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. Towards the end of the game the linear compensation
  49  * becomes zero but we increase the extra komi when winning big. This reduces
  50  * the number of point-wasting moves and makes the game more enjoyable for humans. */
  51
  52 struct dynkomi_linear {
  53         int handicap_value[S_MAX];
  54         int moves[S_MAX];
  55         bool rootbased;
  56         /* Increase the extra komi if my win ratio  > green_zone but always
  57          * keep extra_komi <= komi_ratchet. komi_ratchet starts infinite but decreases
  58          * when we give too much extra komi and this puts us back < orange_zone.
  59          * This is meant only to increase the territory margin when playing against
  60          * weaker opponents. We never take negative komi when losing. The ratchet helps
  61          * avoiding oscillations and keeping us above orange_zone.
  62          * To disable the adaptive phase, set green_zone=2. */
  63         floating_t komi_ratchet;
  64         floating_t green_zone;
  65         floating_t orange_zone;
  66         floating_t drop_step;
  67 };
  68
  69 static floating_t
  70 linear_permove(struct uct_dynkomi *d, struct board *b, struct tree *tree)
  71 {
  72         struct dynkomi_linear *l = d->data;
  73         enum stone color = d->uct->pondering ? tree->root_color : stone_other(tree->root_color);
  74         int lmoves = l->moves[color];
  75         if (b->moves < lmoves) {
  76                 floating_t base_komi = board_effective_handicap(b, l->handicap_value[color]);
  77                 return base_komi * (lmoves - b->moves) / lmoves;
  78         }
  79         /* Force a transition to extra_komi 0 before the adaptive phase. */
  80         if (b->moves <= lmoves + 1) return 0;
  81
  82         /* Do not take decisions on unstable value. */
  83         if (tree->root->u.playouts < GJ_MINGAMES) return tree->extra_komi;
  84
  85         floating_t my_value = tree_node_get_value(tree, 1, tree->root->u.value);
  86         /*  We normalize komi as in komi_by_value(), > 0 when winning. */
  87         floating_t extra_komi = komi_by_color(tree->extra_komi, color);
  88         assert(extra_komi >= 0);
  89
  90         if (my_value < 0.5 && l->komi_ratchet > 0 && l->komi_ratchet != INFINITY) {
  91                 if (DEBUGL(0))
  92                         fprintf(stderr, "losing %f extra %.1f ratchet %.1f -> 0\n",
  93                                 my_value, extra_komi, l->komi_ratchet);
  94                 /* Disable dynkomi completely, too dangerous in this game. */
  95                 extra_komi = l->komi_ratchet = 0;
  96
  97         } else if (my_value < l->orange_zone && extra_komi > 0) {
  98                 extra_komi = l->komi_ratchet  = fmax(extra_komi - l->drop_step, 0.0);
  99                 if (DEBUGL(3))
 100                         fprintf(stderr, "dropping to %f ratchet -> %.1f\n",
 101                                 my_value, extra_komi);
 102
 103         } else if (my_value > l->green_zone && extra_komi +1 <= l->komi_ratchet) {
 104                 extra_komi += 1;
 105                 if (DEBUGL(3))
 106                         fprintf(stderr, "winning %f extra_komi -> %.1f, ratchet %.1f\n",
 107                                 my_value, extra_komi, l->komi_ratchet);
 108         }
 109         return komi_by_color(extra_komi, color);
 110 }
 111
 112 static floating_t
 113 linear_persim(struct uct_dynkomi *d, struct board *b, struct tree *tree, struct tree_node *node)
 114 {
 115         struct dynkomi_linear *l = d->data;
 116         if (l->rootbased)
 117                 return tree->extra_komi;
 118         /* We don't reuse computed value from tree->extra_komi,
 119          * since we want to use value correct for this node depth.
 120          * This also means the values will stay correct after
 121          * node promotion. */
 122         return linear_permove(d, b, tree);
 123 }
 124
 125 struct uct_dynkomi *
 126 uct_dynkomi_init_linear(struct uct *u, char *arg, struct board *b)
 127 {
 128         struct uct_dynkomi *d = calloc2(1, sizeof(*d));
 129         d->uct = u;
 130         d->permove = linear_permove;
 131         d->persim = linear_persim;
 132         d->done = generic_done;
 133
 134         struct dynkomi_linear *l = calloc2(1, sizeof(*l));
 135         d->data = l;
 136
 137         /* Force white to feel behind and try harder, but not to the
 138          * point of resigning immediately in high handicap games.
 139          * By move 100 white should still be behind but should have
 140          * caught up enough to avoid resigning. */
 141         if (board_large(b)) {
 142                 l->moves[S_BLACK] = 100;
 143                 l->moves[S_WHITE] = 50;
 144         }
 145         /* The real value of one stone is twice the komi so about 15 points.
 146          * But use a lower value to avoid being too pessimistic as black
 147          * or too optimistic as white. */
 148         l->handicap_value[S_BLACK] = 8;
 149         l->handicap_value[S_WHITE] = 1;
 150
 151         l->komi_ratchet = INFINITY;
 152         l->green_zone = 0.85;
 153         l->orange_zone = 0.8;
 154         l->drop_step = 4.0;
 155
 156         if (arg) {
 157                 char *optspec, *next = arg;
 158                 while (*next) {
 159                         optspec = next;
 160                         next += strcspn(next, ":");
 161                         if (*next) { *next++ = 0; } else { *next = 0; }
 162
 163                         char *optname = optspec;
 164                         char *optval = strchr(optspec, '=');
 165                         if (optval) *optval++ = 0;
 166
 167                         if (!strcasecmp(optname, "moves") && optval) {
 168                                 /* Dynamic komi in handicap game; linearly
 169                                  * decreases to basic settings until move
 170                                  * #optval. moves=blackmoves%whitemoves */
 171                                 for (int i = S_BLACK; *optval && i <= S_WHITE; i++) {
 172                                         l->moves[i] = atoi(optval);
 173                                         optval += strcspn(optval, "%");
 174                                         if (*optval) optval++;
 175                                 }
 176                         } else if (!strcasecmp(optname, "handicap_value") && optval) {
 177                                 /* Point value of single handicap stone,
 178                                  * for dynkomi computation. */
 179                                 for (int i = S_BLACK; *optval && i <= S_WHITE; i++) {
 180                                         l->handicap_value[i] = atoi(optval);
 181                                         optval += strcspn(optval, "%");
 182                                         if (*optval) optval++;
 183                                 }
 184                         } else if (!strcasecmp(optname, "rootbased")) {
 185                                 /* If set, the extra komi applied will be
 186                                  * the same for all simulations within a move,
 187                                  * instead of being same for all simulations
 188                                  * within the tree node. */
 189                                 l->rootbased = !optval || atoi(optval);
 190                         } else if (!strcasecmp(optname, "green_zone") && optval) {
 191                                 /* Increase komi when win ratio is above green_zone */
 192                                 l->green_zone = atof(optval);
 193                         } else if (!strcasecmp(optname, "orange_zone") && optval) {
 194                                 /* Decrease komi when > 0 and win ratio is below orange_zone */
 195                                 l->orange_zone = atof(optval);
 196                         } else if (!strcasecmp(optname, "drop_step") && optval) {
 197                                 /* Decrease komi by drop_step points */
 198                                 l->drop_step = atof(optval);
 199                         } else {
 200                                 fprintf(stderr, "uct: Invalid dynkomi argument %s or missing value\n", optname);
 201                                 exit(1);
 202                         }
 203                 }
 204         }
 205
 206         return d;
 207 }
 208
 209
 210 /* ADAPTIVE dynkomi strategy - Adaptive Situational Compensation */
 211 /* We adapt the komi based on current situation:
 212  * (i) score-based: We maintain the average score outcome of our
 213  * games and adjust the komi by a fractional step towards the expected
 214  * score;
 215  * (ii) value-based: While winrate is above given threshold, adjust
 216  * the komi by a fixed step in the appropriate direction.
 217  * These adjustments can be
 218  * (a) Move-stepped, new extra komi value is always set only at the
 219  * beginning of the tree search for next move;
 220  * (b) Continuous, new extra komi value is periodically re-determined
 221  * and adjusted throughout a single tree search. */
 222
 223 struct dynkomi_adaptive {
 224         /* Do not take measured average score into regard for
 225          * first @lead_moves - the variance is just too much.
 226          * (Instead, we consider the handicap-based komi provided
 227          * by linear dynkomi.) */
 228         int lead_moves;
 229         /* Maximum komi to pretend the opponent to give. */
 230         floating_t max_losing_komi;
 231         /* Game portion at which losing komi is not allowed anymore. */
 232         floating_t losing_komi_stop;
 233         /* Alternative game portion determination. */
 234         bool adapt_aport;
 235         floating_t (*indicator)(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color);
 236
 237         /* Value-based adaptation. */
 238         floating_t zone_red, zone_green;
 239         int score_step;
 240         floating_t score_step_byavg; // use portion of average score as increment
 241         bool use_komi_ratchet;
 242         bool losing_komi_ratchet; // ratchet even losing komi
 243         int komi_ratchet_maxage;
 244         // runtime, not configuration:
 245         int komi_ratchet_age;
 246         floating_t komi_ratchet;
 247
 248         /* Score-based adaptation. */
 249         floating_t (*adapter)(struct uct_dynkomi *d, struct board *b);
 250         floating_t adapt_base; // [0,1)
 251         /* Sigmoid adaptation rate parameter; see below for details. */
 252         floating_t adapt_phase; // [0,1]
 253         floating_t adapt_rate; // [1,infty)
 254         /* Linear adaptation rate parameter. */
 255         int adapt_moves;
 256         floating_t adapt_dir; // [-1,1]
 257 };
 258 #define TRUSTWORTHY_KOMI_PLAYOUTS 200
 259
 260 static floating_t
 261 board_game_portion(struct dynkomi_adaptive *a, struct board *b)
 262 {
 263         if (!a->adapt_aport) {
 264                 int total_moves = b->moves + 2 * board_estimated_moves_left(b);
 265                 return (floating_t) b->moves / total_moves;
 266         } else {
 267                 int brsize = board_size(b) - 2;
 268                 return 1.0 - (floating_t) b->flen / (brsize * brsize);
 269         }
 270 }
 271
 272 static floating_t
 273 adapter_sigmoid(struct uct_dynkomi *d, struct board *b)
 274 {
 275         struct dynkomi_adaptive *a = d->data;
 276         /* Figure out how much to adjust the komi based on the game
 277          * stage. The adaptation rate is 0 at the beginning,
 278          * at game stage a->adapt_phase crosses though 0.5 and
 279          * approaches 1 at the game end; the slope is controlled
 280          * by a->adapt_rate. */
 281         floating_t game_portion = board_game_portion(a, b);
 282         floating_t l = game_portion - a->adapt_phase;
 283         return 1.0 / (1.0 + exp(-a->adapt_rate * l));
 284 }
 285
 286 static floating_t
 287 adapter_linear(struct uct_dynkomi *d, struct board *b)
 288 {
 289         struct dynkomi_adaptive *a = d->data;
 290         /* Figure out how much to adjust the komi based on the game
 291          * stage. We just linearly increase/decrease the adaptation
 292          * rate for first N moves. */
 293         if (b->moves > a->adapt_moves)
 294                 return 0;
 295         if (a->adapt_dir < 0)
 296                 return 1 - (- a->adapt_dir) * b->moves / a->adapt_moves;
 297         else
 298                 return a->adapt_dir * b->moves / a->adapt_moves;
 299 }
 300
 301 static floating_t
 302 komi_by_score(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color)
 303 {
 304         struct dynkomi_adaptive *a = d->data;
 305         if (d->score.playouts < TRUSTWORTHY_KOMI_PLAYOUTS)
 306                 return tree->extra_komi;
 307
 308         struct move_stats score = d->score;
 309         /* Almost-reset tree->score to gather fresh stats. */
 310         d->score.playouts = 1;
 311
 312         /* Look at average score and push extra_komi in that direction. */
 313         floating_t p = a->adapter(d, b);
 314         p = a->adapt_base + p * (1 - a->adapt_base);
 315         if (p > 0.9) p = 0.9; // don't get too eager!
 316         floating_t extra_komi = tree->extra_komi + p * score.value;
 317         if (DEBUGL(3))
 318                 fprintf(stderr, "mC += %f * %f\n", p, score.value);
 319         return extra_komi;
 320 }
 321
 322 static floating_t
 323 komi_by_value(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color)
 324 {
 325         struct dynkomi_adaptive *a = d->data;
 326         if (d->value.playouts < TRUSTWORTHY_KOMI_PLAYOUTS)
 327                 return tree->extra_komi;
 328
 329         struct move_stats value = d->value;
 330         /* Almost-reset tree->value to gather fresh stats. */
 331         d->value.playouts = 1;
 332         /* Correct color POV. */
 333         if (color == S_WHITE)
 334                 value.value = 1 - value.value;
 335
 336         /* We have three "value zones":
 337          * red zone | yellow zone | green zone
 338          *        ~45%           ~60%
 339          * red zone: reduce komi
 340          * yellow zone: do not touch komi
 341          * green zone: enlage komi.
 342          *
 343          * Also, at some point komi will be tuned in such way
 344          * that it will be in green zone but increasing it will
 345          * be unfeasible. Thus, we have a _ratchet_ - we will
 346          * remember the last komi that has put us into the
 347          * red zone, and not use it or go over it. We use the
 348          * ratchet only when giving extra komi, we always want
 349          * to try to reduce extra komi we take.
 350          *
 351          * TODO: Make the ratchet expire after a while. */
 352
 353         /* We use komi_by_color() first to normalize komi
 354          * additions/subtractions, then apply it again on
 355          * return value to restore original komi parity. */
 356         /* Positive extra_komi means that we are _giving_
 357          * komi (winning), negative extra_komi is _taking_
 358          * komi (losing). */
 359         floating_t extra_komi = komi_by_color(tree->extra_komi, color);
 360         int score_step_red = -a->score_step;
 361         int score_step_green = a->score_step;
 362
 363         if (a->score_step_byavg != 0) {
 364                 struct move_stats score = d->score;
 365                 /* Almost-reset tree->score to gather fresh stats. */
 366                 d->score.playouts = 1;
 367                 /* Correct color POV. */
 368                 if (color == S_WHITE)
 369                         score.value = - score.value;
 370                 if (score.value > 0)
 371                         score_step_green = round(score.value * a->score_step_byavg);
 372                 else
 373                         score_step_red = round(-score.value * a->score_step_byavg);
 374                 if (score_step_green < 0 || score_step_red > 0) {
 375                         /* The steps are in bad direction - keep still. */
 376                         return komi_by_color(extra_komi, color);
 377                 }
 378         }
 379
 380         /* Wear out the ratchet. */
 381         if (a->use_komi_ratchet && a->komi_ratchet_maxage > 0) {
 382                 a->komi_ratchet_age += value.playouts;
 383                 if (a->komi_ratchet_age > a->komi_ratchet_maxage) {
 384                         a->komi_ratchet = 1000;
 385                         a->komi_ratchet_age = 0;
 386                 }
 387         }
 388
 389         if (value.value < a->zone_red) {
 390                 /* Red zone. Take extra komi. */
 391                 if (DEBUGL(3))
 392                         fprintf(stderr, "[red] %f, step %d | komi ratchet %f age %d/%d -> %f\n",
 393                                 value.value, score_step_red, a->komi_ratchet, a->komi_ratchet_age, a->komi_ratchet_maxage, extra_komi);
 394                 if (a->losing_komi_ratchet || extra_komi > 0) {
 395                         a->komi_ratchet = extra_komi;
 396                         a->komi_ratchet_age = 0;
 397                 }
 398                 extra_komi += score_step_red;
 399                 return komi_by_color(extra_komi, color);
 400
 401         } else if (value.value < a->zone_green) {
 402                 /* Yellow zone, do nothing. */
 403                 return komi_by_color(extra_komi, color);
 404
 405         } else {
 406                 /* Green zone. Give extra komi. */
 407                 if (DEBUGL(3))
 408                         fprintf(stderr, "[green] %f, step %d | komi ratchet %f age %d/%d\n",
 409                                 value.value, score_step_green, a->komi_ratchet, a->komi_ratchet_age, a->komi_ratchet_maxage);
 410                 extra_komi += score_step_green;
 411                 if (a->use_komi_ratchet && extra_komi >= a->komi_ratchet)
 412                         extra_komi = a->komi_ratchet - 1;
 413                 return komi_by_color(extra_komi, color);
 414         }
 415 }
 416
 417 static floating_t
 418 bounded_komi(struct dynkomi_adaptive *a, struct board *b,
 419              enum stone color, floating_t komi, floating_t max_losing_komi)
 420 {
 421         /* At the end of game, disallow losing komi. */
 422         if (komi_by_color(komi, color) < 0
 423             && board_game_portion(a, b) > a->losing_komi_stop)
 424                 return 0;
 425
 426         /* Get lower bound on komi we take so that we don't underperform
 427          * too much. */
 428         floating_t min_komi = komi_by_color(- max_losing_komi, color);
 429
 430         if (komi_by_color(komi - min_komi, color) > 0)
 431                 return komi;
 432         else
 433                 return min_komi;
 434 }
 435
 436 static floating_t
 437 adaptive_permove(struct uct_dynkomi *d, struct board *b, struct tree *tree)
 438 {
 439         /* We do not use extra komi at the game end - we are not
 440          * to fool ourselves at this point. */
 441         if (board_estimated_moves_left(b) <= MIN_MOVES_LEFT) {
 442                 tree->use_extra_komi = false;
 443                 return 0;
 444         }
 445         struct dynkomi_adaptive *a = d->data;
 446         enum stone color = stone_other(tree->root_color);
 447         if (DEBUGL(3))
 448                 fprintf(stderr, "m %d/%d ekomi %f permove %f/%d\n",
 449                         b->moves, a->lead_moves, tree->extra_komi,
 450                         d->score.value, d->score.playouts);
 451
 452         if (b->moves <= a->lead_moves)
 453                 return bounded_komi(a, b, color,
 454                                     board_effective_handicap(b, 7 /* XXX */),
 455                                     a->max_losing_komi);
 456
 457         floating_t komi = a->indicator(d, b, tree, color);
 458         if (DEBUGL(3))
 459                 fprintf(stderr, "dynkomi: %f -> %f\n", tree->extra_komi, komi);
 460         return bounded_komi(a, b, color, komi, a->max_losing_komi);
 461 }
 462
 463 static floating_t
 464 adaptive_persim(struct uct_dynkomi *d, struct board *b, struct tree *tree, struct tree_node *node)
 465 {
 466         return tree->extra_komi;
 467 }
 468
 469 struct uct_dynkomi *
 470 uct_dynkomi_init_adaptive(struct uct *u, char *arg, struct board *b)
 471 {
 472         struct uct_dynkomi *d = calloc2(1, sizeof(*d));
 473         d->uct = u;
 474         d->permove = adaptive_permove;
 475         d->persim = adaptive_persim;
 476         d->done = generic_done;
 477
 478         struct dynkomi_adaptive *a = calloc2(1, sizeof(*a));
 479         d->data = a;
 480
 481         a->lead_moves = board_large(b) ? 20 : 4; // XXX
 482         a->max_losing_komi = 30;
 483         a->losing_komi_stop = 1.0f;
 484         a->indicator = komi_by_value;
 485
 486         a->adapter = adapter_sigmoid;
 487         a->adapt_rate = -18;
 488         a->adapt_phase = 0.65;
 489         a->adapt_moves = 200;
 490         a->adapt_dir = -0.5;
 491
 492         a->zone_red = 0.45;
 493         a->zone_green = 0.50;
 494         a->score_step = 1;
 495         a->use_komi_ratchet = true;
 496         a->komi_ratchet_maxage = 0;
 497         a->komi_ratchet = 1000;
 498
 499         if (arg) {
 500                 char *optspec, *next = arg;
 501                 while (*next) {
 502                         optspec = next;
 503                         next += strcspn(next, ":");
 504                         if (*next) { *next++ = 0; } else { *next = 0; }
 505
 506                         char *optname = optspec;
 507                         char *optval = strchr(optspec, '=');
 508                         if (optval) *optval++ = 0;
 509
 510                         if (!strcasecmp(optname, "lead_moves") && optval) {
 511                                 /* Do not adjust komi adaptively for first
 512                                  * N moves. */
 513                                 a->lead_moves = atoi(optval);
 514                         } else if (!strcasecmp(optname, "max_losing_komi") && optval) {
 515                                 a->max_losing_komi = atof(optval);
 516                         } else if (!strcasecmp(optname, "losing_komi_stop") && optval) {
 517                                 a->losing_komi_stop = atof(optval);
 518                         } else if (!strcasecmp(optname, "indicator")) {
 519                                 /* Adaptatation indicator - how to decide
 520                                  * the adaptation rate and direction. */
 521                                 if (!strcasecmp(optval, "value")) {
 522                                         /* Winrate w/ komi so far. */
 523                                         a->indicator = komi_by_value;
 524                                 } else if (!strcasecmp(optval, "score")) {
 525                                         /* Expected score w/ current komi. */
 526                                         a->indicator = komi_by_score;
 527                                 } else {
 528                                         fprintf(stderr, "UCT: Invalid indicator %s\n", optval);
 529                                         exit(1);
 530                                 }
 531
 532                                 /* value indicator settings */
 533                         } else if (!strcasecmp(optname, "zone_red") && optval) {
 534                                 a->zone_red = atof(optval);
 535                         } else if (!strcasecmp(optname, "zone_green") && optval) {
 536                                 a->zone_green = atof(optval);
 537                         } else if (!strcasecmp(optname, "score_step") && optval) {
 538                                 a->score_step = atoi(optval);
 539                         } else if (!strcasecmp(optname, "score_step_byavg") && optval) {
 540                                 a->score_step_byavg = atof(optval);
 541                         } else if (!strcasecmp(optname, "use_komi_ratchet")) {
 542                                 a->use_komi_ratchet = !optval || atoi(optval);
 543                         } else if (!strcasecmp(optname, "losing_komi_ratchet")) {
 544                                 a->losing_komi_ratchet = !optval || atoi(optval);
 545                         } else if (!strcasecmp(optname, "komi_ratchet_age") && optval) {
 546                                 a->komi_ratchet_maxage = atoi(optval);
 547
 548                                 /* score indicator settings */
 549                         } else if (!strcasecmp(optname, "adapter") && optval) {
 550                                 /* Adaptatation method. */
 551                                 if (!strcasecmp(optval, "sigmoid")) {
 552                                         a->adapter = adapter_sigmoid;
 553                                 } else if (!strcasecmp(optval, "linear")) {
 554                                         a->adapter = adapter_linear;
 555                                 } else {
 556                                         fprintf(stderr, "UCT: Invalid adapter %s\n", optval);
 557                                         exit(1);
 558                                 }
 559                         } else if (!strcasecmp(optname, "adapt_base") && optval) {
 560                                 /* Adaptation base rate; see above. */
 561                                 a->adapt_base = atof(optval);
 562                         } else if (!strcasecmp(optname, "adapt_rate") && optval) {
 563                                 /* Adaptation slope; see above. */
 564                                 a->adapt_rate = atof(optval);
 565                         } else if (!strcasecmp(optname, "adapt_phase") && optval) {
 566                                 /* Adaptation phase shift; see above. */
 567                                 a->adapt_phase = atof(optval);
 568                         } else if (!strcasecmp(optname, "adapt_moves") && optval) {
 569                                 /* Adaptation move amount; see above. */
 570                                 a->adapt_moves = atoi(optval);
 571                         } else if (!strcasecmp(optname, "adapt_aport")) {
 572                                 a->adapt_aport = !optval || atoi(optval);
 573                         } else if (!strcasecmp(optname, "adapt_dir") && optval) {
 574                                 /* Adaptation direction vector; see above. */
 575                                 a->adapt_dir = atof(optval);
 576
 577                         } else {
 578                                 fprintf(stderr, "uct: Invalid dynkomi argument %s or missing value\n", optname);
 579                                 exit(1);
 580                         }
 581                 }
 582         }
 583
 584         return d;
 585 }