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
 126 #ifdef BOARD_GAMMA
 127
 128 static void
 129 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)
 130 {
 131 #if 0
 132         struct elo_policy *pp = p->data;
 133         if (pd->total < PROBDIST_EPSILON)
 134                 return;
 135
 136         /* Compare to the manually created distribution. */
 137         /* XXX: This is now broken if callback is used. */
 138
 139         double pdi[board_size2(b)]; memset(pdi, 0, sizeof(pdi));
 140         struct probdist pdx = { .n = board_size2(b), .items = pdi, .total = 0 };
 141         elo_get_probdist(p, &pp->choose, b, to_play, &pdx);
 142         for (int i = 0; i < b->flen; i++) {
 143                 coord_t c = b->f[i];
 144                 if (is_pass(c)) continue;
 145                 // XXX: Hardcoded ignores[] structure
 146                 if (ignores[0] == c) continue;
 147                 double val = pd->items[c];
 148                 if (!is_pass(lc) && coord_is_8adjecent(lc, c, b))
 149                         for (int j = 0; j < lpd->n; j++)
 150                                 if (lpd->coords[j] == c)
 151                                         val = lpd->items[j];
 152                 if (fabs(pdx.items[c] - val) < PROBDIST_EPSILON)
 153                         continue;
 154                 printf("[%s %d] manual %f board %f ", coord2sstr(c, b), b->pat3[c], pdx.items[c], pd->items[c]);
 155                 board_gamma_update(b, c, to_play);
 156                 printf("plainboard %f\n", pd->items[c]);
 157                 assert(0);
 158         }
 159 #endif
 160 }
 161
 162 coord_t
 163 playout_elo_choose(struct playout_policy *p, struct board *b, enum stone to_play)
 164 {
 165         struct elo_policy *pp = p->data;
 166         /* The base board probdist. */
 167         struct probdist *pd = &b->prob[to_play - 1];
 168         double pd_total = pd->total; // precision backup
 169         /* The list of moves we do not consider in pd. */
 170         int ignores[10]; int ignores_n = 0;
 171         /* The list of local moves; we consider these separately. */
 172         struct lprobdist lpd = { .n = 0, .total = 0 };
 173
 174         /* The engine might want to adjust our probdist. */
 175         if (pp->callback)
 176                 pp->callback(pp->callback_data, b, to_play, pd);
 177
 178         /* Make sure ko-prohibited move does not get picked. */
 179         if (!is_pass(b->ko.coord)) {
 180                 assert(b->ko.color == to_play);
 181                 ignores[ignores_n++] = b->ko.coord;
 182                 probdist_mute(pd, b->ko.coord);
 183         }
 184
 185         /* Contiguity detection. */
 186         if (!is_pass(b->last_move.coord)) {
 187                 foreach_8neighbor(b, b->last_move.coord) {
 188                         ignores[ignores_n++] = c;
 189                         probdist_mute(pd, c);
 190
 191                         double val = probdist_one(pd, c) * b->gamma->gamma[FEAT_CONTIGUITY][1];
 192                         lpd.coords[lpd.n] = c;
 193                         lpd.items[lpd.n++] = val;
 194                         lpd.total += val;
 195                 } foreach_8neighbor_end;
 196         }
 197
 198         ignores[ignores_n] = 0;
 199
 200         /* Verify sanity, possibly. */
 201         elo_check_probdist(p, b, to_play, pd, ignores, &lpd, b->last_move.coord);
 202
 203         /* Pick a move. */
 204         coord_t c = pass;
 205         double stab = fast_frandom() * (lpd.total + pd->total);
 206         if (stab < lpd.total - PROBDIST_EPSILON) {
 207                 /* Local probdist. */
 208                 for (int i = 0; i < lpd.n; i++) {
 209                         if (stab <= lpd.items[i]) {
 210                                 c = lpd.coords[i];
 211                                 break;
 212                         }
 213                         stab -= lpd.items[i];
 214                 }
 215                 if (is_pass(c)) {
 216                         fprintf(stderr, "elo: local overstab [%lf]\n", stab);
 217                         abort();
 218                 }
 219
 220         } else if (pd->total >= PROBDIST_EPSILON) {
 221                 /* Global probdist. */
 222                 /* XXX: We re-stab inside. */
 223                 c = probdist_pick(pd, ignores);
 224
 225         } else {
 226                 c = pass;
 227         }
 228
 229         /* Repair the damage. */
 230         if (pp->callback) {
 231                 /* XXX: Do something less horribly inefficient
 232                  * than just recomputing the whole pd. */
 233                 pd->total = 0;
 234                 for (int i = 0; i < b->flen; i++) {
 235                         pd->items[b->f[i]] = 0;
 236                         board_gamma_update(b, b->f[i], to_play);
 237                 }
 238                 assert(fabs(pd->total - pd_total) < PROBDIST_EPSILON);
 239
 240         } else {
 241                 pd->total = pd_total;
 242         }
 243         return c;
 244 }
 245
 246 #else
 247
 248 coord_t
 249 playout_elo_choose(struct playout_policy *p, struct board *b, enum stone to_play)
 250 {
 251         struct elo_policy *pp = p->data;
 252         double pdi[board_size2(b)]; memset(pdi, 0, sizeof(pdi));
 253         struct probdist pd = { .n = board_size2(b), .items = pdi, .total = 0 };
 254         elo_get_probdist(p, &pp->choose, b, to_play, &pd);
 255         if (pp->callback)
 256                 pp->callback(pp->callback_data, b, to_play, &pd);
 257         if (pd.total < PROBDIST_EPSILON)
 258                 return pass;
 259         int ignores[1] = { 0 };
 260         coord_t c = probdist_pick(&pd, ignores);
 261         return c;
 262 }
 263
 264 #endif
 265
 266 void
 267 playout_elo_assess(struct playout_policy *p, struct prior_map *map, int games)
 268 {
 269         struct elo_policy *pp = p->data;
 270         double pdi[board_size2(map->b)]; memset(pdi, 0, sizeof(pdi));
 271         struct probdist pd = { .n = board_size2(map->b), .items = pdi, .total = 0 };
 272
 273         int moves;
 274         moves = elo_get_probdist(p, &pp->assess, map->b, map->to_play, &pd);
 275
 276         /* It is a question how to transform the gamma to won games; we use
 277          * a naive approach currently, but not sure how well it works. */
 278         /* TODO: Try sqrt(p), atan(p)/pi*2. */
 279
 280         for (int f = 0; f < map->b->flen; f++) {
 281                 coord_t c = map->b->f[f];
 282                 if (!map->consider[c])
 283                         continue;
 284                 add_prior_value(map, c, probdist_one(&pd, c) / probdist_total(&pd), games);
 285         }
 286 }
 287
 288 void
 289 playout_elo_done(struct playout_policy *p)
 290 {
 291         struct elo_policy *pp = p->data;
 292         features_gamma_done(pp->choose.fg);
 293         features_gamma_done(pp->assess.fg);
 294 }
 295
 296
 297 void
 298 playout_elo_callback(struct playout_policy *p, playout_elo_callbackp callback, void *data)
 299 {
 300         struct elo_policy *pp = p->data;
 301         pp->callback = callback;
 302         pp->callback_data = data;
 303 }
 304
 305 struct playout_policy *
 306 playout_elo_init(char *arg, struct board *b)
 307 {
 308         struct playout_policy *p = calloc2(1, sizeof(*p));
 309         struct elo_policy *pp = calloc2(1, sizeof(*pp));
 310         p->data = pp;
 311         p->choose = playout_elo_choose;
 312         p->assess = playout_elo_assess;
 313         p->done = playout_elo_done;
 314
 315         const char *gammafile = features_gamma_filename;
 316         /* Some defaults based on the table in Remi Coulom's paper. */
 317         pp->selfatari = 0.06;
 318
 319         struct pattern_config pc = DEFAULT_PATTERN_CONFIG;
 320         int xspat = -1;
 321         bool precise_selfatari = false;
 322
 323         if (arg) {
 324                 char *optspec, *next = arg;
 325                 while (*next) {
 326                         optspec = next;
 327                         next += strcspn(next, ":");
 328                         if (*next) { *next++ = 0; } else { *next = 0; }
 329
 330                         char *optname = optspec;
 331                         char *optval = strchr(optspec, '=');
 332                         if (optval) *optval++ = 0;
 333
 334                         if (!strcasecmp(optname, "selfatari") && optval) {
 335                                 pp->selfatari = atof(optval);
 336                         } else if (!strcasecmp(optname, "precisesa")) {
 337                                 /* Use precise self-atari detection within
 338                                  * fast patterns. */
 339                                 precise_selfatari = !optval || atoi(optval);
 340                         } else if (!strcasecmp(optname, "gammafile") && optval) {
 341                                 /* patterns.gamma by default. We use this,
 342                                  * and need also ${gammafile}f (e.g.
 343                                  * patterns.gammaf) for fast (MC) features. */
 344                                 gammafile = strdup(optval);
 345                         } else if (!strcasecmp(optname, "xspat") && optval) {
 346                                 /* xspat==0: don't match spatial features
 347                                  * xspat==1: match *only* spatial features */
 348                                 xspat = atoi(optval);
 349                         } else {
 350                                 fprintf(stderr, "playout-elo: Invalid policy argument %s or missing value\n", optname);
 351                                 exit(1);
 352                         }
 353                 }
 354         }
 355
 356         pc.spat_dict = spatial_dict_init(false);
 357
 358         pp->assess.pc = pc;
 359         pp->assess.fg = features_gamma_init(&pp->assess.pc, gammafile);
 360         memcpy(pp->assess.ps, PATTERN_SPEC_MATCHALL, sizeof(pattern_spec));
 361         for (int i = 0; i < FEAT_MAX; i++)
 362                 if ((xspat == 0 && i == FEAT_SPATIAL) || (xspat == 1 && i != FEAT_SPATIAL))
 363                         pp->assess.ps[i] = 0;
 364
 365         /* In playouts, we need to operate with much smaller set of features
 366          * in order to keep reasonable speed. */
 367         /* TODO: Configurable. */ /* TODO: Tune. */
 368         pp->choose.pc = FAST_PATTERN_CONFIG;
 369         pp->choose.pc.spat_dict = pc.spat_dict;
 370         char cgammafile[256]; strcpy(stpcpy(cgammafile, gammafile), "f");
 371         pp->choose.fg = features_gamma_init(&pp->choose.pc, cgammafile);
 372         memcpy(pp->choose.ps, PATTERN_SPEC_MATCHFAST, sizeof(pattern_spec));
 373         for (int i = 0; i < FEAT_MAX; i++)
 374                 if ((xspat == 0 && i == FEAT_SPATIAL) || (xspat == 1 && i != FEAT_SPATIAL))
 375                         pp->choose.ps[i] = 0;
 376         if (precise_selfatari)
 377                 pp->choose.ps[FEAT_SELFATARI] = ~(1<<PF_SELFATARI_STUPID);
 378         board_gamma_set(b, pp->choose.fg, precise_selfatari);
 379
 380         return p;
 381 }