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, f, 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, f, g);
 111                 //fprintf(stderr, "<%d> %s %f (E %f)\n", f, coord2sstr(m.coord, b), probdist_one(pd, f), pd->items[f]);
 112         }
 113
 114         return moves;
 115 }
 116
 117
 118 coord_t
 119 playout_elo_choose(struct playout_policy *p, struct board *b, enum stone to_play)
 120 {
 121         struct elo_policy *pp = p->data;
 122 #ifdef BOARD_GAMMA
 123         struct probdist *pd = &b->prob[to_play - 1];
 124         /* Make sure ko-prohibited move does not get picked. */
 125         if (!is_pass(b->ko.coord)) {
 126                 assert(b->ko.color == to_play);
 127                 probdist_set(pd, b->ko.coord, 0);
 128         }
 129         /* Contiguity detection. */
 130         if (!is_pass(b->last_move.coord)) {
 131                 foreach_8neighbor(b, b->last_move.coord) {
 132                         probdist_set(pd, c, pd->items[c] * b->gamma->gamma[FEAT_CONTIGUITY][1]);
 133                 } foreach_8neighbor_end;
 134         }
 135 #if 0
 136         /* Compare to the manually created distribution. */
 137         if (pd->total >= PROBDIST_EPSILON) {
 138                 double pdi[b->flen]; memset(pdi, 0, sizeof(pdi));
 139                 struct probdist pdx = { .n = b->flen, .items = pdi, .total = 0 };
 140                 elo_get_probdist(p, &pp->choose, b, to_play, &pdx);
 141                 for (int i = 0; i < pdx.n; i++) {
 142                         if (is_pass(b->f[i])) continue;
 143                         if (fabs(pdx.items[i] - pd->items[b->f[i]]) >= PROBDIST_EPSILON) {
 144                                 printf("[%s %d] manual %f board %f ", coord2sstr(b->f[i], b), b->pat3[b->f[i]], pdx.items[i], pd->items[b->f[i]]);
 145                                 board_gamma_update(b, b->f[i], to_play);
 146                                 printf("plainboard %f\n", pd->items[b->f[i]]);
 147                                 assert(0);
 148                         }
 149                 }
 150         }
 151 #endif
 152         /* The engine might want to adjust our probdist. */
 153         if (pp->callback)
 154                 pp->callback(pp->callback_data, b, to_play, pd);
 155         /* Pick a move. */
 156         coord_t c = pd->total >= PROBDIST_EPSILON ? probdist_pick(pd) : pass;
 157         /* Repair the damage. */
 158         if (pp->callback) {
 159                 /* XXX: Do something less horribly inefficient
 160                  * than just recomputing the whole pd. */
 161                 pd->total = 0;
 162                 for (int i = 0; i < b->flen; i++) {
 163                         pd->items[b->f[i]] = 0;
 164                         board_gamma_update(b, b->f[i], to_play);
 165                 }
 166         }
 167         if (!is_pass(b->ko.coord))
 168                 board_gamma_update(b, b->ko.coord, to_play);
 169         if (!is_pass(b->last_move.coord)) {
 170                 foreach_8neighbor(b, b->last_move.coord) {
 171                         board_gamma_update(b, c, to_play);
 172                 } foreach_8neighbor_end;
 173         }
 174         return c;
 175
 176 #else
 177         double pdi[b->flen]; memset(pdi, 0, sizeof(pdi));
 178         struct probdist pd = { .n = b->flen, .items = pdi, .total = 0 };
 179         elo_get_probdist(p, &pp->choose, b, to_play, &pd);
 180         if (pp->callback)
 181                 pp->callback(pp->callback_data, b, to_play, &pd);
 182         if (pd.total < PROBDIST_EPSILON)
 183                 return pass;
 184         int f = probdist_pick(&pd);
 185         return b->f[f];
 186 #endif
 187 }
 188
 189 void
 190 playout_elo_assess(struct playout_policy *p, struct prior_map *map, int games)
 191 {
 192         struct elo_policy *pp = p->data;
 193         double pdi[map->b->flen]; memset(pdi, 0, sizeof(pdi));
 194         struct probdist pd = { .n = map->b->flen, .items = pdi, .total = 0 };
 195
 196         int moves;
 197         moves = elo_get_probdist(p, &pp->assess, map->b, map->to_play, &pd);
 198
 199         /* It is a question how to transform the gamma to won games; we use
 200          * a naive approach currently, but not sure how well it works. */
 201         /* TODO: Try sqrt(p), atan(p)/pi*2. */
 202
 203         for (int f = 0; f < map->b->flen; f++) {
 204                 coord_t c = map->b->f[f];
 205                 if (!map->consider[c])
 206                         continue;
 207                 add_prior_value(map, c, probdist_one(&pd, f) / probdist_total(&pd), games);
 208         }
 209 }
 210
 211 void
 212 playout_elo_done(struct playout_policy *p)
 213 {
 214         struct elo_policy *pp = p->data;
 215         features_gamma_done(pp->choose.fg);
 216         features_gamma_done(pp->assess.fg);
 217 }
 218
 219
 220 void
 221 playout_elo_callback(struct playout_policy *p, playout_elo_callbackp callback, void *data)
 222 {
 223         struct elo_policy *pp = p->data;
 224         pp->callback = callback;
 225         pp->callback_data = data;
 226 }
 227
 228 struct playout_policy *
 229 playout_elo_init(char *arg, struct board *b)
 230 {
 231         struct playout_policy *p = calloc2(1, sizeof(*p));
 232         struct elo_policy *pp = calloc2(1, sizeof(*pp));
 233         p->data = pp;
 234         p->choose = playout_elo_choose;
 235         p->assess = playout_elo_assess;
 236         p->done = playout_elo_done;
 237
 238         const char *gammafile = features_gamma_filename;
 239         /* Some defaults based on the table in Remi Coulom's paper. */
 240         pp->selfatari = 0.06;
 241
 242         struct pattern_config pc = DEFAULT_PATTERN_CONFIG;
 243         int xspat = -1;
 244         bool precise_selfatari = false;
 245
 246         if (arg) {
 247                 char *optspec, *next = arg;
 248                 while (*next) {
 249                         optspec = next;
 250                         next += strcspn(next, ":");
 251                         if (*next) { *next++ = 0; } else { *next = 0; }
 252
 253                         char *optname = optspec;
 254                         char *optval = strchr(optspec, '=');
 255                         if (optval) *optval++ = 0;
 256
 257                         if (!strcasecmp(optname, "selfatari") && optval) {
 258                                 pp->selfatari = atof(optval);
 259                         } else if (!strcasecmp(optname, "precisesa")) {
 260                                 /* Use precise self-atari detection within
 261                                  * fast patterns. */
 262                                 precise_selfatari = !optval || atoi(optval);
 263                         } else if (!strcasecmp(optname, "gammafile") && optval) {
 264                                 /* patterns.gamma by default. We use this,
 265                                  * and need also ${gammafile}f (e.g.
 266                                  * patterns.gammaf) for fast (MC) features. */
 267                                 gammafile = strdup(optval);
 268                         } else if (!strcasecmp(optname, "xspat") && optval) {
 269                                 /* xspat==0: don't match spatial features
 270                                  * xspat==1: match *only* spatial features */
 271                                 xspat = atoi(optval);
 272                         } else {
 273                                 fprintf(stderr, "playout-elo: Invalid policy argument %s or missing value\n", optname);
 274                                 exit(1);
 275                         }
 276                 }
 277         }
 278
 279         pc.spat_dict = spatial_dict_init(false);
 280
 281         pp->assess.pc = pc;
 282         pp->assess.fg = features_gamma_init(&pp->assess.pc, gammafile);
 283         memcpy(pp->assess.ps, PATTERN_SPEC_MATCHALL, sizeof(pattern_spec));
 284         for (int i = 0; i < FEAT_MAX; i++)
 285                 if ((xspat == 0 && i == FEAT_SPATIAL) || (xspat == 1 && i != FEAT_SPATIAL))
 286                         pp->assess.ps[i] = 0;
 287
 288         /* In playouts, we need to operate with much smaller set of features
 289          * in order to keep reasonable speed. */
 290         /* TODO: Configurable. */ /* TODO: Tune. */
 291         pp->choose.pc = FAST_PATTERN_CONFIG;
 292         pp->choose.pc.spat_dict = pc.spat_dict;
 293         char cgammafile[256]; strcpy(stpcpy(cgammafile, gammafile), "f");
 294         pp->choose.fg = features_gamma_init(&pp->choose.pc, cgammafile);
 295         memcpy(pp->choose.ps, PATTERN_SPEC_MATCHFAST, sizeof(pattern_spec));
 296         for (int i = 0; i < FEAT_MAX; i++)
 297                 if ((xspat == 0 && i == FEAT_SPATIAL) || (xspat == 1 && i != FEAT_SPATIAL))
 298                         pp->choose.ps[i] = 0;
 299         if (precise_selfatari)
 300                 pp->choose.ps[FEAT_SELFATARI] = ~(1<<PF_SELFATARI_STUPID);
 301         board_gamma_set(b, pp->choose.fg, precise_selfatari);
 302
 303         return p;
 304 }