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[S_MAX];
  52         int moves[S_MAX];
  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         enum stone color = stone_other(tree->root_color);
  61         int lmoves = l->moves[color];
  62         if (b->moves >= lmoves)
  63                 return 0;
  64
  65         floating_t base_komi = board_effective_handicap(b, l->handicap_value[color]);
  66         floating_t extra_komi = base_komi * (lmoves - b->moves) / lmoves;
  67         return extra_komi;
  68 }
  69
  70 static floating_t
  71 linear_persim(struct uct_dynkomi *d, struct board *b, struct tree *tree, struct tree_node *node)
  72 {
  73         struct dynkomi_linear *l = d->data;
  74         if (l->rootbased)
  75                 return tree->extra_komi;
  76         /* We don't reuse computed value from tree->extra_komi,
  77          * since we want to use value correct for this node depth.
  78          * This also means the values will stay correct after
  79          * node promotion. */
  80         return linear_permove(d, b, tree);
  81 }
  82
  83 struct uct_dynkomi *
  84 uct_dynkomi_init_linear(struct uct *u, char *arg, struct board *b)
  85 {
  86         struct uct_dynkomi *d = calloc2(1, sizeof(*d));
  87         d->uct = u;
  88         d->permove = linear_permove;
  89         d->persim = linear_persim;
  90         d->done = generic_done;
  91
  92         struct dynkomi_linear *l = calloc2(1, sizeof(*l));
  93         d->data = l;
  94
  95         /* Force white to feel behind and try harder, but not to the
  96          * point of resigning immediately in high handicap games.
  97          * By move 100 white should still be behind but should have
  98          * caught up enough to avoid resigning. */
  99         if (board_large(b)) {
 100                 l->moves[S_BLACK] = 200;
 101                 l->moves[S_WHITE] = 100;
 102         }
 103         /* The real value of one stone is twice the komi so about 15 points.
 104          * But use a lower value to avoid being too pessimistic as black
 105          * or too optimistic as white. */
 106         l->handicap_value[S_BLACK] = 7;
 107         l->handicap_value[S_WHITE] = 5;
 108
 109         if (arg) {
 110                 char *optspec, *next = arg;
 111                 while (*next) {
 112                         optspec = next;
 113                         next += strcspn(next, ":");
 114                         if (*next) { *next++ = 0; } else { *next = 0; }
 115
 116                         char *optname = optspec;
 117                         char *optval = strchr(optspec, '=');
 118                         if (optval) *optval++ = 0;
 119
 120                         if (!strcasecmp(optname, "moves") && optval) {
 121                                 /* Dynamic komi in handicap game; linearly
 122                                  * decreases to basic settings until move
 123                                  * #optval. moves=blackmoves%whitemoves */
 124                                 for (int i = S_BLACK; *optval && i <= S_WHITE; i++, optval += strcspn(optval, "%")) {
 125                                         optval++;
 126                                         l->moves[i] = atoi(optval);
 127                                 }
 128                         } else if (!strcasecmp(optname, "handicap_value") && optval) {
 129                                 /* Point value of single handicap stone,
 130                                  * for dynkomi computation. */
 131                                 for (int i = S_BLACK; *optval && i <= S_WHITE; i++, optval += strcspn(optval, "%")) {
 132                                         optval++;
 133                                         l->handicap_value[i] = atoi(optval);
 134                                 }
 135                         } else if (!strcasecmp(optname, "rootbased")) {
 136                                 /* If set, the extra komi applied will be
 137                                  * the same for all simulations within a move,
 138                                  * instead of being same for all simulations
 139                                  * within the tree node. */
 140                                 l->rootbased = !optval || atoi(optval);
 141                         } else {
 142                                 fprintf(stderr, "uct: Invalid dynkomi argument %s or missing value\n", optname);
 143                                 exit(1);
 144                         }
 145                 }
 146         }
 147
 148         return d;
 149 }
 150
 151
 152 /* ADAPTIVE dynkomi strategy - Adaptive Situational Compensation */
 153 /* We adapt the komi based on current situation:
 154  * (i) score-based: We maintain the average score outcome of our
 155  * games and adjust the komi by a fractional step towards the expected
 156  * score;
 157  * (ii) value-based: While winrate is above given threshold, adjust
 158  * the komi by a fixed step in the appropriate direction.
 159  * These adjustments can be
 160  * (a) Move-stepped, new extra komi value is always set only at the
 161  * beginning of the tree search for next move;
 162  * (b) Continuous, new extra komi value is periodically re-determined
 163  * and adjusted throughout a single tree search. */
 164
 165 struct dynkomi_adaptive {
 166         /* Do not take measured average score into regard for
 167          * first @lead_moves - the variance is just too much.
 168          * (Instead, we consider the handicap-based komi provided
 169          * by linear dynkomi.) */
 170         int lead_moves;
 171         /* Maximum komi to pretend the opponent to give. */
 172         floating_t max_losing_komi;
 173         /* Game portion at which losing komi is not allowed anymore. */
 174         floating_t losing_komi_stop;
 175         /* Alternative game portion determination. */
 176         bool adapt_aport;
 177         floating_t (*indicator)(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color);
 178
 179         /* Value-based adaptation. */
 180         floating_t zone_red, zone_green;
 181         int score_step;
 182         floating_t score_step_byavg; // use portion of average score as increment
 183         bool use_komi_ratchet;
 184         bool losing_komi_ratchet; // ratchet even losing komi
 185         int komi_ratchet_maxage;
 186         // runtime, not configuration:
 187         int komi_ratchet_age;
 188         floating_t komi_ratchet;
 189
 190         /* Score-based adaptation. */
 191         floating_t (*adapter)(struct uct_dynkomi *d, struct board *b);
 192         floating_t adapt_base; // [0,1)
 193         /* Sigmoid adaptation rate parameter; see below for details. */
 194         floating_t adapt_phase; // [0,1]
 195         floating_t adapt_rate; // [1,infty)
 196         /* Linear adaptation rate parameter. */
 197         int adapt_moves;
 198         floating_t adapt_dir; // [-1,1]
 199 };
 200 #define TRUSTWORTHY_KOMI_PLAYOUTS 200
 201
 202 static floating_t
 203 board_game_portion(struct dynkomi_adaptive *a, struct board *b)
 204 {
 205         if (!a->adapt_aport) {
 206                 int total_moves = b->moves + 2 * board_estimated_moves_left(b);
 207                 return (floating_t) b->moves / total_moves;
 208         } else {
 209                 int brsize = board_size(b) - 2;
 210                 return 1.0 - (floating_t) b->flen / (brsize * brsize);
 211         }
 212 }
 213
 214 static floating_t
 215 adapter_sigmoid(struct uct_dynkomi *d, struct board *b)
 216 {
 217         struct dynkomi_adaptive *a = d->data;
 218         /* Figure out how much to adjust the komi based on the game
 219          * stage. The adaptation rate is 0 at the beginning,
 220          * at game stage a->adapt_phase crosses though 0.5 and
 221          * approaches 1 at the game end; the slope is controlled
 222          * by a->adapt_rate. */
 223         floating_t game_portion = board_game_portion(a, b);
 224         floating_t l = game_portion - a->adapt_phase;
 225         return 1.0 / (1.0 + exp(-a->adapt_rate * l));
 226 }
 227
 228 static floating_t
 229 adapter_linear(struct uct_dynkomi *d, struct board *b)
 230 {
 231         struct dynkomi_adaptive *a = d->data;
 232         /* Figure out how much to adjust the komi based on the game
 233          * stage. We just linearly increase/decrease the adaptation
 234          * rate for first N moves. */
 235         if (b->moves > a->adapt_moves)
 236                 return 0;
 237         if (a->adapt_dir < 0)
 238                 return 1 - (- a->adapt_dir) * b->moves / a->adapt_moves;
 239         else
 240                 return a->adapt_dir * b->moves / a->adapt_moves;
 241 }
 242
 243 static floating_t
 244 komi_by_score(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color)
 245 {
 246         struct dynkomi_adaptive *a = d->data;
 247         if (d->score.playouts < TRUSTWORTHY_KOMI_PLAYOUTS)
 248                 return tree->extra_komi;
 249
 250         struct move_stats score = d->score;
 251         /* Almost-reset tree->score to gather fresh stats. */
 252         d->score.playouts = 1;
 253
 254         /* Look at average score and push extra_komi in that direction. */
 255         floating_t p = a->adapter(d, b);
 256         p = a->adapt_base + p * (1 - a->adapt_base);
 257         if (p > 0.9) p = 0.9; // don't get too eager!
 258         floating_t extra_komi = tree->extra_komi + p * score.value;
 259         if (DEBUGL(3))
 260                 fprintf(stderr, "mC += %f * %f\n", p, score.value);
 261         return extra_komi;
 262 }
 263
 264 static floating_t
 265 komi_by_value(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color)
 266 {
 267         struct dynkomi_adaptive *a = d->data;
 268         if (d->value.playouts < TRUSTWORTHY_KOMI_PLAYOUTS)
 269                 return tree->extra_komi;
 270
 271         struct move_stats value = d->value;
 272         /* Almost-reset tree->value to gather fresh stats. */
 273         d->value.playouts = 1;
 274         /* Correct color POV. */
 275         if (color == S_WHITE)
 276                 value.value = 1 - value.value;
 277
 278         /* We have three "value zones":
 279          * red zone | yellow zone | green zone
 280          *        ~45%           ~60%
 281          * red zone: reduce komi
 282          * yellow zone: do not touch komi
 283          * green zone: enlage komi.
 284          *
 285          * Also, at some point komi will be tuned in such way
 286          * that it will be in green zone but increasing it will
 287          * be unfeasible. Thus, we have a _ratchet_ - we will
 288          * remember the last komi that has put us into the
 289          * red zone, and not use it or go over it. We use the
 290          * ratchet only when giving extra komi, we always want
 291          * to try to reduce extra komi we take.
 292          *
 293          * TODO: Make the ratchet expire after a while. */
 294
 295         /* We use komi_by_color() first to normalize komi
 296          * additions/subtractions, then apply it again on
 297          * return value to restore original komi parity. */
 298         /* Positive extra_komi means that we are _giving_
 299          * komi (winning), negative extra_komi is _taking_
 300          * komi (losing). */
 301         floating_t extra_komi = komi_by_color(tree->extra_komi, color);
 302         int score_step_red = -a->score_step;
 303         int score_step_green = a->score_step;
 304
 305         if (a->score_step_byavg != 0) {
 306                 struct move_stats score = d->score;
 307                 /* Almost-reset tree->score to gather fresh stats. */
 308                 d->score.playouts = 1;
 309                 /* Correct color POV. */
 310                 if (color == S_WHITE)
 311                         score.value = - score.value;
 312                 if (score.value > 0)
 313                         score_step_green = round(score.value * a->score_step_byavg);
 314                 else
 315                         score_step_red = round(-score.value * a->score_step_byavg);
 316                 if (score_step_green < 0 || score_step_red > 0) {
 317                         /* The steps are in bad direction - keep still. */
 318                         return komi_by_color(extra_komi, color);
 319                 }
 320         }
 321
 322         /* Wear out the ratchet. */
 323         if (a->use_komi_ratchet && a->komi_ratchet_maxage > 0) {
 324                 a->komi_ratchet_age += value.playouts;
 325                 if (a->komi_ratchet_age > a->komi_ratchet_maxage) {
 326                         a->komi_ratchet = 1000;
 327                         a->komi_ratchet_age = 0;
 328                 }
 329         }
 330
 331         if (value.value < a->zone_red) {
 332                 /* Red zone. Take extra komi. */
 333                 if (DEBUGL(3))
 334                         fprintf(stderr, "[red] %f, step %d | komi ratchet %f age %d/%d -> %f\n",
 335                                 value.value, score_step_red, a->komi_ratchet, a->komi_ratchet_age, a->komi_ratchet_maxage, extra_komi);
 336                 if (a->losing_komi_ratchet || extra_komi > 0) {
 337                         a->komi_ratchet = extra_komi;
 338                         a->komi_ratchet_age = 0;
 339                 }
 340                 extra_komi += score_step_red;
 341                 return komi_by_color(extra_komi, color);
 342
 343         } else if (value.value < a->zone_green) {
 344                 /* Yellow zone, do nothing. */
 345                 return komi_by_color(extra_komi, color);
 346
 347         } else {
 348                 /* Green zone. Give extra komi. */
 349                 if (DEBUGL(3))
 350                         fprintf(stderr, "[green] %f, step %d | komi ratchet %f age %d/%d\n",
 351                                 value.value, score_step_green, a->komi_ratchet, a->komi_ratchet_age, a->komi_ratchet_maxage);
 352                 extra_komi += score_step_green;
 353                 if (a->use_komi_ratchet && extra_komi >= a->komi_ratchet)
 354                         extra_komi = a->komi_ratchet - 1;
 355                 return komi_by_color(extra_komi, color);
 356         }
 357 }
 358
 359 static floating_t
 360 bounded_komi(struct dynkomi_adaptive *a, struct board *b,
 361              enum stone color, floating_t komi, floating_t max_losing_komi)
 362 {
 363         /* At the end of game, disallow losing komi. */
 364         if (komi_by_color(komi, color) < 0
 365             && board_game_portion(a, b) > a->losing_komi_stop)
 366                 return 0;
 367
 368         /* Get lower bound on komi we take so that we don't underperform
 369          * too much. */
 370         floating_t min_komi = komi_by_color(- max_losing_komi, color);
 371
 372         if (komi_by_color(komi - min_komi, color) > 0)
 373                 return komi;
 374         else
 375                 return min_komi;
 376 }
 377
 378 static floating_t
 379 adaptive_permove(struct uct_dynkomi *d, struct board *b, struct tree *tree)
 380 {
 381         struct dynkomi_adaptive *a = d->data;
 382         enum stone color = stone_other(tree->root_color);
 383         if (DEBUGL(3))
 384                 fprintf(stderr, "m %d/%d ekomi %f permove %f/%d\n",
 385                         b->moves, a->lead_moves, tree->extra_komi,
 386                         d->score.value, d->score.playouts);
 387
 388         if (b->moves <= a->lead_moves)
 389                 return bounded_komi(a, b, color,
 390                                     board_effective_handicap(b, 7 /* XXX */),
 391                                     a->max_losing_komi);
 392
 393         floating_t komi = a->indicator(d, b, tree, color);
 394         if (DEBUGL(3))
 395                 fprintf(stderr, "dynkomi: %f -> %f\n", tree->extra_komi, komi);
 396         return bounded_komi(a, b, color, komi, a->max_losing_komi);
 397 }
 398
 399 static floating_t
 400 adaptive_persim(struct uct_dynkomi *d, struct board *b, struct tree *tree, struct tree_node *node)
 401 {
 402         return tree->extra_komi;
 403 }
 404
 405 struct uct_dynkomi *
 406 uct_dynkomi_init_adaptive(struct uct *u, char *arg, struct board *b)
 407 {
 408         struct uct_dynkomi *d = calloc2(1, sizeof(*d));
 409         d->uct = u;
 410         d->permove = adaptive_permove;
 411         d->persim = adaptive_persim;
 412         d->done = generic_done;
 413
 414         struct dynkomi_adaptive *a = calloc2(1, sizeof(*a));
 415         d->data = a;
 416
 417         a->lead_moves = board_large(b) ? 20 : 4; // XXX
 418         a->max_losing_komi = 30;
 419         a->losing_komi_stop = 1.0f;
 420         a->indicator = komi_by_value;
 421
 422         a->adapter = adapter_sigmoid;
 423         a->adapt_rate = -18;
 424         a->adapt_phase = 0.65;
 425         a->adapt_moves = 200;
 426         a->adapt_dir = -0.5;
 427
 428         a->zone_red = 0.45;
 429         a->zone_green = 0.50;
 430         a->score_step = 1;
 431         a->use_komi_ratchet = true;
 432         a->komi_ratchet_maxage = 0;
 433         a->komi_ratchet = 1000;
 434
 435         if (arg) {
 436                 char *optspec, *next = arg;
 437                 while (*next) {
 438                         optspec = next;
 439                         next += strcspn(next, ":");
 440                         if (*next) { *next++ = 0; } else { *next = 0; }
 441
 442                         char *optname = optspec;
 443                         char *optval = strchr(optspec, '=');
 444                         if (optval) *optval++ = 0;
 445
 446                         if (!strcasecmp(optname, "lead_moves") && optval) {
 447                                 /* Do not adjust komi adaptively for first
 448                                  * N moves. */
 449                                 a->lead_moves = atoi(optval);
 450                         } else if (!strcasecmp(optname, "max_losing_komi") && optval) {
 451                                 a->max_losing_komi = atof(optval);
 452                         } else if (!strcasecmp(optname, "losing_komi_stop") && optval) {
 453                                 a->losing_komi_stop = atof(optval);
 454                         } else if (!strcasecmp(optname, "indicator")) {
 455                                 /* Adaptatation indicator - how to decide
 456                                  * the adaptation rate and direction. */
 457                                 if (!strcasecmp(optval, "value")) {
 458                                         /* Winrate w/ komi so far. */
 459                                         a->indicator = komi_by_value;
 460                                 } else if (!strcasecmp(optval, "score")) {
 461                                         /* Expected score w/ current komi. */
 462                                         a->indicator = komi_by_score;
 463                                 } else {
 464                                         fprintf(stderr, "UCT: Invalid indicator %s\n", optval);
 465                                         exit(1);
 466                                 }
 467
 468                                 /* value indicator settings */
 469                         } else if (!strcasecmp(optname, "zone_red") && optval) {
 470                                 a->zone_red = atof(optval);
 471                         } else if (!strcasecmp(optname, "zone_green") && optval) {
 472                                 a->zone_green = atof(optval);
 473                         } else if (!strcasecmp(optname, "score_step") && optval) {
 474                                 a->score_step = atoi(optval);
 475                         } else if (!strcasecmp(optname, "score_step_byavg") && optval) {
 476                                 a->score_step_byavg = atof(optval);
 477                         } else if (!strcasecmp(optname, "use_komi_ratchet")) {
 478                                 a->use_komi_ratchet = !optval || atoi(optval);
 479                         } else if (!strcasecmp(optname, "losing_komi_ratchet")) {
 480                                 a->losing_komi_ratchet = !optval || atoi(optval);
 481                         } else if (!strcasecmp(optname, "komi_ratchet_age") && optval) {
 482                                 a->komi_ratchet_maxage = atoi(optval);
 483
 484                                 /* score indicator settings */
 485                         } else if (!strcasecmp(optname, "adapter") && optval) {
 486                                 /* Adaptatation method. */
 487                                 if (!strcasecmp(optval, "sigmoid")) {
 488                                         a->adapter = adapter_sigmoid;
 489                                 } else if (!strcasecmp(optval, "linear")) {
 490                                         a->adapter = adapter_linear;
 491                                 } else {
 492                                         fprintf(stderr, "UCT: Invalid adapter %s\n", optval);
 493                                         exit(1);
 494                                 }
 495                         } else if (!strcasecmp(optname, "adapt_base") && optval) {
 496                                 /* Adaptation base rate; see above. */
 497                                 a->adapt_base = atof(optval);
 498                         } else if (!strcasecmp(optname, "adapt_rate") && optval) {
 499                                 /* Adaptation slope; see above. */
 500                                 a->adapt_rate = atof(optval);
 501                         } else if (!strcasecmp(optname, "adapt_phase") && optval) {
 502                                 /* Adaptation phase shift; see above. */
 503                                 a->adapt_phase = atof(optval);
 504                         } else if (!strcasecmp(optname, "adapt_moves") && optval) {
 505                                 /* Adaptation move amount; see above. */
 506                                 a->adapt_moves = atoi(optval);
 507                         } else if (!strcasecmp(optname, "adapt_aport")) {
 508                                 a->adapt_aport = !optval || atoi(optval);
 509                         } else if (!strcasecmp(optname, "adapt_dir") && optval) {
 510                                 /* Adaptation direction vector; see above. */
 511                                 a->adapt_dir = atof(optval);
 512
 513                         } else {
 514                                 fprintf(stderr, "uct: Invalid dynkomi argument %s or missing value\n", optname);
 515                                 exit(1);
 516                         }
 517                 }
 518         }
 519
 520         return d;
 521 }