtotals_n++] = pd->rowtotals[rowi]; \
 189         probdist_mute(pd, c_); \
 190 } while (0)
 191
 192         /* Make sure ko-prohibited move does not get picked. */
 193         if (!is_pass(b->ko.coord)) {
 194                 assert(b->ko.color == to_play);
 195                 ignore_move(b->ko.coord);
 196         }
 197
 198         /* Contiguity detection. */
 199         if (!is_pass(b->last_move.coord)) {
 200                 foreach_8neighbor(b, b->last_move.coord) {
 201                         ignore_move(c);
 202
 203                         double val = probdist_one(pd, c) * b->gamma->gamma[FEAT_CONTIGUITY][1];
 204                         lpd.coords[lpd.n] = c;
 205                         lpd.items[lpd.n++] = val;
 206                         lpd.total += val;
 207                 } foreach_8neighbor_end;
 208         }
 209
 210         ignores[ignores_n] = pass;
 211
 212         /* Verify sanity, possibly. */
 213         elo_check_probdist(p, b, to_play, pd, ignores, &lpd, b->last_move.coord);
 214
 215         /* Pick a move. */
 216         coord_t c = pass;
 217         double stab = fast_frandom() * (lpd.total + pd->total);
 218         if (stab < lpd.total - PROBDIST_EPSILON) {
 219                 /* Local probdist. */
 220                 for (int i = 0; i < lpd.n; i++) {
 221                         if (stab <= lpd.items[i]) {
 222                                 c = lpd.coords[i];
 223                                 break;
 224                         }
 225                         stab -= lpd.items[i];
 226                 }
 227                 if (is_pass(c)) {
 228                         fprintf(stderr, "elo: local overstab [%lf]\n", stab);
 229                         abort();
 230                 }
 231
 232         } else if (pd->total >= PROBDIST_EPSILON) {
 233                 /* Global probdist. */
 234                 /* XXX: We re-stab inside. */
 235                 c = probdist_pick(pd, ignores);
 236
 237         } else {
 238                 c = pass;
 239         }
 240
 241         /* Repair the damage. */
 242         if (pp->callback) {
 243                 /* XXX: Do something less horribly inefficient
 244                  * than just recomputing the whole pd. */
 245                 pd->total = 0;
 246                 for (int i = 0; i < pd->n / pd->n1; i++)
 247                         pd->rowtotals[i] = 0;
 248                 for (int i = 0; i < b->flen; i++) {
 249                         pd->items[b->f[i]] = 0;
 250                         board_gamma_update(b, b->f[i], to_play);
 251                 }
 252                 assert(fabs(pd->total - lpd.btotal) < PROBDIST_EPSILON);
 253
 254         } else {
 255                 pd->total = lpd.btotal;
 256                 /* If we touched a row multiple times (and we sure will),
 257                  * the latter value is obsolete; but since we go through
 258                  * the backups in reverse order, all is good. */
 259                 for (int j = lpd.browtotals_n - 1; j >= 0; j--)
 260                         pd->rowtotals[lpd.browtotals_i[j]] = lpd.browtotals_v[j];
 261         }
 262         return c;
 263 }
 264
 265 #else
 266
 267 coord_t
 268 playout_elo_choose(struct playout_policy *p, struct board *b, enum stone to_play)
 269 {
 270         struct elo_policy *pp = p->data;
 271         double pdi[board_size2(b)]; memset(pdi, 0, sizeof(pdi));
 272         double pdr[board_size(b)]; memset(pdr, 0, sizeof(pdr));
 273         struct probdist pd = { .n = board_size2(b), .n1 = board_size(b), .items = pdi, .rowtotals = pdr, .total = 0 };
 274         elo_get_probdist(p, &pp->choose, b, to_play, &pd);
 275         if (pp->callback)
 276                 pp->callback(pp->callback_data, b, to_play, &pd);
 277         if (pd.total < PROBDIST_EPSILON)
 278                 return pass;
 279         int ignores[1] = { pass };
 280         coord_t c = probdist_pick(&pd, ignores);
 281         return c;
 282 }
 283
 284 #endif
 285
 286 void
 287 playout_elo_assess(struct playout_policy *p, struct prior_map *map, int games)
 288 {
 289         struct elo_policy *pp = p->data;
 290         double pdi[board_size2(map->b)]; memset(pdi, 0, sizeof(pdi));
 291         double pdr[board_size(map->b)]; memset(pdr, 0, sizeof(pdr));
 292         struct probdist pd = { .n = board_size2(map->b), .n1 = board_size(map->b), .items = pdi, .rowtotals = pdr, .total = 0 };
 293
 294         int moves;
 295         moves = elo_get_probdist(p, &pp->assess, map->b, map->to_play, &pd);
 296
 297         /* It is a question how to transform the gamma to won games; we use
 298          * a naive approach currently, but not sure how well it works. */
 299         /* TODO: Try sqrt(p), atan(p)/pi*2. */
 300
 301         for (int f = 0; f < map->b->flen; f++) {
 302                 coord_t c = map->b->f[f];
 303                 if (!map->consider[c])
 304                         continue;
 305                 add_prior_value(map, c, probdist_one(&pd, c) / probdist_total(&pd), games);
 306         }
 307 }
 308
 309 void
 310 playout_elo_done(struct playout_policy *p)
 311 {
 312         struct elo_policy *pp = p->data;
 313         features_gamma_done(pp->choose.fg);
 314         features_gamma_done(pp->assess.fg);
 315 }
 316
 317
 318 void
 319 playout_elo_callback(struct playout_policy *p, playout_elo_callbackp callback, void *data)
 320 {
 321         struct elo_policy *pp = p->data;
 322         pp->callback = callback;
 323         pp->callback_data = data;
 324 }
 325
 326 struct playout_policy *
 327 playout_elo_init(char *arg, struct board *b)
 328 {
 329         struct playout_policy *p = calloc2(1, sizeof(*p));
 330         struct elo_policy *pp = calloc2(1, sizeof(*pp));
 331         p->data = pp;
 332         p->choose = playout_elo_choose;
 333         p->assess = playout_elo_assess;
 334         p->done = playout_elo_done;
 335
 336         const char *gammafile = features_gamma_filename;
 337         /* Some defaults based on the table in Remi Coulom's paper. */
 338         pp->selfatari = 0.06;
 339
 340         struct pattern_config pc = DEFAULT_PATTERN_CONFIG;
 341         int xspat = -1;
 342         bool precise_selfatari = false;
 343
 344         if (arg) {
 345                 char *optspec, *next = arg;
 346                 while (*next) {
 347                         optspec = next;
 348                         next += strcspn(next, ":");
 349                         if (*next) { *next++ = 0; } else { *next = 0; }
 350
 351                         char *optname = optspec;
 352                         char *optval = strchr(optspec, '=');
 353                         if (optval) *optval++ = 0;
 354
 355                         if (!strcasecmp(optname, "selfatari") && optval) {
 356                                 pp->selfatari = atof(optval);
 357                         } else if (!strcasecmp(optname, "precisesa")) {
 358                                 /* Use precise self-atari detection within
 359                                  * fast patterns. */
 360                                 precise_selfatari = !optval || atoi(optval);
 361                         } else if (!strcasecmp(optname, "gammafile") && optval) {
 362                                 /* patterns.gamma by default. We use this,
 363                                  * and need also ${gammafile}f (e.g.
 364                                  * patterns.gammaf) for fast (MC) features. */
 365                                 gammafile = strdup(optval);
 366                         } else if (!strcasecmp(optname, "xspat") && optval) {
 367                                 /* xspat==0: don't match spatial features
 368                                  * xspat==1: match *only* spatial features */
 369                                 xspat = atoi(optval);
 370                         } else {
 371                                 fprintf(stderr, "playout-elo: Invalid policy argument %s or missing value\n", optname);
 372                                 exit(1);
 373                         }
 374                 }
 375         }
 376
 377         pc.spat_dict = spatial_dict_init(false);
 378
 379         pp->assess.pc = pc;
 380         pp->assess.fg = features_gamma_init(&pp->assess.pc, gammafile);
 381         memcpy(pp->assess.ps, PATTERN_SPEC_MATCHALL, sizeof(pattern_spec));
 382         for (int i = 0; i < FEAT_MAX; i++)
 383                 if ((xspat == 0 && i == FEAT_SPATIAL) || (xspat == 1 && i != FEAT_SPATIAL))
 384                         pp->assess.ps[i] = 0;
 385
 386         /* In playouts, we need to operate with much smaller set of features
 387          * in order to keep reasonable speed. */
 388         /* TODO: Configurable. */ /* TODO: Tune. */
 389         pp->choose.pc = FAST_PATTERN_CONFIG;
 390         pp->choose.pc.spat_dict = pc.spat_dict;
 391         char cgammafile[256]; strcpy(stpcpy(cgammafile, gammafile), "f");
 392         pp->choose.fg = features_gamma_init(&pp->choose.pc, cgammafile);
 393         memcpy(pp->choose.ps, PATTERN_SPEC_MATCHFAST, sizeof(pattern_spec));
 394         for (int i = 0; i < FEAT_MAX; i++)
 395                 if ((xspat == 0 && i == FEAT_SPATIAL) || (xspat == 1 && i != FEAT_SPATIAL))
 396                         pp->choose.ps[i] = 0;
 397         if (precise_selfatari)
 398                 pp->choose.ps[FEAT_SELFATARI] = ~(1<<PF_SELFATARI_STUPID);
 399         board_gamma_set(b, pp->choose.fg, precise_selfatari);
 400
 401         return p;
 402 }