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 };
  51
  52
  53 /* This is the core of the policy - initializes and constructs the
  54  * probability distribution over the move candidates. */
  55
  56 int
  57 elo_get_probdist(struct playout_policy *p, struct patternset *ps, struct board *b, enum stone to_play, struct probdist *pd)
  58 {
  59         //struct elo_policy *pp = p->data;
  60         int moves = 0;
  61
  62         probdist_init(pd, board_size2(b));
  63
  64         /* First, assign per-point probabilities. */
  65
  66         for (int f = 0; f < b->flen; f++) {
  67                 struct move m = { .coord = b->f[f], .color = to_play };
  68
  69                 /* Skip pass (for now)? */
  70                 if (is_pass(m.coord))
  71                         continue;
  72                 //fprintf(stderr, "<%d> %s\n", f, coord2sstr(m.coord, b));
  73
  74                 /* Skip invalid moves. */
  75                 if (!board_is_valid_move(b, &m))
  76                         continue;
  77
  78                 /* We shall never fill our own single-point eyes. */
  79                 /* XXX: In some rare situations, this prunes the best move:
  80                  * Bulk-five nakade with eye at 1-1 point. */
  81                 if (board_is_one_point_eye(b, &m.coord, to_play)) {
  82                         continue;
  83                 }
  84
  85                 moves++;
  86                 /* Each valid move starts with gamma 1. */
  87                 probdist_add(pd, m.coord, 1.f);
  88
  89                 /* Some easy features: */
  90                 /* XXX: We just disable them for now since we call the
  91                  * pattern matcher; you need the gammas file. */
  92 #if 0
  93                 if (is_bad_selfatari(b, to_play, m.coord))
  94                         probdist_mul(pd, m.coord, pp->selfatari);
  95 #endif
  96
  97                 /* Match pattern features: */
  98                 struct pattern p;
  99                 pattern_match(&ps->pc, ps->ps, &p, b, &m);
 100                 for (int i = 0; i < p.n; i++) {
 101                         /* Multiply together gammas of all pattern features. */
 102                         float gamma = feature_gamma(ps->fg, &p.f[i], NULL);
 103                         //char buf[256] = ""; feature2str(buf, &p.f[i]);
 104                         //fprintf(stderr, "<%d> %s feat %s gamma %f\n", f, coord2sstr(m.coord, b), buf, gamma);
 105                         probdist_mul(pd, m.coord, gamma);
 106                 }
 107                 //fprintf(stderr, "<%d> %s %f\n", f, coord2sstr(m.coord, b), pd->moves[m.coord]);
 108         }
 109
 110         return moves;
 111 }
 112
 113
 114 coord_t
 115 playout_elo_choose(struct playout_policy *p, struct board *b, enum stone to_play)
 116 {
 117         struct elo_policy *pp = p->data;
 118         struct probdist pd;
 119         elo_get_probdist(p, &pp->choose, b, to_play, &pd);
 120         coord_t c = probdist_pick(&pd);
 121         probdist_done(&pd);
 122         return c;
 123 }
 124
 125 void
 126 playout_elo_assess(struct playout_policy *p, struct prior_map *map, int games)
 127 {
 128         struct elo_policy *pp = p->data;
 129         struct probdist pd;
 130         int moves;
 131
 132         moves = elo_get_probdist(p, &pp->assess, map->b, map->to_play, &pd);
 133
 134         /* It is a question how to transform the gamma to won games; we use
 135          * a naive approach currently, but not sure how well it works. */
 136         /* TODO: Try sqrt(p), atan(p)/pi*2. */
 137
 138         for (int f = 0; f < map->b->flen; f++) {
 139                 coord_t c = map->b->f[f];
 140                 if (!map->consider[c])
 141                         continue;
 142                 add_prior_value(map, c, pd.moves[c] / pd.total, games);
 143         }
 144
 145         probdist_done(&pd);
 146 }
 147
 148
 149 struct playout_policy *
 150 playout_elo_init(char *arg)
 151 {
 152         struct playout_policy *p = calloc(1, sizeof(*p));
 153         struct elo_policy *pp = calloc(1, sizeof(*pp));
 154         p->data = pp;
 155         p->choose = playout_elo_choose;
 156         p->assess = playout_elo_assess;
 157
 158         const char *gammafile = features_gamma_filename;
 159         /* Some defaults based on the table in Remi Coulom's paper. */
 160         pp->selfatari = 0.06;
 161
 162         struct pattern_config pc = DEFAULT_PATTERN_CONFIG;
 163
 164         if (arg) {
 165                 char *optspec, *next = arg;
 166                 while (*next) {
 167                         optspec = next;
 168                         next += strcspn(next, ":");
 169                         if (*next) { *next++ = 0; } else { *next = 0; }
 170
 171                         char *optname = optspec;
 172                         char *optval = strchr(optspec, '=');
 173                         if (optval) *optval++ = 0;
 174
 175                         if (!strcasecmp(optname, "selfatari") && optval) {
 176                                 pp->selfatari = atof(optval);
 177                         } else if (!strcasecmp(optname, "gammafile") && optval) {
 178                                 /* patterns.gamma by default. We use this,
 179                                  * and need also ${gammafile}f (e.g.
 180                                  * patterns.gammaf) for fast (MC) features. */
 181                                 gammafile = strdup(optval);
 182                         } else {
 183                                 fprintf(stderr, "playout-elo: Invalid policy argument %s or missing value\n", optname);
 184                                 exit(1);
 185                         }
 186                 }
 187         }
 188
 189         pc.spat_dict = spatial_dict_init(false);
 190
 191         pp->assess.pc = pc;
 192         pp->assess.fg = features_gamma_init(&pp->assess.pc, gammafile);
 193         memcpy(pp->assess.ps, PATTERN_SPEC_MATCHALL, sizeof(pattern_spec));
 194
 195         /* In playouts, we need to operate with much smaller set of features
 196          * in order to keep reasonable speed. */
 197         /* TODO: Configurable. */ /* TODO: Tune. */
 198         pp->choose.pc = FAST_PATTERN_CONFIG;
 199         pp->choose.pc.spat_dict = pc.spat_dict;
 200         char cgammafile[256]; strcpy(stpcpy(cgammafile, gammafile), "f");
 201         pp->choose.fg = features_gamma_init(&pp->choose.pc, cgammafile);
 202         memcpy(pp->choose.ps, PATTERN_SPEC_MATCHFAST, sizeof(pattern_spec));
 203
 204         return p;
 205 }