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         floating_t extra_komi;
  76
  77         if (b->moves < lmoves) {
  78                 floating_t base_komi = board_effective_handicap(b, l->handicap_value[color]);
  79                 extra_komi = base_komi * (lmoves - b->moves) / lmoves;
  80                 return extra_komi;
  81         } else {
  82                 extra_komi = floor(tree->extra_komi);
  83         }
  84
  85         /* Do not take decisions on unstable value. */
  86         if (tree->root->u.playouts < GJ_MINGAMES) return extra_komi;
  87
  88         floating_t my_value = tree_node_get_value(tree, 1, tree->root->u.value);
  89         /*  We normalize komi as in komi_by_value(), > 0 when winning. */
  90         extra_komi = komi_by_color(extra_komi, color);
  91         assert(extra_komi >= 0);
  92         floating_t orig_komi = extra_komi;
  93
  94         if (my_value < 0.5 && l->komi_ratchet > 0 && l->komi_ratchet != INFINITY) {
  95                 if (DEBUGL(0))
  96                         fprintf(stderr, "losing %f extra %.1f ratchet %.1f -> 0\n",
  97                                 my_value, extra_komi, l->komi_ratchet);
  98                 /* Disable dynkomi completely, too dangerous in this game. */
  99                 extra_komi = l->komi_ratchet = 0;
 100
 101         } else if (my_value < l->orange_zone && extra_komi > 0) {
 102                 extra_komi = l->komi_ratchet  = fmax(extra_komi - l->drop_step, 0.0);
 103                 if (extra_komi != orig_komi && DEBUGL(3))
 104                         fprintf(stderr, "dropping to %f ratchet -> %.1f\n",
 105                                 my_value, extra_komi);
 106
 107         } else if (my_value > l->green_zone && extra_komi +1 <= l->komi_ratchet) {
 108                 extra_komi += 1;
 109                 if (extra_komi != orig_komi && DEBUGL(3))
 110                         fprintf(stderr, "winning %f extra_komi -> %.1f, ratchet %.1f\n",
 111                                 my_value, extra_komi, l->komi_ratchet);
 112         }
 113         return komi_by_color(extra_komi, color);
 114 }
 115
 116 static floating_t
 117 linear_persim(struct uct_dynkomi *d, struct board *b, struct tree *tree, struct tree_node *node)
 118 {
 119         struct dynkomi_linear *l = d->data;
 120         if (l->rootbased)
 121                 return tree->extra_komi;
 122         /* We don't reuse computed value from tree->extra_komi,
 123          * since we want to use value correct for this node depth.
 124          * This also means the values will stay correct after
 125          * node promotion. */
 126         return linear_permove(d, b, tree);
 127 }
 128
 129 struct uct_dynkomi *
 130 uct_dynkomi_init_linear(struct uct *u, char *arg, struct board *b)
 131 {
 132         struct uct_dynkomi *d = calloc2(1, sizeof(*d));
 133         d->uct = u;
 134         d->permove = linear_permove;
 135         d->persim = linear_persim;
 136         d->done = generic_done;
 137
 138         struct dynkomi_linear *l = calloc2(1, sizeof(*l));
 139         d->data = l;
 140
 141         /* Force white to feel behind and try harder, but not to the
 142          * point of resigning immediately in high handicap games.
 143          * By move 100 white should still be behind but should have
 144          * caught up enough to avoid resigning. */
 145         if (board_large(b)) {
 146                 l->moves[S_BLACK] = 100;
 147                 l->moves[S_WHITE] = 50;
 148         }
 149         /* The real value of one stone is twice the komi so about 15 points.
 150          * But use a lower value to avoid being too pessimistic as black
 151          * or too optimistic as white. */
 152         l->handicap_value[S_BLACK] = 8;
 153         l->handicap_value[S_WHITE] = 1;
 154
 155         l->komi_ratchet = INFINITY;
 156         l->green_zone = 0.85;
 157         l->orange_zone = 0.8;
 158         l->drop_step = 4.0;
 159
 160         if (arg) {
 161                 char *optspec, *next = arg;
 162                 while (*next) {
 163                         optspec = next;
 164                         next += strcspn(next, ":");
 165                         if (*next) { *next++ = 0; } else { *next = 0; }
 166
 167                         char *optname = optspec;
 168                         char *optval = strchr(optspec, '=');
 169                         if (optval) *optval++ = 0;
 170
 171                         if (!strcasecmp(optname, "moves") && optval) {
 172                                 /* Dynamic komi in handicap game; linearly
 173                                  * decreases to basic settings until move
 174                                  * #optval. moves=blackmoves%whitemoves */
 175                                 for (int i = S_BLACK; *optval && i <= S_WHITE; i++) {
 176                                         l->moves[i] = atoi(optval);
 177                                         optval += strcspn(optval, "%");
 178                                         if (*optval) optval++;
 179                                 }
 180                         } else if (!strcasecmp(optname, "handicap_value") && optval) {
 181                                 /* Point value of single handicap stone,
 182                                  * for dynkomi computation. */
 183                                 for (int i = S_BLACK; *optval && i <= S_WHITE; i++) {
 184                                         l->handicap_value[i] = atoi(optval);
 185                                         optval += strcspn(optval, "%");
 186                                         if (*optval) optval++;
 187                                 }
 188                         } else if (!strcasecmp(optname, "rootbased")) {
 189                                 /* If set, the extra komi applied will be
 190                                  * the same for all simulations within a move,
 191                                  * instead of being same for all simulations
 192                                  * within the tree node. */
 193                                 l->rootbased = !optval || atoi(optval);
 194                         } else if (!strcasecmp(optname, "green_zone") && optval) {
 195                                 /* Increase komi when win ratio is above green_zone */
 196                                 l->green_zone = atof(optval);
 197                         } else if (!strcasecmp(optname, "orange_zone") && optval) {
 198                                 /* Decrease komi when > 0 and win ratio is below orange_zone */
 199                                 l->orange_zone = atof(optval);
 200                         } else if (!strcasecmp(optname, "drop_step") && optval) {
 201                                 /* Decrease komi by drop_step points */
 202                                 l->drop_step = atof(optval);
 203                         } else {
 204                                 fprintf(stderr, "uct: Invalid dynkomi argument %s or missing value\n", optname);
 205                                 exit(1);
 206                         }
 207                 }
 208         }
 209
 210         return d;
 211 }
 212
 213
 214 /* ADAPTIVE dynkomi strategy - Adaptive Situational Compensation */
 215 /* We adapt the komi based on current situation:
 216  * (i) score-based: We maintain the average score outcome of our
 217  * games and adjust the komi by a fractional step towards the expected
 218  * score;
 219  * (ii) value-based: While winrate is above given threshold, adjust
 220  * the komi by a fixed step in the appropriate direction.
 221  * These adjustments can be
 222  * (a) Move-stepped, new extra komi value is always set only at the
 223  * beginning of the tree search for next move;
 224  * (b) Continuous, new extra komi value is periodically re-determined
 225  * and adjusted throughout a single tree search. */
 226
 227 struct dynkomi_adaptive {
 228         /* Do not take measured average score into regard for
 229          * first @lead_moves - the variance is just too much.
 230          * (Instead, we consider the handicap-based komi provided
 231          * by linear dynkomi.) */
 232         int lead_moves;
 233         /* Maximum komi to pretend the opponent to give. */
 234         floating_t max_losing_komi;
 235         /* Game portion at which losing komi is not allowed anymore. */
 236         floating_t losing_komi_stop;
 237         /* Alternative game portion determination. */
 238         bool adapt_aport;
 239         floating_t (*indicator)(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color);
 240
 241         /* Value-based adaptation. */
 242         floating_t zone_red, zone_green;
 243         int score_step;
 244         floating_t score_step_byavg; // use portion of average score as increment
 245         bool use_komi_ratchet;
 246         bool losing_komi_ratchet; // ratchet even losing komi
 247         int komi_ratchet_maxage;
 248         // runtime, not configuration:
 249         int komi_ratchet_age;
 250         floating_t komi_ratchet;
 251
 252         /* Score-based adaptation. */
 253         floating_t (*adapter)(struct uct_dynkomi *d, struct board *b);
 254         floating_t adapt_base; // [0,1)
 255         /* Sigmoid adaptation rate parameter; see below for details. */
 256         floating_t adapt_phase; // [0,1]
 257         floating_t adapt_rate; // [1,infty)
 258         /* Linear adaptation rate parameter. */
 259         int adapt_moves;
 260         floating_t adapt_dir; // [-1,1]
 261 };
 262 #define TRUSTWORTHY_KOMI_PLAYOUTS 200
 263
 264 static floating_t
 265 board_game_portion(struct dynkomi_adaptive *a, struct board *b)
 266 {
 267         if (!a->adapt_aport) {
 268                 int total_moves = b->moves + 2 * board_estimated_moves_left(b);
 269                 return (floating_t) b->moves / total_moves;
 270         } else {
 271                 int brsize = board_size(b) - 2;
 272                 return 1.0 - (floating_t) b->flen / (brsize * brsize);
 273         }
 274 }
 275
 276 static floating_t
 277 adapter_sigmoid(struct uct_dynkomi *d, struct board *b)
 278 {
 279         struct dynkomi_adaptive *a = d->data;
 280         /* Figure out how much to adjust the komi based on the game
 281          * stage. The adaptation rate is 0 at the beginning,
 282          * at game stage a->adapt_phase crosses though 0.5 and
 283          * approaches 1 at the game end; the slope is controlled
 284          * by a->adapt_rate. */
 285         floating_t game_portion = board_game_portion(a, b);
 286         floating_t l = game_portion - a->adapt_phase;
 287         return 1.0 / (1.0 + exp(-a->adapt_rate * l));
 288 }
 289
 290 static floating_t
 291 adapter_linear(struct uct_dynkomi *d, struct board *b)
 292 {
 293         struct dynkomi_adaptive *a = d->data;
 294         /* Figure out how much to adjust the komi based on the game
 295          * stage. We just linearly increase/decrease the adaptation
 296          * rate for first N moves. */
 297         if (b->moves > a->adapt_moves)
 298                 return 0;
 299         if (a->adapt_dir < 0)
 300                 return 1 - (- a->adapt_dir) * b->moves / a->adapt_moves;
 301         else
 302                 return a->adapt_dir * b->moves / a->adapt_moves;
 303 }
 304
 305 static floating_t
 306 komi_by_score(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color)
 307 {
 308         struct dynkomi_adaptive *a = d->data;
 309         if (d->score.playouts < TRUSTWORTHY_KOMI_PLAYOUTS)
 310                 return tree->extra_komi;
 311
 312         struct move_stats score = d->score;
 313         /* Almost-reset tree->score to gather fresh stats. */
 314         d->score.playouts = 1;
 315
 316         /* Look at average score and push extra_komi in that direction. */
 317         floating_t p = a->adapter(d, b);
 318         p = a->adapt_base + p * (1 - a->adapt_base);
 319         if (p > 0.9) p = 0.9; // don't get too eager!
 320         floating_t extra_komi = tree->extra_komi + p * score.value;
 321         if (DEBUGL(3))
 322                 fprintf(stderr, "mC += %f * %f\n", p, score.value);
 323         return extra_komi;
 324 }
 325
 326 static floating_t
 327 komi_by_value(struct uct_dynkomi *d, struct board *b, struct tree *tree, enum stone color)
 328 {
 329         struct dynkomi_adaptive *a = d->data;
 330         if (d->value.playouts < TRUSTWORTHY_KOMI_PLAYOUTS)
 331                 return tree->extra_komi;
 332
 333         struct move_stats value = d->value;
 334         /* Almost-reset tree->value to gather fresh stats. */
 335         d->value.playouts = 1;
 336         /* Correct color POV. */
 337         if (color == S_WHITE)
 338                 value.value = 1 - value.value;
 339
 340         /* We have three "value zones":
 341          * red zone | yellow zone | green zone
 342          *        ~45%           ~60%
 343          * red zone: reduce komi
 344          * yellow zone: do not touch komi
 345          * green zone: enlage komi.
 346          *
 347          * Also, at some point komi will be tuned in such way
 348          * that it will be in green zone but increasing it will
 349          * be unfeasible. Thus, we have a _ratchet_ - we will
 350          * remember the last komi that has put us into the
 351          * red zone, and not use it or go over it. We use the
 352          * ratchet only when giving extra komi, we always want
 353          * to try to reduce extra komi we take.
 354          *
 355          * TODO: Make the ratchet expire after a while. */
 356
 357         /* We use komi_by_color() first to normalize komi
 358          * additions/subtractions, then apply it again on
 359          * return value to restore original komi parity. */
 360         /* Positive extra_komi means that we are _giving_
 361          * komi (winning), negative extra_komi is _taking_
 362          * komi (losing). */
 363         floating_t extra_komi = komi_by_color(tree->extra_komi, color);
 364         int score_step_red = -a->score_step;
 365         int score_step_green = a->score_step;
 366
 367         if (a->score_step_byavg != 0) {
 368                 struct move_stats score = d->score;
 369                 /* Almost-reset tree->score to gather fresh stats. */
 370                 d->score.playouts = 1;
 371                 /* Correct color POV. */
 372                 if (color == S_WHITE)
 373                         score.value = - score.value;
 374                 if (score.value > 0)
 375                         score_step_green = round(score.value * a->score_step_byavg);
 376                 else
 377                         score_step_red = round(-score.value * a->score_step_byavg);
 378                 if (score_step_green < 0 || score_step_red > 0) {
 379                         /* The steps are in bad direction - keep still. */
 380                         return komi_by_color(extra_komi, color);
 381                 }
 382         }
 383
 384         /* Wear out the ratchet. */
 385         if (a->use_komi_ratchet && a->komi_ratchet_maxage > 0) {
 386                 a->komi_ratchet_age += value.playouts;
 387                 if (a->komi_ratchet_age > a->komi_ratchet_maxage) {
 388                         a->komi_ratchet = 1000;
 389                         a->komi_ratchet_age = 0;
 390                 }
 391         }
 392
 393         if (value.value < a->zone_red) {
 394                 /* Red zone. Take extra komi. */
 395                 if (DEBUGL(3))
 396                         fprintf(stderr, "[red] %f, step %d | komi ratchet %f age %d/%d -> %f\n",
 397                                 value.value, score_step_red, a->komi_ratchet, a->komi_ratchet_age, a->komi_ratchet_maxage, extra_komi);
 398                 if (a->losing_komi_ratchet || extra_komi > 0) {
 399                         a->komi_ratchet = extra_komi;
 400                         a->komi_ratchet_age = 0;
 401                 }
 402                 extra_komi += score_step_red;
 403                 return komi_by_color(extra_komi, color);
 404
 405         } else if (value.value < a->zone_green) {
 406                 /* Yellow zone, do nothing. */
 407                 return komi_by_color(extra_komi, color);
 408
 409         } else {
 410                 /* Green zone. Give extra komi. */
 411                 if (DEBUGL(3))
 412                         fprintf(stderr, "[green] %f, step %d | komi ratchet %f age %d/%d\n",
 413                                 value.value, score_step_green, a->komi_ratchet, a->komi_ratchet_age, a->komi_ratchet_maxage);
 414                 extra_komi += score_step_green;
 415                 if (a->use_komi_ratchet && extra_komi >= a->komi_ratchet)
 416                         extra_komi = a->komi_ratchet - 1;
 417                 return komi_by_color(extra_komi, color);
 418         }
 419 }
 420
 421 static floating_t
 422 bounded_komi(struct dynkomi_adaptive *a, struct board *b,
 423              enum stone color, floating_t komi, floating_t max_losing_komi)
 424 {
 425         /* At the end of game, disallow losing komi. */
 426         if (komi_by_color(komi, color) < 0
 427             && board_game_portion(a, b) > a->losing_komi_stop)
 428                 return 0;
 429
 430         /* Get lower bound on komi we take so that we don't underperform
 431          * too much. */
 432         floating_t min_komi = komi_by_color(- max_losing_komi, color);
 433
 434         if (komi_by_color(komi - min_komi, color) > 0)
 435                 return komi;
 436         else
 437                 return min_komi;
 438 }
 439
 440 static floating_t
 441 adaptive_permove(struct uct_dynkomi *d, struct board *b, struct tree *tree)
 442 {
 443         /* We do not use extra komi at the game end - we are not
 444          * to fool ourselves at this point. */
 445         if (board_estimated_moves_left(b) <= MIN_MOVES_LEFT) {
 446                 tree->use_extra_komi = false;
 447                 return 0;
 448         }
 449         struct dynkomi_adaptive *a = d->data;
 450         enum stone color = stone_other(tree->root_color);
 451         if (DEBUGL(3))
 452                 fprintf(stderr, "m %d/%d ekomi %f permove %f/%d\n",
 453                         b->moves, a->lead_moves, tree->extra_komi,
 454                         d->score.value, d->score.playouts);
 455
 456         if (b->moves <= a->lead_moves)
 457                 return bounded_komi(a, b, color,
 458                                     board_effective_handicap(b, 7 /* XXX */),
 459                                     a->max_losing_komi);
 460
 461         floating_t komi = a->indicator(d, b, tree, color);
 462         if (DEBUGL(3))
 463                 fprintf(stderr, "dynkomi: %f -> %f\n", tree->extra_komi, komi);
 464         return bounded_komi(a, b, color, komi, a->max_losing_komi);
 465 }
 466
 467 static floating_t
 468 adaptive_persim(struct uct_dynkomi *d, struct board *b, struct tree *tree, struct tree_node *node)
 469 {
 470         return tree->extra_komi;
 471 }
 472
 473 struct uct_dynkomi *
 474 uct_dynkomi_init_adaptive(struct uct *u, char *arg, struct board *b)
 475 {
 476         struct uct_dynkomi *d = calloc2(1, sizeof(*d));
 477         d->uct = u;
 478         d->permove = adaptive_permove;
 479         d->persim = adaptive_persim;
 480         d->done = generic_done;
 481
 482         struct dynkomi_adaptive *a = calloc2(1, sizeof(*a));
 483         d->data = a;
 484
 485         a->lead_moves = board_large(b) ? 20 : 4; // XXX
 486         a->max_losing_komi = 30;
 487         a->losing_komi_stop = 1.0f;
 488         a->indicator = komi_by_value;
 489
 490         a->adapter = adapter_sigmoid;
 491         a->adapt_rate = -18;
 492         a->adapt_phase = 0.65;
 493         a->adapt_moves = 200;
 494         a->adapt_dir = -0.5;
 495
 496         a->zone_red = 0.45;
 497         a->zone_green = 0.50;
 498         a->score_step = 1;
 499         a->use_komi_ratchet = true;
 500         a->komi_ratchet_maxage = 0;
 501         a->komi_ratchet = 1000;
 502
 503         if (arg) {
 504                 char *optspec, *next = arg;
 505                 while (*next) {
 506                         optspec = next;
 507                         next += strcspn(next, ":");
 508                         if (*next) { *next++ = 0; } else { *next = 0; }
 509
 510                         char *optname = optspec;
 511                         char *optval = strchr(optspec, '=');
 512                         if (optval) *optval++ = 0;
 513
 514                         if (!strcasecmp(optname, "lead_moves") && optval) {
 515                                 /* Do not adjust komi adaptively for first
 516                                  * N moves. */
 517                                 a->lead_moves = atoi(optval);
 518                         } else if (!strcasecmp(optname, "max_losing_komi") && optval) {
 519                                 a->max_losing_komi = atof(optval);
 520                         } else if (!strcasecmp(optname, "losing_komi_stop") && optval) {
 521                                 a->losing_komi_stop = atof(optval);
 522                         } else if (!strcasecmp(optname, "indicator")) {
 523                                 /* Adaptatation indicator - how to decide
 524                                  * the adaptation rate and direction. */
 525                                 if (!strcasecmp(optval, "value")) {
 526                                         /* Winrate w/ komi so far. */
 527                                         a->indicator = komi_by_value;
 528                                 } else if (!strcasecmp(optval, "score")) {
 529                                         /* Expected score w/ current komi. */
 530                                         a->indicator = komi_by_score;
 531                                 } else {
 532                                         fprintf(stderr, "UCT: Invalid indicator %s\n", optval);
 533                                         exit(1);
 534                                 }
 535
 536                                 /* value indicator settings */
 537                         } else if (!strcasecmp(optname, "zone_red") && optval) {
 538                                 a->zone_red = atof(optval);
 539                         } else if (!strcasecmp(optname, "zone_green") && optval) {
 540                                 a->zone_green = atof(optval);
 541                         } else if (!strcasecmp(optname, "score_step") && optval) {
 542                                 a->score_step = atoi(optval);
 543                         } else if (!strcasecmp(optname, "score_step_byavg") && optval) {
 544                                 a->score_step_byavg = atof(optval);
 545                         } else if (!strcasecmp(optname, "use_komi_ratchet")) {
 546                                 a->use_komi_ratchet = !optval || atoi(optval);
 547                         } else if (!strcasecmp(optname, "losing_komi_ratchet")) {
 548                                 a->losing_komi_ratchet = !optval || atoi(optval);
 549                         } else if (!strcasecmp(optname, "komi_ratchet_age") && optval) {
 550                                 a->komi_ratchet_maxage = atoi(optval);
 551
 552                                 /* score indicator settings */
 553                         } else if (!strcasecmp(optname, "adapter") && optval) {
 554                                 /* Adaptatation method. */
 555                                 if (!strcasecmp(optval, "sigmoid")) {
 556                                         a->adapter = adapter_sigmoid;
 557                                 } else if (!strcasecmp(optval, "linear")) {
 558                                         a->adapter = adapter_linear;
 559                                 } else {
 560                                         fprintf(stderr, "UCT: Invalid adapter %s\n", optval);
 561                                         exit(1);
 562                                 }
 563                         } else if (!strcasecmp(optname, "adapt_base") && optval) {
 564                                 /* Adaptation base rate; see above. */
 565                                 a->adapt_base = atof(optval);
 566                         } else if (!strcasecmp(optname, "adapt_rate") && optval) {
 567                                 /* Adaptation slope; see above. */
 568                                 a->adapt_rate = atof(optval);
 569                         } else if (!strcasecmp(optname, "adapt_phase") && optval) {
 570                                 /* Adaptation phase shift; see above. */
 571                                 a->adapt_phase = atof(optval);
 572                         } else if (!strcasecmp(optname, "adapt_moves") && optval) {
 573                                 /* Adaptation move amount; see above. */
 574                                 a->adapt_moves = atoi(optval);
 575                         } else if (!strcasecmp(optname, "adapt_aport")) {
 576                                 a->adapt_aport = !optval || atoi(optval);
 577                         } else if (!strcasecmp(optname, "adapt_dir") && optval) {
 578                                 /* Adaptation direction vector; see above. */
 579                                 a->adapt_dir = atof(optval);
 580
 581                         } else {
 582                                 fprintf(stderr, "uct: Invalid dynkomi argument %s or missing value\n", optname);
 583                                 exit(1);
 584                         }
 585                 }
 586         }
 587
 588         return d;
 589 }