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         float score_step_byavg; // use portion of average score as increment
 161         bool use_komi_ratchet;
 162         int komi_ratchet_maxage;
 163         // runtime, not configuration:
 164         int komi_ratchet_age;
 165         float komi_ratchet;
 166
 167         /* Score-based adaptation. */
 168         float (*adapter)(struct uct_dynkomi *d, struct board *b);
 169         float adapt_base; // [0,1)
 170         /* Sigmoid adaptation rate parameter; see below for details. */
 171         float adapt_phase; // [0,1]
 172         float adapt_rate; // [1,infty)
 173         bool adapt_aport; // alternative game portion determination
 174         /* Linear adaptation rate parameter. */
 175         int adapt_moves;
 176         float adapt_dir; // [-1,1]
 177 };
 178 #define TRUSTWORTHY_KOMI_PLAYOUTS 200
 179
 180 static float
 181 adapter_sigmoid(struct uct_dynkomi *d, struct board *b)
 182 {
 183         struct dynkomi_adaptive *a = d->data;
 184         /* Figure out how much to adjust the komi based on the game
 185          * stage. The adaptation rate is 0 at the beginning,
 186          * at game stage a->adapt_phase crosses though 0.5 and
 187          * approaches 1 at the game end; the slope is controlled
 188          * by a->adapt_rate. */
 189         float game_portion;
 190         if (!a->adapt_aport) {
 191                 int total_moves = b->moves + 2 * board_estimated_moves_left(b);
 192                 game_portion = (float) b->moves / total_moves;
 193         } else {
 194                 int brsize = board_size(b) - 2;
 195                 game_portion = 1.0 - (float) b->flen / (brsize * brsize);
 196         }
 197         float l = game_portion - a->adapt_phase;
 198         return 1.0 / (1.0 + exp(-a->adapt_rate * l));
 199 }
 200
 201 static float
 202 adapter_linear(struct uct_dynkomi *d, struct board *b)
 203 {
 204         struct dynkomi_adaptive *a = d->data;
 205         /* Figure out how much to adjust the komi based on the game
 206          * stage. We just linearly increase/decrease the adaptation
 207          * rate for first N moves. */
 208         if (b->moves > a->adapt_moves)
 209                 return 0;
 210         if (a->adapt_dir < 0)
 211                 return 1 - (- a->adapt_dir) * b->moves / a->adapt_moves;
 212         else
 213                 return a->adapt_dir * b->moves / a->adapt_moves;
 214 }
 215
 216 static float
 217 komi_by_score(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color)
 218 {
 219         struct dynkomi_adaptive *a = d->data;
 220         if (d->score.playouts < TRUSTWORTHY_KOMI_PLAYOUTS)
 221                 return tree->extra_komi;
 222
 223         struct move_stats score = d->score;
 224         /* Almost-reset tree->score to gather fresh stats. */
 225         d->score.playouts = 1;
 226
 227         /* Look at average score and push extra_komi in that direction. */
 228         float p = a->adapter(d, b);
 229         p = a->adapt_base + p * (1 - a->adapt_base);
 230         if (p > 0.9) p = 0.9; // don't get too eager!
 231         float extra_komi = tree->extra_komi + p * score.value;
 232         if (DEBUGL(3))
 233                 fprintf(stderr, "mC += %f * %f\n", p, score.value);
 234         return extra_komi;
 235 }
 236
 237 static float
 238 komi_by_value(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color)
 239 {
 240         struct dynkomi_adaptive *a = d->data;
 241         if (d->value.playouts < TRUSTWORTHY_KOMI_PLAYOUTS)
 242                 return tree->extra_komi;
 243
 244         struct move_stats value = d->value;
 245         /* Almost-reset tree->value to gather fresh stats. */
 246         d->value.playouts = 1;
 247         /* Correct color POV. */
 248         if (color == S_WHITE)
 249                 value.value = 1 - value.value;
 250
 251         /* We have three "value zones":
 252          * red zone | yellow zone | green zone
 253          *        ~45%           ~60%
 254          * red zone: reduce komi
 255          * yellow zone: do not touch komi
 256          * green zone: enlage komi.
 257          *
 258          * Also, at some point komi will be tuned in such way
 259          * that it will be in green zone but increasing it will
 260          * be unfeasible. Thus, we have a _ratchet_ - we will
 261          * remember the last komi that has put us into the
 262          * red zone, and not use it or go over it. We use the
 263          * ratchet only when giving extra komi, we always want
 264          * to try to reduce extra komi we take.
 265          *
 266          * TODO: Make the ratchet expire after a while. */
 267         /* We use komi_by_color() first to normalize komi
 268          * additions/subtractions, then apply it again on
 269          * return value to restore original komi parity. */
 270         float extra_komi = komi_by_color(tree->extra_komi, color);
 271         int score_step_red = -a->score_step;
 272         int score_step_green = a->score_step;
 273
 274         if (a->score_step_byavg != 0) {
 275                 struct move_stats score = d->score;
 276                 /* Almost-reset tree->score to gather fresh stats. */
 277                 d->score.playouts = 1;
 278                 /* Correct color POV. */
 279                 if (color == S_WHITE)
 280                         score.value = - score.value;
 281                 if (score.value > 0)
 282                         score_step_green = round(score.value * a->score_step_byavg);
 283                 else
 284                         score_step_red = round(-score.value * a->score_step_byavg);
 285         }
 286
 287         if (value.value < a->zone_red) {
 288                 /* Red zone. Take extra komi. */
 289                 if (DEBUGL(3))
 290                         fprintf(stderr, "[red] %f, step %d | komi ratchet %f age %d/%d -> %f\n",
 291                                 value.value, score_step_red, a->komi_ratchet, a->komi_ratchet_age, a->komi_ratchet_maxage, extra_komi);
 292                 if (extra_komi > 0) {
 293                         assert(extra_komi < a->komi_ratchet);
 294                         a->komi_ratchet = extra_komi;
 295                         a->komi_ratchet_age = 0;
 296                 }
 297                 extra_komi += score_step_red;
 298                 return komi_by_color(extra_komi, color);
 299
 300         } else if (value.value < a->zone_green) {
 301                 /* Yellow zone, do nothing. */
 302                 return komi_by_color(extra_komi, color);
 303
 304         } else {
 305                 /* Green zone. Give extra komi. */
 306                 extra_komi += score_step_green;
 307                 if (DEBUGL(3))
 308                         fprintf(stderr, "[green] %f, step %d | komi ratchet %f age %d/%d\n",
 309                                 value.value, score_step_green, a->komi_ratchet, a->komi_ratchet_age, a->komi_ratchet_maxage);
 310                 if (a->komi_ratchet_maxage > 0 && a->komi_ratchet_age > a->komi_ratchet_maxage) {
 311                         a->komi_ratchet = 1000;
 312                         a->komi_ratchet_age = 0;
 313                 }
 314                 if (a->use_komi_ratchet && extra_komi >= a->komi_ratchet) {
 315                         extra_komi = a->komi_ratchet - 1;
 316                         a->komi_ratchet_age++;
 317                 }
 318                 return komi_by_color(extra_komi, color);
 319         }
 320 }
 321
 322 static float
 323 bounded_komi(enum stone color, float komi, float max_losing_komi)
 324 {
 325         /* Get lower bound on komi we take so that we don't underperform
 326          * too much. */
 327         float min_komi = komi_by_color(- max_losing_komi, color);
 328
 329         if (komi_by_color(komi - min_komi, color) > 0)
 330                 return komi;
 331         else
 332                 return min_komi;
 333 }
 334
 335 static float
 336 adaptive_permove(struct uct_dynkomi *d, struct board *b, struct tree *tree)
 337 {
 338         struct dynkomi_adaptive *a = d->data;
 339         enum stone color = stone_other(tree->root_color);
 340         if (DEBUGL(3))
 341                 fprintf(stderr, "m %d/%d ekomi %f permove %f/%d\n",
 342                         b->moves, a->lead_moves, tree->extra_komi,
 343                         d->score.value, d->score.playouts);
 344
 345         if (b->moves <= a->lead_moves)
 346                 return bounded_komi(color, board_effective_handicap(b, 7 /* XXX */), a->max_losing_komi);
 347
 348         float komi = a->indicator(d, b, tree, color);
 349         if (DEBUGL(3))
 350                 fprintf(stderr, "dynkomi: %f -> %f\n", tree->extra_komi, komi);
 351         return bounded_komi(color, komi, a->max_losing_komi);
 352 }
 353
 354 static float
 355 adaptive_persim(struct uct_dynkomi *d, struct board *b, struct tree *tree, struct tree_node *node)
 356 {
 357         return tree->extra_komi;
 358 }
 359
 360 struct uct_dynkomi *
 361 uct_dynkomi_init_adaptive(struct uct *u, char *arg, struct board *b)
 362 {
 363         struct uct_dynkomi *d = calloc2(1, sizeof(*d));
 364         d->uct = u;
 365         d->permove = adaptive_permove;
 366         d->persim = adaptive_persim;
 367         d->done = generic_done;
 368
 369         struct dynkomi_adaptive *a = calloc2(1, sizeof(*a));
 370         d->data = a;
 371
 372         if (board_size(b) - 2 >= 19)
 373                 a->lead_moves = 20;
 374         else
 375                 a->lead_moves = 4; // XXX
 376         a->max_losing_komi = 0;
 377         a->indicator = komi_by_score;
 378
 379         a->adapter = adapter_sigmoid;
 380         a->adapt_rate = -18;
 381         a->adapt_phase = 0.65;
 382         a->adapt_moves = 200;
 383         a->adapt_dir = -0.5;
 384
 385         a->zone_red = 0.45;
 386         a->zone_green = 0.55;
 387         a->score_step = 2;
 388         a->use_komi_ratchet = true;
 389         a->komi_ratchet_maxage = 0;
 390         a->komi_ratchet = 1000;
 391
 392         if (arg) {
 393                 char *optspec, *next = arg;
 394                 while (*next) {
 395                         optspec = next;
 396                         next += strcspn(next, ":");
 397                         if (*next) { *next++ = 0; } else { *next = 0; }
 398
 399                         char *optname = optspec;
 400                         char *optval = strchr(optspec, '=');
 401                         if (optval) *optval++ = 0;
 402
 403                         if (!strcasecmp(optname, "lead_moves") && optval) {
 404                                 /* Do not adjust komi adaptively for first
 405                                  * N moves. */
 406                                 a->lead_moves = atoi(optval);
 407                         } else if (!strcasecmp(optname, "max_losing_komi") && optval) {
 408                                 a->max_losing_komi = atof(optval);
 409                         } else if (!strcasecmp(optname, "indicator")) {
 410                                 /* Adaptatation indicator - how to decide
 411                                  * the adaptation rate and direction. */
 412                                 if (!strcasecmp(optval, "value")) {
 413                                         /* Winrate w/ komi so far. */
 414                                         a->indicator = komi_by_value;
 415                                 } else if (!strcasecmp(optval, "score")) {
 416                                         /* Expected score w/ current komi. */
 417                                         a->indicator = komi_by_score;
 418                                 } else {
 419                                         fprintf(stderr, "UCT: Invalid indicator %s\n", optval);
 420                                         exit(1);
 421                                 }
 422
 423                                 /* value indicator settings */
 424                         } else if (!strcasecmp(optname, "zone_red") && optval) {
 425                                 a->zone_red = atof(optval);
 426                         } else if (!strcasecmp(optname, "zone_green") && optval) {
 427                                 a->zone_green = atof(optval);
 428                         } else if (!strcasecmp(optname, "score_step") && optval) {
 429                                 a->score_step = atoi(optval);
 430                         } else if (!strcasecmp(optname, "score_step_byavg") && optval) {
 431                                 a->score_step_byavg = atof(optval);
 432                         } else if (!strcasecmp(optname, "use_komi_ratchet")) {
 433                                 a->use_komi_ratchet = !optval || atoi(optval);
 434                         } else if (!strcasecmp(optname, "komi_ratchet_age") && optval) {
 435                                 a->komi_ratchet_maxage = atoi(optval);
 436
 437                                 /* score indicator settings */
 438                         } else if (!strcasecmp(optname, "adapter") && optval) {
 439                                 /* Adaptatation method. */
 440                                 if (!strcasecmp(optval, "sigmoid")) {
 441                                         a->adapter = adapter_sigmoid;
 442                                 } else if (!strcasecmp(optval, "linear")) {
 443                                         a->adapter = adapter_linear;
 444                                 } else {
 445                                         fprintf(stderr, "UCT: Invalid adapter %s\n", optval);
 446                                         exit(1);
 447                                 }
 448                         } else if (!strcasecmp(optname, "adapt_base") && optval) {
 449                                 /* Adaptation base rate; see above. */
 450                                 a->adapt_base = atof(optval);
 451                         } else if (!strcasecmp(optname, "adapt_rate") && optval) {
 452                                 /* Adaptation slope; see above. */
 453                                 a->adapt_rate = atof(optval);
 454                         } else if (!strcasecmp(optname, "adapt_phase") && optval) {
 455                                 /* Adaptation phase shift; see above. */
 456                                 a->adapt_phase = atof(optval);
 457                         } else if (!strcasecmp(optname, "adapt_moves") && optval) {
 458                                 /* Adaptation move amount; see above. */
 459                                 a->adapt_moves = atoi(optval);
 460                         } else if (!strcasecmp(optname, "adapt_aport")) {
 461                                 a->adapt_aport = !optval || atoi(optval);
 462                         } else if (!strcasecmp(optname, "adapt_dir") && optval) {
 463                                 /* Adaptation direction vector; see above. */
 464                                 a->adapt_dir = atof(optval);
 465
 466                         } else {
 467                                 fprintf(stderr, "uct: Invalid dynkomi argument %s or missing value\n", optname);
 468                                 exit(1);
 469                         }
 470                 }
 471         }
 472
 473         return d;
 474 }