uct/dynkomi.c

   1 #define DEBUG
   2 #include <assert.h>
   3 #include <math.h>
   4 #include <stdio.h>
   5 #include <stdlib.h>
   6 #include <string.h>
   7
   8 #include "board.h"
   9 #include "debug.h"
  10 #include "tactics.h"
  11 #include "uct/dynkomi.h"
  12 #include "uct/internal.h"
  13 #include "uct/tree.h"
  14
  15
  16 static void
  17 generic_done(struct uct_dynkomi *d)
  18 {
  19         if (d->data) free(d->data);
  20         free(d);
  21 }
  22
  23
  24 /* NONE dynkomi strategy - never fiddle with komi values. */
  25
  26 struct uct_dynkomi *
  27 uct_dynkomi_init_none(struct uct *u, char *arg, struct board *b)
  28 {
  29         struct uct_dynkomi *d = calloc2(1, sizeof(*d));
  30         d->uct = u;
  31         d->permove = NULL;
  32         d->persim = NULL;
  33         d->done = generic_done;
  34         d->data = NULL;
  35
  36         if (arg) {
  37                 fprintf(stderr, "uct: Dynkomi method none accepts no arguments\n");
  38                 exit(1);
  39         }
  40
  41         return d;
  42 }
  43
  44
  45 /* LINEAR dynkomi strategy - Linearly Decreasing Handicap Compensation. */
  46 /* At move 0, we impose extra komi of handicap_count*handicap_value, then
  47  * we linearly decrease this extra komi throughout the game down to 0
  48  * at @moves moves. */
  49
  50 struct dynkomi_linear {
  51         int handicap_value;
  52         int moves;
  53         bool rootbased;
  54 };
  55
  56 static float
  57 linear_permove(struct uct_dynkomi *d, struct board *b, struct tree *tree)
  58 {
  59         struct dynkomi_linear *l = d->data;
  60         if (b->moves >= l->moves)
  61                 return 0;
  62
  63         float base_komi = board_effective_handicap(b, l->handicap_value);
  64         float extra_komi = base_komi * (l->moves - b->moves) / l->moves;
  65         return extra_komi;
  66 }
  67
  68 static float
  69 linear_persim(struct uct_dynkomi *d, struct board *b, struct tree *tree, struct tree_node *node)
  70 {
  71         struct dynkomi_linear *l = d->data;
  72         if (l->rootbased)
  73                 return tree->extra_komi;
  74         /* We don't reuse computed value from tree->extra_komi,
  75          * since we want to use value correct for this node depth.
  76          * This also means the values will stay correct after
  77          * node promotion. */
  78         return linear_permove(d, b, tree);
  79 }
  80
  81 struct uct_dynkomi *
  82 uct_dynkomi_init_linear(struct uct *u, char *arg, struct board *b)
  83 {
  84         struct uct_dynkomi *d = calloc2(1, sizeof(*d));
  85         d->uct = u;
  86         d->permove = linear_permove;
  87         d->persim = linear_persim;
  88         d->done = generic_done;
  89
  90         struct dynkomi_linear *l = calloc2(1, sizeof(*l));
  91         d->data = l;
  92
  93         if (board_size(b) - 2 >= 19)
  94                 l->moves = 200;
  95         l->handicap_value = 7;
  96
  97         if (arg) {
  98                 char *optspec, *next = arg;
  99                 while (*next) {
 100                         optspec = next;
 101                         next += strcspn(next, ":");
 102                         if (*next) { *next++ = 0; } else { *next = 0; }
 103
 104                         char *optname = optspec;
 105                         char *optval = strchr(optspec, '=');
 106                         if (optval) *optval++ = 0;
 107
 108                         if (!strcasecmp(optname, "moves") && optval) {
 109                                 /* Dynamic komi in handicap game; linearly
 110                                  * decreases to basic settings until move
 111                                  * #optval. */
 112                                 l->moves = atoi(optval);
 113                         } else if (!strcasecmp(optname, "handicap_value") && optval) {
 114                                 /* Point value of single handicap stone,
 115                                  * for dynkomi computation. */
 116                                 l->handicap_value = atoi(optval);
 117                         } else if (!strcasecmp(optname, "rootbased")) {
 118                                 /* If set, the extra komi applied will be
 119                                  * the same for all simulations within a move,
 120                                  * instead of being same for all simulations
 121                                  * within the tree node. */
 122                                 l->rootbased = !optval || atoi(optval);
 123                         } else {
 124                                 fprintf(stderr, "uct: Invalid dynkomi argument %s or missing value\n", optname);
 125                                 exit(1);
 126                         }
 127                 }
 128         }
 129
 130         return d;
 131 }
 132
 133
 134 /* ADAPTIVE dynkomi strategy - Adaptive Situational Compensation */
 135 /* We adapt the komi based on current situation:
 136  * (i) score-based: We maintain the average score outcome of our
 137  * games and adjust the komi by a fractional step towards the expected
 138  * score;
 139  * (ii) value-based: While winrate is above given threshold, adjust
 140  * the komi by a fixed step in the appropriate direction.
 141  * These adjustments can be
 142  * (a) Move-stepped, new extra komi value is always set only at the
 143  * beginning of the tree search for next move;
 144  * (b) Continuous, new extra komi value is periodically re-determined
 145  * and adjusted throughout a single tree search. */
 146
 147 struct dynkomi_adaptive {
 148         /* Do not take measured average score into regard for
 149          * first @lead_moves - the variance is just too much.
 150          * (Instead, we consider the handicap-based komi provided
 151          * by linear dynkomi.) */
 152         int lead_moves;
 153         /* Maximum komi to pretend the opponent to give. */
 154         float max_losing_komi;
 155         float (*indicator)(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color);
 156
 157         /* Value-based adaptation. */
 158         float zone_red, zone_green;
 159         int score_step;
 160         bool use_komi_latch;
 161         float komi_latch; // runtime, not configuration
 162
 163         /* Score-based adaptation. */
 164         float (*adapter)(struct uct_dynkomi *d, struct board *b);
 165         float adapt_base; // [0,1)
 166         /* Sigmoid adaptation rate parameter; see below for details. */
 167         float adapt_phase; // [0,1]
 168         float adapt_rate; // [1,infty)
 169         bool adapt_aport; // alternative game portion determination
 170         /* Linear adaptation rate parameter. */
 171         int adapt_moves;
 172         float adapt_dir; // [-1,1]
 173 };
 174 #define TRUSTWORTHY_KOMI_PLAYOUTS 200
 175
 176 static float
 177 adapter_sigmoid(struct uct_dynkomi *d, struct board *b)
 178 {
 179         struct dynkomi_adaptive *a = d->data;
 180         /* Figure out how much to adjust the komi based on the game
 181          * stage. The adaptation rate is 0 at the beginning,
 182          * at game stage a->adapt_phase crosses though 0.5 and
 183          * approaches 1 at the game end; the slope is controlled
 184          * by a->adapt_rate. */
 185         float game_portion;
 186         if (!a->adapt_aport) {
 187                 int total_moves = b->moves + 2 * board_estimated_moves_left(b);
 188                 game_portion = (float) b->moves / total_moves;
 189         } else {
 190                 int brsize = board_size(b) - 2;
 191                 game_portion = 1.0 - (float) b->flen / (brsize * brsize);
 192         }
 193         float l = game_portion - a->adapt_phase;
 194         return 1.0 / (1.0 + exp(-a->adapt_rate * l));
 195 }
 196
 197 static float
 198 adapter_linear(struct uct_dynkomi *d, struct board *b)
 199 {
 200         struct dynkomi_adaptive *a = d->data;
 201         /* Figure out how much to adjust the komi based on the game
 202          * stage. We just linearly increase/decrease the adaptation
 203          * rate for first N moves. */
 204         if (b->moves > a->adapt_moves)
 205                 return 0;
 206         if (a->adapt_dir < 0)
 207                 return 1 - (- a->adapt_dir) * b->moves / a->adapt_moves;
 208         else
 209                 return a->adapt_dir * b->moves / a->adapt_moves;
 210 }
 211
 212 static float
 213 komi_by_score(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color)
 214 {
 215         struct dynkomi_adaptive *a = d->data;
 216         if (d->score.playouts < TRUSTWORTHY_KOMI_PLAYOUTS)
 217                 return tree->extra_komi;
 218
 219         struct move_stats score = d->score;
 220         /* Almost-reset tree->score to gather fresh stats. */
 221         d->score.playouts = 1;
 222
 223         /* Look at average score and push extra_komi in that direction. */
 224         float p = a->adapter(d, b);
 225         p = a->adapt_base + p * (1 - a->adapt_base);
 226         if (p > 0.9) p = 0.9; // don't get too eager!
 227         float extra_komi = tree->extra_komi + p * score.value;
 228         if (DEBUGL(3))
 229                 fprintf(stderr, "mC += %f * %f\n", p, score.value);
 230         return extra_komi;
 231 }
 232
 233 static float
 234 komi_by_value(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color)
 235 {
 236         struct dynkomi_adaptive *a = d->data;
 237         if (d->value.playouts < TRUSTWORTHY_KOMI_PLAYOUTS)
 238                 return tree->extra_komi;
 239
 240         struct move_stats value = d->value;
 241         /* Almost-reset tree->value to gather fresh stats. */
 242         d->value.playouts = 1;
 243         /* Correct color POV. */
 244         if (color == S_WHITE)
 245                 value.value = 1 - value.value;
 246
 247         /* We have three "value zones":
 248          * red zone | yellow zone | green zone
 249          *        ~45%           ~60%
 250          * red zone: reduce komi
 251          * yellow zone: do not touch komi
 252          * green zone: enlage komi.
 253          *
 254          * Also, at some point komi will be tuned in such way
 255          * that it will be in green zone but increasing it will
 256          * be unfeasible. Thus, we have a _latch_ - we will
 257          * remember the last komi that has put us into the
 258          * red zone, and not use it or go over it. We use the
 259          * latch only when giving extra komi, we always want
 260          * to try to reduce extra komi we take.
 261          *
 262          * TODO: Make the latch expire after a while. */
 263         /* We use komi_by_color() first to normalize komi
 264          * additions/subtractions, then apply it again on
 265          * return value to restore original komi parity. */
 266         float extra_komi = komi_by_color(tree->extra_komi, color);
 267
 268         if (value.value < a->zone_red) {
 269                 /* Red zone. Take extra komi. */
 270                 if (DEBUGL(3))
 271                         fprintf(stderr, "[red] %f, komi latch %f -> %f\n",
 272                                 value.value, a->komi_latch, extra_komi);
 273                 if (extra_komi > 0) a->komi_latch = extra_komi;
 274                 extra_komi -= a->score_step;
 275                 return komi_by_color(extra_komi, color);
 276
 277         } else if (value.value < a->zone_green) {
 278                 /* Yellow zone, do nothing. */
 279                 return komi_by_color(extra_komi, color);
 280
 281         } else {
 282                 /* Green zone. Give extra komi. */
 283                 extra_komi += a->score_step;
 284                 if (DEBUGL(3))
 285                         fprintf(stderr, "[green] %f, += %d | komi latch %f\n",
 286                                 value.value, a->score_step, a->komi_latch);
 287                 if (a->use_komi_latch && extra_komi >= a->komi_latch)
 288                         extra_komi = a->komi_latch - 1;
 289                 return komi_by_color(extra_komi, color);
 290         }
 291 }
 292
 293 static float
 294 adaptive_permove(struct uct_dynkomi *d, struct board *b, struct tree *tree)
 295 {
 296         struct dynkomi_adaptive *a = d->data;
 297         if (DEBUGL(3))
 298                 fprintf(stderr, "m %d/%d ekomi %f permove %f/%d\n",
 299                         b->moves, a->lead_moves, tree->extra_komi,
 300                         d->score.value, d->score.playouts);
 301         if (b->moves <= a->lead_moves)
 302                 return board_effective_handicap(b, 7 /* XXX */);
 303
 304         enum stone color = stone_other(tree->root_color);
 305         /* Get lower bound on komi we take so that we don't underperform
 306          * too much. */
 307         float min_komi = komi_by_color(- a->max_losing_komi, color);
 308
 309         float komi = a->indicator(d, b, tree, color);
 310         if (DEBUGL(3))
 311                 fprintf(stderr, "dynkomi: %f -> %f\n", tree->extra_komi, komi);
 312         return komi_by_color(komi - min_komi, color) > 0 ? komi : min_komi;
 313 }
 314
 315 static float
 316 adaptive_persim(struct uct_dynkomi *d, struct board *b, struct tree *tree, struct tree_node *node)
 317 {
 318         return tree->extra_komi;
 319 }
 320
 321 struct uct_dynkomi *
 322 uct_dynkomi_init_adaptive(struct uct *u, char *arg, struct board *b)
 323 {
 324         struct uct_dynkomi *d = calloc2(1, sizeof(*d));
 325         d->uct = u;
 326         d->permove = adaptive_permove;
 327         d->persim = adaptive_persim;
 328         d->done = generic_done;
 329
 330         struct dynkomi_adaptive *a = calloc2(1, sizeof(*a));
 331         d->data = a;
 332
 333         if (board_size(b) - 2 >= 19)
 334                 a->lead_moves = 20;
 335         else
 336                 a->lead_moves = 4; // XXX
 337         a->max_losing_komi = 10;
 338         a->indicator = komi_by_score;
 339
 340         a->adapter = adapter_sigmoid;
 341         a->adapt_rate = 20;
 342         a->adapt_phase = 0.5;
 343         a->adapt_moves = 200;
 344         a->adapt_dir = -0.5;
 345
 346         a->zone_red = 0.45;
 347         a->zone_green = 0.6;
 348         a->score_step = 2;
 349         a->use_komi_latch = true;
 350         a->komi_latch = 1000;
 351
 352         if (arg) {
 353                 char *optspec, *next = arg;
 354                 while (*next) {
 355                         optspec = next;
 356                         next += strcspn(next, ":");
 357                         if (*next) { *next++ = 0; } else { *next = 0; }
 358
 359                         char *optname = optspec;
 360                         char *optval = strchr(optspec, '=');
 361                         if (optval) *optval++ = 0;
 362
 363                         if (!strcasecmp(optname, "lead_moves") && optval) {
 364                                 /* Do not adjust komi adaptively for first
 365                                  * N moves. */
 366                                 a->lead_moves = atoi(optval);
 367                         } else if (!strcasecmp(optname, "max_losing_komi") && optval) {
 368                                 a->max_losing_komi = atof(optval);
 369                         } else if (!strcasecmp(optname, "indicator")) {
 370                                 /* Adaptatation indicator - how to decide
 371                                  * the adaptation rate and direction. */
 372                                 if (!strcasecmp(optval, "value")) {
 373                                         /* Winrate w/ komi so far. */
 374                                         a->indicator = komi_by_value;
 375                                 } else if (!strcasecmp(optval, "score")) {
 376                                         /* Expected score w/ current komi. */
 377                                         a->indicator = komi_by_score;
 378                                 } else {
 379                                         fprintf(stderr, "UCT: Invalid indicator %s\n", optval);
 380                                         exit(1);
 381                                 }
 382
 383                                 /* value indicator settings */
 384                         } else if (!strcasecmp(optname, "zone_red") && optval) {
 385                                 a->zone_red = atof(optval);
 386                         } else if (!strcasecmp(optname, "zone_green") && optval) {
 387                                 a->zone_green = atof(optval);
 388                         } else if (!strcasecmp(optname, "score_step") && optval) {
 389                                 a->score_step = atoi(optval);
 390                         } else if (!strcasecmp(optname, "use_komi_latch")) {
 391                                 a->use_komi_latch = !optval || atoi(optval);
 392
 393                                 /* score indicator settings */
 394                         } else if (!strcasecmp(optname, "adapter") && optval) {
 395                                 /* Adaptatation method. */
 396                                 if (!strcasecmp(optval, "sigmoid")) {
 397                                         a->adapter = adapter_sigmoid;
 398                                 } else if (!strcasecmp(optval, "linear")) {
 399                                         a->adapter = adapter_linear;
 400                                 } else {
 401                                         fprintf(stderr, "UCT: Invalid adapter %s\n", optval);
 402                                         exit(1);
 403                                 }
 404                         } else if (!strcasecmp(optname, "adapt_base") && optval) {
 405                                 /* Adaptation base rate; see above. */
 406                                 a->adapt_base = atof(optval);
 407                         } else if (!strcasecmp(optname, "adapt_rate") && optval) {
 408                                 /* Adaptation slope; see above. */
 409                                 a->adapt_rate = atof(optval);
 410                         } else if (!strcasecmp(optname, "adapt_phase") && optval) {
 411                                 /* Adaptation phase shift; see above. */
 412                                 a->adapt_phase = atof(optval);
 413                         } else if (!strcasecmp(optname, "adapt_moves") && optval) {
 414                                 /* Adaptation move amount; see above. */
 415                                 a->adapt_moves = atoi(optval);
 416                         } else if (!strcasecmp(optname, "adapt_aport")) {
 417                                 a->adapt_aport = !optval || atoi(optval);
 418                         } else if (!strcasecmp(optname, "adapt_dir") && optval) {
 419                                 /* Adaptation direction vector; see above. */
 420                                 a->adapt_dir = atof(optval);
 421
 422                         } else {
 423                                 fprintf(stderr, "uct: Invalid dynkomi argument %s or missing value\n", optname);
 424                                 exit(1);
 425                         }
 426                 }
 427         }
 428
 429         return d;
 430 }