10 #include "tactics/util.h"
11 #include "uct/dynkomi.h"
12 #include "uct/internal.h"
17 generic_done(struct uct_dynkomi
*d
)
19 if (d
->data
) free(d
->data
);
24 /* NONE dynkomi strategy - never fiddle with komi values. */
27 uct_dynkomi_init_none(struct uct
*u
, char *arg
, struct board
*b
)
29 struct uct_dynkomi
*d
= calloc2(1, sizeof(*d
));
33 d
->done
= generic_done
;
37 fprintf(stderr
, "uct: Dynkomi method none accepts no arguments\n");
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. */
52 struct dynkomi_linear
{
53 int handicap_value
[S_MAX
];
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
;
70 linear_simple(struct dynkomi_linear
*l
, struct board
*b
, enum stone color
)
72 int lmoves
= l
->moves
[color
];
73 floating_t base_komi
= board_effective_handicap(b
, l
->handicap_value
[color
]);
74 return base_komi
* (lmoves
- b
->moves
) / lmoves
;
78 linear_permove(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
)
80 struct dynkomi_linear
*l
= d
->data
;
81 enum stone color
= d
->uct
->pondering
? tree
->root_color
: stone_other(tree
->root_color
);
82 int lmoves
= l
->moves
[color
];
84 if (b
->moves
< lmoves
)
85 return linear_simple(l
, b
, color
);
87 /* Allow simple adaptation in extreme endgame situations. */
89 floating_t extra_komi
= floor(tree
->extra_komi
);
91 /* Do not take decisions on unstable value. */
92 if (tree
->root
->u
.playouts
< GJ_MINGAMES
)
95 floating_t my_value
= tree_node_get_value(tree
, 1, tree
->root
->u
.value
);
96 /* We normalize komi as in komi_by_value(), > 0 when winning. */
97 extra_komi
= komi_by_color(extra_komi
, color
);
98 if (extra_komi
< 0 && DEBUGL(3))
99 fprintf(stderr
, "XXX: extra_komi %.3f < 0 (color %s tree ek %.3f)\n", extra_komi
, stone2str(color
), tree
->extra_komi
);
100 // assert(extra_komi >= 0);
101 floating_t orig_komi
= extra_komi
;
103 if (my_value
< 0.5 && l
->komi_ratchet
> 0 && l
->komi_ratchet
!= INFINITY
) {
105 fprintf(stderr
, "losing %f extra komi %.1f ratchet %.1f -> 0\n",
106 my_value
, extra_komi
, l
->komi_ratchet
);
107 /* Disable dynkomi completely, too dangerous in this game. */
108 extra_komi
= l
->komi_ratchet
= 0;
110 } else if (my_value
< l
->orange_zone
&& extra_komi
> 0) {
111 extra_komi
= l
->komi_ratchet
= fmax(extra_komi
- l
->drop_step
, 0.0);
112 if (extra_komi
!= orig_komi
&& DEBUGL(3))
113 fprintf(stderr
, "dropping to %f, extra komi %.1f -> %.1f\n",
114 my_value
, orig_komi
, extra_komi
);
116 } else if (my_value
> l
->green_zone
&& extra_komi
+ 1 <= l
->komi_ratchet
) {
118 if (extra_komi
!= orig_komi
&& DEBUGL(3))
119 fprintf(stderr
, "winning %f extra_komi %.1f -> %.1f, ratchet %.1f\n",
120 my_value
, orig_komi
, extra_komi
, l
->komi_ratchet
);
122 return komi_by_color(extra_komi
, color
);
126 linear_persim(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
, struct tree_node
*node
)
128 struct dynkomi_linear
*l
= d
->data
;
130 return tree
->extra_komi
;
132 /* We don't reuse computed value from tree->extra_komi,
133 * since we want to use value correct for this node depth.
134 * This also means the values will stay correct after
137 enum stone color
= d
->uct
->pondering
? tree
->root_color
: stone_other(tree
->root_color
);
138 int lmoves
= l
->moves
[color
];
139 if (b
->moves
< lmoves
)
140 return linear_simple(l
, b
, color
);
141 return tree
->extra_komi
;
145 uct_dynkomi_init_linear(struct uct
*u
, char *arg
, struct board
*b
)
147 struct uct_dynkomi
*d
= calloc2(1, sizeof(*d
));
149 d
->permove
= linear_permove
;
150 d
->persim
= linear_persim
;
151 d
->done
= generic_done
;
153 struct dynkomi_linear
*l
= calloc2(1, sizeof(*l
));
156 /* Force white to feel behind and try harder, but not to the
157 * point of resigning immediately in high handicap games.
158 * By move 100 white should still be behind but should have
159 * caught up enough to avoid resigning. */
160 int moves
= board_large(b
) ? 100 : 50;
161 if (!board_small(b
)) {
162 l
->moves
[S_BLACK
] = moves
;
163 l
->moves
[S_WHITE
] = moves
;
166 /* The real value of one stone is twice the komi so about 15 points.
167 * But use a lower value to avoid being too pessimistic as black
168 * or too optimistic as white. */
169 l
->handicap_value
[S_BLACK
] = 8;
170 l
->handicap_value
[S_WHITE
] = 1;
172 l
->komi_ratchet
= INFINITY
;
173 l
->green_zone
= 0.85;
174 l
->orange_zone
= 0.8;
178 char *optspec
, *next
= arg
;
181 next
+= strcspn(next
, ":");
182 if (*next
) { *next
++ = 0; } else { *next
= 0; }
184 char *optname
= optspec
;
185 char *optval
= strchr(optspec
, '=');
186 if (optval
) *optval
++ = 0;
188 if (!strcasecmp(optname
, "moves") && optval
) {
189 /* Dynamic komi in handicap game; linearly
190 * decreases to basic settings until move
191 * #optval. moves=blackmoves%whitemoves */
192 for (int i
= S_BLACK
; *optval
&& i
<= S_WHITE
; i
++) {
193 l
->moves
[i
] = atoi(optval
);
194 optval
+= strcspn(optval
, "%");
195 if (*optval
) optval
++;
197 } else if (!strcasecmp(optname
, "handicap_value") && optval
) {
198 /* Point value of single handicap stone,
199 * for dynkomi computation. */
200 for (int i
= S_BLACK
; *optval
&& i
<= S_WHITE
; i
++) {
201 l
->handicap_value
[i
] = atoi(optval
);
202 optval
+= strcspn(optval
, "%");
203 if (*optval
) optval
++;
205 } else if (!strcasecmp(optname
, "rootbased")) {
206 /* If set, the extra komi applied will be
207 * the same for all simulations within a move,
208 * instead of being same for all simulations
209 * within the tree node. */
210 l
->rootbased
= !optval
|| atoi(optval
);
211 } else if (!strcasecmp(optname
, "green_zone") && optval
) {
212 /* Increase komi when win ratio is above green_zone */
213 l
->green_zone
= atof(optval
);
214 } else if (!strcasecmp(optname
, "orange_zone") && optval
) {
215 /* Decrease komi when > 0 and win ratio is below orange_zone */
216 l
->orange_zone
= atof(optval
);
217 } else if (!strcasecmp(optname
, "drop_step") && optval
) {
218 /* Decrease komi by drop_step points */
219 l
->drop_step
= atof(optval
);
221 fprintf(stderr
, "uct: Invalid dynkomi argument %s or missing value\n", optname
);
231 /* ADAPTIVE dynkomi strategy - Adaptive Situational Compensation */
232 /* We adapt the komi based on current situation:
233 * (i) score-based: We maintain the average score outcome of our
234 * games and adjust the komi by a fractional step towards the expected
236 * (ii) value-based: While winrate is above given threshold, adjust
237 * the komi by a fixed step in the appropriate direction.
238 * These adjustments can be
239 * (a) Move-stepped, new extra komi value is always set only at the
240 * beginning of the tree search for next move;
241 * (b) Continuous, new extra komi value is periodically re-determined
242 * and adjusted throughout a single tree search. */
244 struct dynkomi_adaptive
{
245 /* Do not take measured average score into regard for
246 * first @lead_moves - the variance is just too much.
247 * (Instead, we consider the handicap-based komi provided
248 * by linear dynkomi.) */
250 /* Maximum komi to pretend the opponent to give. */
251 floating_t max_losing_komi
;
252 /* Game portion at which losing komi is not allowed anymore. */
253 floating_t losing_komi_stop
;
254 /* Turn off dynkomi at the (perceived) closing of the game
255 * (last few moves). */
256 bool no_komi_at_game_end
;
257 /* Alternative game portion determination. */
259 floating_t (*indicator
)(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
, enum stone color
);
261 /* Value-based adaptation. */
262 floating_t zone_red
, zone_green
;
264 floating_t score_step_byavg
; // use portion of average score as increment
265 bool use_komi_ratchet
;
266 bool losing_komi_ratchet
; // ratchet even losing komi
267 int komi_ratchet_maxage
;
268 // runtime, not configuration:
269 int komi_ratchet_age
;
270 floating_t komi_ratchet
;
272 /* Score-based adaptation. */
273 floating_t (*adapter
)(struct uct_dynkomi
*d
, struct board
*b
);
274 floating_t adapt_base
; // [0,1)
275 /* Sigmoid adaptation rate parameter; see below for details. */
276 floating_t adapt_phase
; // [0,1]
277 floating_t adapt_rate
; // [1,infty)
278 /* Linear adaptation rate parameter. */
280 floating_t adapt_dir
; // [-1,1]
282 #define TRUSTWORTHY_KOMI_PLAYOUTS 200
285 board_game_portion(struct dynkomi_adaptive
*a
, struct board
*b
)
287 if (!a
->adapt_aport
) {
288 int total_moves
= b
->moves
+ 2 * board_estimated_moves_left(b
);
289 return (floating_t
) b
->moves
/ total_moves
;
291 int brsize
= board_size(b
) - 2;
292 return 1.0 - (floating_t
) b
->flen
/ (brsize
* brsize
);
297 adapter_sigmoid(struct uct_dynkomi
*d
, struct board
*b
)
299 struct dynkomi_adaptive
*a
= d
->data
;
300 /* Figure out how much to adjust the komi based on the game
301 * stage. The adaptation rate is 0 at the beginning,
302 * at game stage a->adapt_phase crosses though 0.5 and
303 * approaches 1 at the game end; the slope is controlled
304 * by a->adapt_rate. */
305 floating_t game_portion
= board_game_portion(a
, b
);
306 floating_t l
= game_portion
- a
->adapt_phase
;
307 return 1.0 / (1.0 + exp(-a
->adapt_rate
* l
));
311 adapter_linear(struct uct_dynkomi
*d
, struct board
*b
)
313 struct dynkomi_adaptive
*a
= d
->data
;
314 /* Figure out how much to adjust the komi based on the game
315 * stage. We just linearly increase/decrease the adaptation
316 * rate for first N moves. */
317 if (b
->moves
> a
->adapt_moves
)
319 if (a
->adapt_dir
< 0)
320 return 1 - (- a
->adapt_dir
) * b
->moves
/ a
->adapt_moves
;
322 return a
->adapt_dir
* b
->moves
/ a
->adapt_moves
;
326 komi_by_score(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
, enum stone color
)
328 struct dynkomi_adaptive
*a
= d
->data
;
329 if (d
->score
.playouts
< TRUSTWORTHY_KOMI_PLAYOUTS
)
330 return tree
->extra_komi
;
332 struct move_stats score
= d
->score
;
333 /* Almost-reset tree->score to gather fresh stats. */
334 d
->score
.playouts
= 1;
336 /* Look at average score and push extra_komi in that direction. */
337 floating_t p
= a
->adapter(d
, b
);
338 p
= a
->adapt_base
+ p
* (1 - a
->adapt_base
);
339 if (p
> 0.9) p
= 0.9; // don't get too eager!
340 floating_t extra_komi
= tree
->extra_komi
+ p
* score
.value
;
342 fprintf(stderr
, "mC += %f * %f\n", p
, score
.value
);
347 komi_by_value(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
, enum stone color
)
349 struct dynkomi_adaptive
*a
= d
->data
;
350 if (d
->value
.playouts
< TRUSTWORTHY_KOMI_PLAYOUTS
)
351 return tree
->extra_komi
;
353 struct move_stats value
= d
->value
;
354 /* Almost-reset tree->value to gather fresh stats. */
355 d
->value
.playouts
= 1;
356 /* Correct color POV. */
357 if (color
== S_WHITE
)
358 value
.value
= 1 - value
.value
;
360 /* We have three "value zones":
361 * red zone | yellow zone | green zone
363 * red zone: reduce komi
364 * yellow zone: do not touch komi
365 * green zone: enlage komi.
367 * Also, at some point komi will be tuned in such way
368 * that it will be in green zone but increasing it will
369 * be unfeasible. Thus, we have a _ratchet_ - we will
370 * remember the last komi that has put us into the
371 * red zone, and not use it or go over it. We use the
372 * ratchet only when giving extra komi, we always want
373 * to try to reduce extra komi we take.
375 * TODO: Make the ratchet expire after a while. */
377 /* We use komi_by_color() first to normalize komi
378 * additions/subtractions, then apply it again on
379 * return value to restore original komi parity. */
380 /* Positive extra_komi means that we are _giving_
381 * komi (winning), negative extra_komi is _taking_
383 floating_t extra_komi
= komi_by_color(tree
->extra_komi
, color
);
384 int score_step_red
= -a
->score_step
;
385 int score_step_green
= a
->score_step
;
387 if (a
->score_step_byavg
!= 0) {
388 struct move_stats score
= d
->score
;
389 /* Almost-reset tree->score to gather fresh stats. */
390 d
->score
.playouts
= 1;
391 /* Correct color POV. */
392 if (color
== S_WHITE
)
393 score
.value
= - score
.value
;
395 score_step_green
= round(score
.value
* a
->score_step_byavg
);
397 score_step_red
= round(-score
.value
* a
->score_step_byavg
);
398 if (score_step_green
< 0 || score_step_red
> 0) {
399 /* The steps are in bad direction - keep still. */
400 return komi_by_color(extra_komi
, color
);
404 /* Wear out the ratchet. */
405 if (a
->use_komi_ratchet
&& a
->komi_ratchet_maxage
> 0) {
406 a
->komi_ratchet_age
+= value
.playouts
;
407 if (a
->komi_ratchet_age
> a
->komi_ratchet_maxage
) {
408 a
->komi_ratchet
= 1000;
409 a
->komi_ratchet_age
= 0;
413 if (value
.value
< a
->zone_red
) {
414 /* Red zone. Take extra komi. */
416 fprintf(stderr
, "[red] %f, step %d | komi ratchet %f age %d/%d -> %f\n",
417 value
.value
, score_step_red
, a
->komi_ratchet
, a
->komi_ratchet_age
, a
->komi_ratchet_maxage
, extra_komi
);
418 if (a
->losing_komi_ratchet
|| extra_komi
> 0) {
419 a
->komi_ratchet
= extra_komi
;
420 a
->komi_ratchet_age
= 0;
422 extra_komi
+= score_step_red
;
423 return komi_by_color(extra_komi
, color
);
425 } else if (value
.value
< a
->zone_green
) {
426 /* Yellow zone, do nothing. */
427 return komi_by_color(extra_komi
, color
);
430 /* Green zone. Give extra komi. */
432 fprintf(stderr
, "[green] %f, step %d | komi ratchet %f age %d/%d\n",
433 value
.value
, score_step_green
, a
->komi_ratchet
, a
->komi_ratchet_age
, a
->komi_ratchet_maxage
);
434 extra_komi
+= score_step_green
;
435 if (a
->use_komi_ratchet
&& extra_komi
>= a
->komi_ratchet
)
436 extra_komi
= a
->komi_ratchet
- 1;
437 return komi_by_color(extra_komi
, color
);
442 bounded_komi(struct dynkomi_adaptive
*a
, struct board
*b
,
443 enum stone color
, floating_t komi
, floating_t max_losing_komi
)
445 /* At the end of game, disallow losing komi. */
446 if (komi_by_color(komi
, color
) < 0
447 && board_game_portion(a
, b
) > a
->losing_komi_stop
)
450 /* Get lower bound on komi we take so that we don't underperform
452 floating_t min_komi
= komi_by_color(- max_losing_komi
, color
);
454 if (komi_by_color(komi
- min_komi
, color
) > 0)
461 adaptive_permove(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
)
463 struct dynkomi_adaptive
*a
= d
->data
;
464 enum stone color
= stone_other(tree
->root_color
);
466 /* We do not use extra komi at the game end - we are not
467 * to fool ourselves at this point. */
468 if (a
->no_komi_at_game_end
&& board_estimated_moves_left(b
) <= MIN_MOVES_LEFT
) {
469 tree
->use_extra_komi
= false;
474 fprintf(stderr
, "m %d/%d ekomi %f permove %f/%d\n",
475 b
->moves
, a
->lead_moves
, tree
->extra_komi
,
476 d
->score
.value
, d
->score
.playouts
);
478 if (b
->moves
<= a
->lead_moves
)
479 return bounded_komi(a
, b
, color
,
480 board_effective_handicap(b
, 7 /* XXX */),
483 floating_t komi
= a
->indicator(d
, b
, tree
, color
);
485 fprintf(stderr
, "dynkomi: %f -> %f\n", tree
->extra_komi
, komi
);
486 return bounded_komi(a
, b
, color
, komi
, a
->max_losing_komi
);
490 adaptive_persim(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
, struct tree_node
*node
)
492 return tree
->extra_komi
;
496 uct_dynkomi_init_adaptive(struct uct
*u
, char *arg
, struct board
*b
)
498 struct uct_dynkomi
*d
= calloc2(1, sizeof(*d
));
500 d
->permove
= adaptive_permove
;
501 d
->persim
= adaptive_persim
;
502 d
->done
= generic_done
;
504 struct dynkomi_adaptive
*a
= calloc2(1, sizeof(*a
));
507 a
->lead_moves
= board_large(b
) ? 20 : 4; // XXX
508 a
->max_losing_komi
= 30;
509 a
->losing_komi_stop
= 1.0f
;
510 a
->no_komi_at_game_end
= true;
511 a
->indicator
= komi_by_value
;
513 a
->adapter
= adapter_sigmoid
;
515 a
->adapt_phase
= 0.65;
516 a
->adapt_moves
= 200;
520 a
->zone_green
= 0.50;
522 a
->use_komi_ratchet
= true;
523 a
->komi_ratchet_maxage
= 0;
524 a
->komi_ratchet
= 1000;
527 char *optspec
, *next
= arg
;
530 next
+= strcspn(next
, ":");
531 if (*next
) { *next
++ = 0; } else { *next
= 0; }
533 char *optname
= optspec
;
534 char *optval
= strchr(optspec
, '=');
535 if (optval
) *optval
++ = 0;
537 if (!strcasecmp(optname
, "lead_moves") && optval
) {
538 /* Do not adjust komi adaptively for first
540 a
->lead_moves
= atoi(optval
);
541 } else if (!strcasecmp(optname
, "max_losing_komi") && optval
) {
542 a
->max_losing_komi
= atof(optval
);
543 } else if (!strcasecmp(optname
, "losing_komi_stop") && optval
) {
544 a
->losing_komi_stop
= atof(optval
);
545 } else if (!strcasecmp(optname
, "no_komi_at_game_end")) {
546 a
->no_komi_at_game_end
= !optval
|| atoi(optval
);
547 } else if (!strcasecmp(optname
, "indicator")) {
548 /* Adaptatation indicator - how to decide
549 * the adaptation rate and direction. */
550 if (!strcasecmp(optval
, "value")) {
551 /* Winrate w/ komi so far. */
552 a
->indicator
= komi_by_value
;
553 } else if (!strcasecmp(optval
, "score")) {
554 /* Expected score w/ current komi. */
555 a
->indicator
= komi_by_score
;
557 fprintf(stderr
, "UCT: Invalid indicator %s\n", optval
);
561 /* value indicator settings */
562 } else if (!strcasecmp(optname
, "zone_red") && optval
) {
563 a
->zone_red
= atof(optval
);
564 } else if (!strcasecmp(optname
, "zone_green") && optval
) {
565 a
->zone_green
= atof(optval
);
566 } else if (!strcasecmp(optname
, "score_step") && optval
) {
567 a
->score_step
= atoi(optval
);
568 } else if (!strcasecmp(optname
, "score_step_byavg") && optval
) {
569 a
->score_step_byavg
= atof(optval
);
570 } else if (!strcasecmp(optname
, "use_komi_ratchet")) {
571 a
->use_komi_ratchet
= !optval
|| atoi(optval
);
572 } else if (!strcasecmp(optname
, "losing_komi_ratchet")) {
573 a
->losing_komi_ratchet
= !optval
|| atoi(optval
);
574 } else if (!strcasecmp(optname
, "komi_ratchet_age") && optval
) {
575 a
->komi_ratchet_maxage
= atoi(optval
);
577 /* score indicator settings */
578 } else if (!strcasecmp(optname
, "adapter") && optval
) {
579 /* Adaptatation method. */
580 if (!strcasecmp(optval
, "sigmoid")) {
581 a
->adapter
= adapter_sigmoid
;
582 } else if (!strcasecmp(optval
, "linear")) {
583 a
->adapter
= adapter_linear
;
585 fprintf(stderr
, "UCT: Invalid adapter %s\n", optval
);
588 } else if (!strcasecmp(optname
, "adapt_base") && optval
) {
589 /* Adaptation base rate; see above. */
590 a
->adapt_base
= atof(optval
);
591 } else if (!strcasecmp(optname
, "adapt_rate") && optval
) {
592 /* Adaptation slope; see above. */
593 a
->adapt_rate
= atof(optval
);
594 } else if (!strcasecmp(optname
, "adapt_phase") && optval
) {
595 /* Adaptation phase shift; see above. */
596 a
->adapt_phase
= atof(optval
);
597 } else if (!strcasecmp(optname
, "adapt_moves") && optval
) {
598 /* Adaptation move amount; see above. */
599 a
->adapt_moves
= atoi(optval
);
600 } else if (!strcasecmp(optname
, "adapt_aport")) {
601 a
->adapt_aport
= !optval
|| atoi(optval
);
602 } else if (!strcasecmp(optname
, "adapt_dir") && optval
) {
603 /* Adaptation direction vector; see above. */
604 a
->adapt_dir
= atof(optval
);
607 fprintf(stderr
, "uct: Invalid dynkomi argument %s or missing value\n", optname
);