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/util.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 floating_t
  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         floating_t base_komi = board_effective_handicap(b, l->handicap_value);
  64         floating_t extra_komi = base_komi * (l->moves - b->moves) / l->moves;
  65         return extra_komi;
  66 }
  67
  68 static floating_t
  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_large(b))
  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         floating_t max_losing_komi;
 155         /* Game portion at which losing komi is not allowed anymore. */
 156         floating_t losing_komi_stop;
 157         /* Alternative game portion determination. */
 158         bool adapt_aport;
 159         floating_t (*indicator)(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color);
 160
 161         /* Value-based adaptation. */
 162         floating_t zone_red, zone_green;
 163         int score_step;
 164         floating_t score_step_byavg; // use portion of average score as increment
 165         bool use_komi_ratchet;
 166         bool losing_komi_ratchet; // ratchet even losing komi
 167         int komi_ratchet_maxage;
 168         // runtime, not configuration:
 169         int komi_ratchet_age;
 170         floating_t komi_ratchet;
 171
 172         /* Score-based adaptation. */
 173         floating_t (*adapter)(struct uct_dynkomi *d, struct board *b);
 174         floating_t adapt_base; // [0,1)
 175         /* Sigmoid adaptation rate parameter; see below for details. */
 176         floating_t adapt_phase; // [0,1]
 177         floating_t adapt_rate; // [1,infty)
 178         /* Linear adaptation rate parameter. */
 179         int adapt_moves;
 180         floating_t adapt_dir; // [-1,1]
 181 };
 182 #define TRUSTWORTHY_KOMI_PLAYOUTS 200
 183
 184 static floating_t
 185 board_game_portion(struct dynkomi_adaptive *a, struct board *b)
 186 {
 187         if (!a->adapt_aport) {
 188                 int total_moves = b->moves + 2 * board_estimated_moves_left(b);
 189                 return (floating_t) b->moves / total_moves;
 190         } else {
 191                 int brsize = board_size(b) - 2;
 192                 return 1.0 - (floating_t) b->flen / (brsize * brsize);
 193         }
 194 }
 195
 196 static floating_t
 197 adapter_sigmoid(struct uct_dynkomi *d, struct board *b)
 198 {
 199         struct dynkomi_adaptive *a = d->data;
 200         /* Figure out how much to adjust the komi based on the game
 201          * stage. The adaptation rate is 0 at the beginning,
 202          * at game stage a->adapt_phase crosses though 0.5 and
 203          * approaches 1 at the game end; the slope is controlled
 204          * by a->adapt_rate. */
 205         floating_t game_portion = board_game_portion(a, b);
 206         floating_t l = game_portion - a->adapt_phase;
 207         return 1.0 / (1.0 + exp(-a->adapt_rate * l));
 208 }
 209
 210 static floating_t
 211 adapter_linear(struct uct_dynkomi *d, struct board *b)
 212 {
 213         struct dynkomi_adaptive *a = d->data;
 214         /* Figure out how much to adjust the komi based on the game
 215          * stage. We just linearly increase/decrease the adaptation
 216          * rate for first N moves. */
 217         if (b->moves > a->adapt_moves)
 218                 return 0;
 219         if (a->adapt_dir < 0)
 220                 return 1 - (- a->adapt_dir) * b->moves / a->adapt_moves;
 221         else
 222                 return a->adapt_dir * b->moves / a->adapt_moves;
 223 }
 224
 225 static floating_t
 226 komi_by_score(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color)
 227 {
 228         struct dynkomi_adaptive *a = d->data;
 229         if (d->score.playouts < TRUSTWORTHY_KOMI_PLAYOUTS)
 230                 return tree->extra_komi;
 231
 232         struct move_stats score = d->score;
 233         /* Almost-reset tree->score to gather fresh stats. */
 234         d->score.playouts = 1;
 235
 236         /* Look at average score and push extra_komi in that direction. */
 237         floating_t p = a->adapter(d, b);
 238         p = a->adapt_base + p * (1 - a->adapt_base);
 239         if (p > 0.9) p = 0.9; // don't get too eager!
 240         floating_t extra_komi = tree->extra_komi + p * score.value;
 241         if (DEBUGL(3))
 242                 fprintf(stderr, "mC += %f * %f\n", p, score.value);
 243         return extra_komi;
 244 }
 245
 246 static floating_t
 247 komi_by_value(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color)
 248 {
 249         struct dynkomi_adaptive *a = d->data;
 250         if (d->value.playouts < TRUSTWORTHY_KOMI_PLAYOUTS)
 251                 return tree->extra_komi;
 252
 253         struct move_stats value = d->value;
 254         /* Almost-reset tree->value to gather fresh stats. */
 255         d->value.playouts = 1;
 256         /* Correct color POV. */
 257         if (color == S_WHITE)
 258                 value.value = 1 - value.value;
 259
 260         /* We have three "value zones":
 261          * red zone | yellow zone | green zone
 262          *        ~45%           ~60%
 263          * red zone: reduce komi
 264          * yellow zone: do not touch komi
 265          * green zone: enlage komi.
 266          *
 267          * Also, at some point komi will be tuned in such way
 268          * that it will be in green zone but increasing it will
 269          * be unfeasible. Thus, we have a _ratchet_ - we will
 270          * remember the last komi that has put us into the
 271          * red zone, and not use it or go over it. We use the
 272          * ratchet only when giving extra komi, we always want
 273          * to try to reduce extra komi we take.
 274          *
 275          * TODO: Make the ratchet expire after a while. */
 276
 277         /* We use komi_by_color() first to normalize komi
 278          * additions/subtractions, then apply it again on
 279          * return value to restore original komi parity. */
 280         /* Positive extra_komi means that we are _giving_
 281          * komi (winning), negative extra_komi is _taking_
 282          * komi (losing). */
 283         floating_t extra_komi = komi_by_color(tree->extra_komi, color);
 284         int score_step_red = -a->score_step;
 285         int score_step_green = a->score_step;
 286
 287         if (a->score_step_byavg != 0) {
 288                 struct move_stats score = d->score;
 289                 /* Almost-reset tree->score to gather fresh stats. */
 290                 d->score.playouts = 1;
 291                 /* Correct color POV. */
 292                 if (color == S_WHITE)
 293                         score.value = - score.value;
 294                 if (score.value > 0)
 295                         score_step_green = round(score.value * a->score_step_byavg);
 296                 else
 297                         score_step_red = round(-score.value * a->score_step_byavg);
 298                 if (score_step_green < 0 || score_step_red > 0) {
 299                         /* The steps are in bad direction - keep still. */
 300                         return komi_by_color(extra_komi, color);
 301                 }
 302         }
 303
 304         /* Wear out the ratchet. */
 305         if (a->use_komi_ratchet && a->komi_ratchet_maxage > 0) {
 306                 a->komi_ratchet_age += value.playouts;
 307                 if (a->komi_ratchet_age > a->komi_ratchet_maxage) {
 308                         a->komi_ratchet = 1000;
 309                         a->komi_ratchet_age = 0;
 310                 }
 311         }
 312
 313         if (value.value < a->zone_red) {
 314                 /* Red zone. Take extra komi. */
 315                 if (DEBUGL(3))
 316                         fprintf(stderr, "[red] %f, step %d | komi ratchet %f age %d/%d -> %f\n",
 317                                 value.value, score_step_red, a->komi_ratchet, a->komi_ratchet_age, a->komi_ratchet_maxage, extra_komi);
 318                 if (a->losing_komi_ratchet || extra_komi > 0) {
 319                         a->komi_ratchet = extra_komi;
 320                         a->komi_ratchet_age = 0;
 321                 }
 322                 extra_komi += score_step_red;
 323                 return komi_by_color(extra_komi, color);
 324
 325         } else if (value.value < a->zone_green) {
 326                 /* Yellow zone, do nothing. */
 327                 return komi_by_color(extra_komi, color);
 328
 329         } else {
 330                 /* Green zone. Give extra komi. */
 331                 if (DEBUGL(3))
 332                         fprintf(stderr, "[green] %f, step %d | komi ratchet %f age %d/%d\n",
 333                                 value.value, score_step_green, a->komi_ratchet, a->komi_ratchet_age, a->komi_ratchet_maxage);
 334                 extra_komi += score_step_green;
 335                 if (a->use_komi_ratchet && extra_komi >= a->komi_ratchet)
 336                         extra_komi = a->komi_ratchet - 1;
 337                 return komi_by_color(extra_komi, color);
 338         }
 339 }
 340
 341 static floating_t
 342 bounded_komi(struct dynkomi_adaptive *a, struct board *b,
 343              enum stone color, floating_t komi, floating_t max_losing_komi)
 344 {
 345         /* At the end of game, disallow losing komi. */
 346         if (komi_by_color(komi, color) < 0
 347             && board_game_portion(a, b) > a->losing_komi_stop)
 348                 return 0;
 349
 350         /* Get lower bound on komi we take so that we don't underperform
 351          * too much. */
 352         floating_t min_komi = komi_by_color(- max_losing_komi, color);
 353
 354         if (komi_by_color(komi - min_komi, color) > 0)
 355                 return komi;
 356         else
 357                 return min_komi;
 358 }
 359
 360 static floating_t
 361 adaptive_permove(struct uct_dynkomi *d, struct board *b, struct tree *tree)
 362 {
 363         struct dynkomi_adaptive *a = d->data;
 364         enum stone color = stone_other(tree->root_color);
 365         if (DEBUGL(3))
 366                 fprintf(stderr, "m %d/%d ekomi %f permove %f/%d\n",
 367                         b->moves, a->lead_moves, tree->extra_komi,
 368                         d->score.value, d->score.playouts);
 369
 370         if (b->moves <= a->lead_moves)
 371                 return bounded_komi(a, b, color,
 372                                     board_effective_handicap(b, 7 /* XXX */),
 373                                     a->max_losing_komi);
 374
 375         floating_t komi = a->indicator(d, b, tree, color);
 376         if (DEBUGL(3))
 377                 fprintf(stderr, "dynkomi: %f -> %f\n", tree->extra_komi, komi);
 378         return bounded_komi(a, b, color, komi, a->max_losing_komi);
 379 }
 380
 381 static floating_t
 382 adaptive_persim(struct uct_dynkomi *d, struct board *b, struct tree *tree, struct tree_node *node)
 383 {
 384         return tree->extra_komi;
 385 }
 386
 387 struct uct_dynkomi *
 388 uct_dynkomi_init_adaptive(struct uct *u, char *arg, struct board *b)
 389 {
 390         struct uct_dynkomi *d = calloc2(1, sizeof(*d));
 391         d->uct = u;
 392         d->permove = adaptive_permove;
 393         d->persim = adaptive_persim;
 394         d->done = generic_done;
 395
 396         struct dynkomi_adaptive *a = calloc2(1, sizeof(*a));
 397         d->data = a;
 398
 399         a->lead_moves = board_large(b) ? 20 : 4; // XXX
 400         a->max_losing_komi = 30;
 401         a->losing_komi_stop = 1.0f;
 402         a->indicator = komi_by_value;
 403
 404         a->adapter = adapter_sigmoid;
 405         a->adapt_rate = -18;
 406         a->adapt_phase = 0.65;
 407         a->adapt_moves = 200;
 408         a->adapt_dir = -0.5;
 409
 410         a->zone_red = 0.45;
 411         a->zone_green = 0.50;
 412         a->score_step = 1;
 413         a->use_komi_ratchet = true;
 414         a->komi_ratchet_maxage = 0;
 415         a->komi_ratchet = 1000;
 416
 417         if (arg) {
 418                 char *optspec, *next = arg;
 419                 while (*next) {
 420                         optspec = next;
 421                         next += strcspn(next, ":");
 422                         if (*next) { *next++ = 0; } else { *next = 0; }
 423
 424                         char *optname = optspec;
 425                         char *optval = strchr(optspec, '=');
 426                         if (optval) *optval++ = 0;
 427
 428                         if (!strcasecmp(optname, "lead_moves") && optval) {
 429                                 /* Do not adjust komi adaptively for first
 430                                  * N moves. */
 431                                 a->lead_moves = atoi(optval);
 432                         } else if (!strcasecmp(optname, "max_losing_komi") && optval) {
 433                                 a->max_losing_komi = atof(optval);
 434                         } else if (!strcasecmp(optname, "losing_komi_stop") && optval) {
 435                                 a->losing_komi_stop = atof(optval);
 436                         } else if (!strcasecmp(optname, "indicator")) {
 437                                 /* Adaptatation indicator - how to decide
 438                                  * the adaptation rate and direction. */
 439                                 if (!strcasecmp(optval, "value")) {
 440                                         /* Winrate w/ komi so far. */
 441                                         a->indicator = komi_by_value;
 442                                 } else if (!strcasecmp(optval, "score")) {
 443                                         /* Expected score w/ current komi. */
 444                                         a->indicator = komi_by_score;
 445                                 } else {
 446                                         fprintf(stderr, "UCT: Invalid indicator %s\n", optval);
 447                                         exit(1);
 448                                 }
 449
 450                                 /* value indicator settings */
 451                         } else if (!strcasecmp(optname, "zone_red") && optval) {
 452                                 a->zone_red = atof(optval);
 453                         } else if (!strcasecmp(optname, "zone_green") && optval) {
 454                                 a->zone_green = atof(optval);
 455                         } else if (!strcasecmp(optname, "score_step") && optval) {
 456                                 a->score_step = atoi(optval);
 457                         } else if (!strcasecmp(optname, "score_step_byavg") && optval) {
 458                                 a->score_step_byavg = atof(optval);
 459                         } else if (!strcasecmp(optname, "use_komi_ratchet")) {
 460                                 a->use_komi_ratchet = !optval || atoi(optval);
 461                         } else if (!strcasecmp(optname, "losing_komi_ratchet")) {
 462                                 a->losing_komi_ratchet = !optval || atoi(optval);
 463                         } else if (!strcasecmp(optname, "komi_ratchet_age") && optval) {
 464                                 a->komi_ratchet_maxage = atoi(optval);
 465
 466                                 /* score indicator settings */
 467                         } else if (!strcasecmp(optname, "adapter") && optval) {
 468                                 /* Adaptatation method. */
 469                                 if (!strcasecmp(optval, "sigmoid")) {
 470                                         a->adapter = adapter_sigmoid;
 471                                 } else if (!strcasecmp(optval, "linear")) {
 472                                         a->adapter = adapter_linear;
 473                                 } else {
 474                                         fprintf(stderr, "UCT: Invalid adapter %s\n", optval);
 475                                         exit(1);
 476                                 }
 477                         } else if (!strcasecmp(optname, "adapt_base") && optval) {
 478                                 /* Adaptation base rate; see above. */
 479                                 a->adapt_base = atof(optval);
 480                         } else if (!strcasecmp(optname, "adapt_rate") && optval) {
 481                                 /* Adaptation slope; see above. */
 482                                 a->adapt_rate = atof(optval);
 483                         } else if (!strcasecmp(optname, "adapt_phase") && optval) {
 484                                 /* Adaptation phase shift; see above. */
 485                                 a->adapt_phase = atof(optval);
 486                         } else if (!strcasecmp(optname, "adapt_moves") && optval) {
 487                                 /* Adaptation move amount; see above. */
 488                                 a->adapt_moves = atoi(optval);
 489                         } else if (!strcasecmp(optname, "adapt_aport")) {
 490                                 a->adapt_aport = !optval || atoi(optval);
 491                         } else if (!strcasecmp(optname, "adapt_dir") && optval) {
 492                                 /* Adaptation direction vector; see above. */
 493                                 a->adapt_dir = atof(optval);
 494
 495                         } else {
 496                                 fprintf(stderr, "uct: Invalid dynkomi argument %s or missing value\n", optname);
 497                                 exit(1);
 498                         }
 499                 }
 500         }
 501
 502         return d;
 503 }