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_permove(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
)
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
;
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
;
82 extra_komi
= floor(tree
->extra_komi
);
85 /* Do not take decisions on unstable value. */
86 if (tree
->root
->u
.playouts
< GJ_MINGAMES
) return extra_komi
;
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 if (extra_komi
< 0 && DEBUGL(3))
92 fprintf(stderr
, "XXX: extra_komi %.3f < 0 (color %s tree ek %.3f)\n", extra_komi
, stone2str(color
), tree
->extra_komi
);
93 // assert(extra_komi >= 0);
94 floating_t orig_komi
= extra_komi
;
96 if (my_value
< 0.5 && l
->komi_ratchet
> 0 && l
->komi_ratchet
!= INFINITY
) {
98 fprintf(stderr
, "losing %f extra komi %.1f ratchet %.1f -> 0\n",
99 my_value
, extra_komi
, l
->komi_ratchet
);
100 /* Disable dynkomi completely, too dangerous in this game. */
101 extra_komi
= l
->komi_ratchet
= 0;
103 } else if (my_value
< l
->orange_zone
&& extra_komi
> 0) {
104 extra_komi
= l
->komi_ratchet
= fmax(extra_komi
- l
->drop_step
, 0.0);
105 if (extra_komi
!= orig_komi
&& DEBUGL(3))
106 fprintf(stderr
, "dropping to %f, extra komi %.1f -> ratchet %.1f\n",
107 my_value
, orig_komi
, extra_komi
);
109 } else if (my_value
> l
->green_zone
&& extra_komi
+ 1 <= l
->komi_ratchet
) {
111 if (extra_komi
!= orig_komi
&& DEBUGL(3))
112 fprintf(stderr
, "winning %f extra_komi %.1f -> %.1f, ratchet %.1f\n",
113 my_value
, orig_komi
, extra_komi
, l
->komi_ratchet
);
115 return komi_by_color(extra_komi
, color
);
119 linear_persim(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
, struct tree_node
*node
)
121 struct dynkomi_linear
*l
= d
->data
;
123 return tree
->extra_komi
;
124 /* We don't reuse computed value from tree->extra_komi,
125 * since we want to use value correct for this node depth.
126 * This also means the values will stay correct after
128 return linear_permove(d
, b
, tree
);
132 uct_dynkomi_init_linear(struct uct
*u
, char *arg
, struct board
*b
)
134 struct uct_dynkomi
*d
= calloc2(1, sizeof(*d
));
136 d
->permove
= linear_permove
;
137 d
->persim
= linear_persim
;
138 d
->done
= generic_done
;
140 struct dynkomi_linear
*l
= calloc2(1, sizeof(*l
));
143 /* Force white to feel behind and try harder, but not to the
144 * point of resigning immediately in high handicap games.
145 * By move 100 white should still be behind but should have
146 * caught up enough to avoid resigning. */
147 int moves
= board_large(b
) ? 100 : 50;
148 if (!board_small(b
)) {
149 l
->moves
[S_BLACK
] = moves
;
150 l
->moves
[S_WHITE
] = moves
;
153 /* The real value of one stone is twice the komi so about 15 points.
154 * But use a lower value to avoid being too pessimistic as black
155 * or too optimistic as white. */
156 l
->handicap_value
[S_BLACK
] = 8;
157 l
->handicap_value
[S_WHITE
] = 1;
159 l
->komi_ratchet
= INFINITY
;
160 l
->green_zone
= 0.85;
161 l
->orange_zone
= 0.8;
165 char *optspec
, *next
= arg
;
168 next
+= strcspn(next
, ":");
169 if (*next
) { *next
++ = 0; } else { *next
= 0; }
171 char *optname
= optspec
;
172 char *optval
= strchr(optspec
, '=');
173 if (optval
) *optval
++ = 0;
175 if (!strcasecmp(optname
, "moves") && optval
) {
176 /* Dynamic komi in handicap game; linearly
177 * decreases to basic settings until move
178 * #optval. moves=blackmoves%whitemoves */
179 for (int i
= S_BLACK
; *optval
&& i
<= S_WHITE
; i
++) {
180 l
->moves
[i
] = atoi(optval
);
181 optval
+= strcspn(optval
, "%");
182 if (*optval
) optval
++;
184 } else if (!strcasecmp(optname
, "handicap_value") && optval
) {
185 /* Point value of single handicap stone,
186 * for dynkomi computation. */
187 for (int i
= S_BLACK
; *optval
&& i
<= S_WHITE
; i
++) {
188 l
->handicap_value
[i
] = atoi(optval
);
189 optval
+= strcspn(optval
, "%");
190 if (*optval
) optval
++;
192 } else if (!strcasecmp(optname
, "rootbased")) {
193 /* If set, the extra komi applied will be
194 * the same for all simulations within a move,
195 * instead of being same for all simulations
196 * within the tree node. */
197 l
->rootbased
= !optval
|| atoi(optval
);
198 } else if (!strcasecmp(optname
, "green_zone") && optval
) {
199 /* Increase komi when win ratio is above green_zone */
200 l
->green_zone
= atof(optval
);
201 } else if (!strcasecmp(optname
, "orange_zone") && optval
) {
202 /* Decrease komi when > 0 and win ratio is below orange_zone */
203 l
->orange_zone
= atof(optval
);
204 } else if (!strcasecmp(optname
, "drop_step") && optval
) {
205 /* Decrease komi by drop_step points */
206 l
->drop_step
= atof(optval
);
208 fprintf(stderr
, "uct: Invalid dynkomi argument %s or missing value\n", optname
);
218 /* ADAPTIVE dynkomi strategy - Adaptive Situational Compensation */
219 /* We adapt the komi based on current situation:
220 * (i) score-based: We maintain the average score outcome of our
221 * games and adjust the komi by a fractional step towards the expected
223 * (ii) value-based: While winrate is above given threshold, adjust
224 * the komi by a fixed step in the appropriate direction.
225 * These adjustments can be
226 * (a) Move-stepped, new extra komi value is always set only at the
227 * beginning of the tree search for next move;
228 * (b) Continuous, new extra komi value is periodically re-determined
229 * and adjusted throughout a single tree search. */
231 struct dynkomi_adaptive
{
232 /* Do not take measured average score into regard for
233 * first @lead_moves - the variance is just too much.
234 * (Instead, we consider the handicap-based komi provided
235 * by linear dynkomi.) */
237 /* Maximum komi to pretend the opponent to give. */
238 floating_t max_losing_komi
;
239 /* Game portion at which losing komi is not allowed anymore. */
240 floating_t losing_komi_stop
;
241 /* Turn off dynkomi at the (perceived) closing of the game
242 * (last few moves). */
243 bool no_komi_at_game_end
;
244 /* Alternative game portion determination. */
246 floating_t (*indicator
)(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
, enum stone color
);
248 /* Value-based adaptation. */
249 floating_t zone_red
, zone_green
;
251 floating_t score_step_byavg
; // use portion of average score as increment
252 bool use_komi_ratchet
;
253 bool losing_komi_ratchet
; // ratchet even losing komi
254 int komi_ratchet_maxage
;
255 // runtime, not configuration:
256 int komi_ratchet_age
;
257 floating_t komi_ratchet
;
259 /* Score-based adaptation. */
260 floating_t (*adapter
)(struct uct_dynkomi
*d
, struct board
*b
);
261 floating_t adapt_base
; // [0,1)
262 /* Sigmoid adaptation rate parameter; see below for details. */
263 floating_t adapt_phase
; // [0,1]
264 floating_t adapt_rate
; // [1,infty)
265 /* Linear adaptation rate parameter. */
267 floating_t adapt_dir
; // [-1,1]
269 #define TRUSTWORTHY_KOMI_PLAYOUTS 200
272 board_game_portion(struct dynkomi_adaptive
*a
, struct board
*b
)
274 if (!a
->adapt_aport
) {
275 int total_moves
= b
->moves
+ 2 * board_estimated_moves_left(b
);
276 return (floating_t
) b
->moves
/ total_moves
;
278 int brsize
= board_size(b
) - 2;
279 return 1.0 - (floating_t
) b
->flen
/ (brsize
* brsize
);
284 adapter_sigmoid(struct uct_dynkomi
*d
, struct board
*b
)
286 struct dynkomi_adaptive
*a
= d
->data
;
287 /* Figure out how much to adjust the komi based on the game
288 * stage. The adaptation rate is 0 at the beginning,
289 * at game stage a->adapt_phase crosses though 0.5 and
290 * approaches 1 at the game end; the slope is controlled
291 * by a->adapt_rate. */
292 floating_t game_portion
= board_game_portion(a
, b
);
293 floating_t l
= game_portion
- a
->adapt_phase
;
294 return 1.0 / (1.0 + exp(-a
->adapt_rate
* l
));
298 adapter_linear(struct uct_dynkomi
*d
, struct board
*b
)
300 struct dynkomi_adaptive
*a
= d
->data
;
301 /* Figure out how much to adjust the komi based on the game
302 * stage. We just linearly increase/decrease the adaptation
303 * rate for first N moves. */
304 if (b
->moves
> a
->adapt_moves
)
306 if (a
->adapt_dir
< 0)
307 return 1 - (- a
->adapt_dir
) * b
->moves
/ a
->adapt_moves
;
309 return a
->adapt_dir
* b
->moves
/ a
->adapt_moves
;
313 komi_by_score(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
, enum stone color
)
315 struct dynkomi_adaptive
*a
= d
->data
;
316 if (d
->score
.playouts
< TRUSTWORTHY_KOMI_PLAYOUTS
)
317 return tree
->extra_komi
;
319 struct move_stats score
= d
->score
;
320 /* Almost-reset tree->score to gather fresh stats. */
321 d
->score
.playouts
= 1;
323 /* Look at average score and push extra_komi in that direction. */
324 floating_t p
= a
->adapter(d
, b
);
325 p
= a
->adapt_base
+ p
* (1 - a
->adapt_base
);
326 if (p
> 0.9) p
= 0.9; // don't get too eager!
327 floating_t extra_komi
= tree
->extra_komi
+ p
* score
.value
;
329 fprintf(stderr
, "mC += %f * %f\n", p
, score
.value
);
334 komi_by_value(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
, enum stone color
)
336 struct dynkomi_adaptive
*a
= d
->data
;
337 if (d
->value
.playouts
< TRUSTWORTHY_KOMI_PLAYOUTS
)
338 return tree
->extra_komi
;
340 struct move_stats value
= d
->value
;
341 /* Almost-reset tree->value to gather fresh stats. */
342 d
->value
.playouts
= 1;
343 /* Correct color POV. */
344 if (color
== S_WHITE
)
345 value
.value
= 1 - value
.value
;
347 /* We have three "value zones":
348 * red zone | yellow zone | green zone
350 * red zone: reduce komi
351 * yellow zone: do not touch komi
352 * green zone: enlage komi.
354 * Also, at some point komi will be tuned in such way
355 * that it will be in green zone but increasing it will
356 * be unfeasible. Thus, we have a _ratchet_ - we will
357 * remember the last komi that has put us into the
358 * red zone, and not use it or go over it. We use the
359 * ratchet only when giving extra komi, we always want
360 * to try to reduce extra komi we take.
362 * TODO: Make the ratchet expire after a while. */
364 /* We use komi_by_color() first to normalize komi
365 * additions/subtractions, then apply it again on
366 * return value to restore original komi parity. */
367 /* Positive extra_komi means that we are _giving_
368 * komi (winning), negative extra_komi is _taking_
370 floating_t extra_komi
= komi_by_color(tree
->extra_komi
, color
);
371 int score_step_red
= -a
->score_step
;
372 int score_step_green
= a
->score_step
;
374 if (a
->score_step_byavg
!= 0) {
375 struct move_stats score
= d
->score
;
376 /* Almost-reset tree->score to gather fresh stats. */
377 d
->score
.playouts
= 1;
378 /* Correct color POV. */
379 if (color
== S_WHITE
)
380 score
.value
= - score
.value
;
382 score_step_green
= round(score
.value
* a
->score_step_byavg
);
384 score_step_red
= round(-score
.value
* a
->score_step_byavg
);
385 if (score_step_green
< 0 || score_step_red
> 0) {
386 /* The steps are in bad direction - keep still. */
387 return komi_by_color(extra_komi
, color
);
391 /* Wear out the ratchet. */
392 if (a
->use_komi_ratchet
&& a
->komi_ratchet_maxage
> 0) {
393 a
->komi_ratchet_age
+= value
.playouts
;
394 if (a
->komi_ratchet_age
> a
->komi_ratchet_maxage
) {
395 a
->komi_ratchet
= 1000;
396 a
->komi_ratchet_age
= 0;
400 if (value
.value
< a
->zone_red
) {
401 /* Red zone. Take extra komi. */
403 fprintf(stderr
, "[red] %f, step %d | komi ratchet %f age %d/%d -> %f\n",
404 value
.value
, score_step_red
, a
->komi_ratchet
, a
->komi_ratchet_age
, a
->komi_ratchet_maxage
, extra_komi
);
405 if (a
->losing_komi_ratchet
|| extra_komi
> 0) {
406 a
->komi_ratchet
= extra_komi
;
407 a
->komi_ratchet_age
= 0;
409 extra_komi
+= score_step_red
;
410 return komi_by_color(extra_komi
, color
);
412 } else if (value
.value
< a
->zone_green
) {
413 /* Yellow zone, do nothing. */
414 return komi_by_color(extra_komi
, color
);
417 /* Green zone. Give extra komi. */
419 fprintf(stderr
, "[green] %f, step %d | komi ratchet %f age %d/%d\n",
420 value
.value
, score_step_green
, a
->komi_ratchet
, a
->komi_ratchet_age
, a
->komi_ratchet_maxage
);
421 extra_komi
+= score_step_green
;
422 if (a
->use_komi_ratchet
&& extra_komi
>= a
->komi_ratchet
)
423 extra_komi
= a
->komi_ratchet
- 1;
424 return komi_by_color(extra_komi
, color
);
429 bounded_komi(struct dynkomi_adaptive
*a
, struct board
*b
,
430 enum stone color
, floating_t komi
, floating_t max_losing_komi
)
432 /* At the end of game, disallow losing komi. */
433 if (komi_by_color(komi
, color
) < 0
434 && board_game_portion(a
, b
) > a
->losing_komi_stop
)
437 /* Get lower bound on komi we take so that we don't underperform
439 floating_t min_komi
= komi_by_color(- max_losing_komi
, color
);
441 if (komi_by_color(komi
- min_komi
, color
) > 0)
448 adaptive_permove(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
)
450 struct dynkomi_adaptive
*a
= d
->data
;
451 enum stone color
= stone_other(tree
->root_color
);
453 /* We do not use extra komi at the game end - we are not
454 * to fool ourselves at this point. */
455 if (a
->no_komi_at_game_end
&& board_estimated_moves_left(b
) <= MIN_MOVES_LEFT
) {
456 tree
->use_extra_komi
= false;
461 fprintf(stderr
, "m %d/%d ekomi %f permove %f/%d\n",
462 b
->moves
, a
->lead_moves
, tree
->extra_komi
,
463 d
->score
.value
, d
->score
.playouts
);
465 if (b
->moves
<= a
->lead_moves
)
466 return bounded_komi(a
, b
, color
,
467 board_effective_handicap(b
, 7 /* XXX */),
470 floating_t komi
= a
->indicator(d
, b
, tree
, color
);
472 fprintf(stderr
, "dynkomi: %f -> %f\n", tree
->extra_komi
, komi
);
473 return bounded_komi(a
, b
, color
, komi
, a
->max_losing_komi
);
477 adaptive_persim(struct uct_dynkomi
*d
, struct board
*b
, struct tree
*tree
, struct tree_node
*node
)
479 return tree
->extra_komi
;
483 uct_dynkomi_init_adaptive(struct uct
*u
, char *arg
, struct board
*b
)
485 struct uct_dynkomi
*d
= calloc2(1, sizeof(*d
));
487 d
->permove
= adaptive_permove
;
488 d
->persim
= adaptive_persim
;
489 d
->done
= generic_done
;
491 struct dynkomi_adaptive
*a
= calloc2(1, sizeof(*a
));
494 a
->lead_moves
= board_large(b
) ? 20 : 4; // XXX
495 a
->max_losing_komi
= 30;
496 a
->losing_komi_stop
= 1.0f
;
497 a
->no_komi_at_game_end
= true;
498 a
->indicator
= komi_by_value
;
500 a
->adapter
= adapter_sigmoid
;
502 a
->adapt_phase
= 0.65;
503 a
->adapt_moves
= 200;
507 a
->zone_green
= 0.50;
509 a
->use_komi_ratchet
= true;
510 a
->komi_ratchet_maxage
= 0;
511 a
->komi_ratchet
= 1000;
514 char *optspec
, *next
= arg
;
517 next
+= strcspn(next
, ":");
518 if (*next
) { *next
++ = 0; } else { *next
= 0; }
520 char *optname
= optspec
;
521 char *optval
= strchr(optspec
, '=');
522 if (optval
) *optval
++ = 0;
524 if (!strcasecmp(optname
, "lead_moves") && optval
) {
525 /* Do not adjust komi adaptively for first
527 a
->lead_moves
= atoi(optval
);
528 } else if (!strcasecmp(optname
, "max_losing_komi") && optval
) {
529 a
->max_losing_komi
= atof(optval
);
530 } else if (!strcasecmp(optname
, "losing_komi_stop") && optval
) {
531 a
->losing_komi_stop
= atof(optval
);
532 } else if (!strcasecmp(optname
, "no_komi_at_game_end")) {
533 a
->no_komi_at_game_end
= !optval
|| atoi(optval
);
534 } else if (!strcasecmp(optname
, "indicator")) {
535 /* Adaptatation indicator - how to decide
536 * the adaptation rate and direction. */
537 if (!strcasecmp(optval
, "value")) {
538 /* Winrate w/ komi so far. */
539 a
->indicator
= komi_by_value
;
540 } else if (!strcasecmp(optval
, "score")) {
541 /* Expected score w/ current komi. */
542 a
->indicator
= komi_by_score
;
544 fprintf(stderr
, "UCT: Invalid indicator %s\n", optval
);
548 /* value indicator settings */
549 } else if (!strcasecmp(optname
, "zone_red") && optval
) {
550 a
->zone_red
= atof(optval
);
551 } else if (!strcasecmp(optname
, "zone_green") && optval
) {
552 a
->zone_green
= atof(optval
);
553 } else if (!strcasecmp(optname
, "score_step") && optval
) {
554 a
->score_step
= atoi(optval
);
555 } else if (!strcasecmp(optname
, "score_step_byavg") && optval
) {
556 a
->score_step_byavg
= atof(optval
);
557 } else if (!strcasecmp(optname
, "use_komi_ratchet")) {
558 a
->use_komi_ratchet
= !optval
|| atoi(optval
);
559 } else if (!strcasecmp(optname
, "losing_komi_ratchet")) {
560 a
->losing_komi_ratchet
= !optval
|| atoi(optval
);
561 } else if (!strcasecmp(optname
, "komi_ratchet_age") && optval
) {
562 a
->komi_ratchet_maxage
= atoi(optval
);
564 /* score indicator settings */
565 } else if (!strcasecmp(optname
, "adapter") && optval
) {
566 /* Adaptatation method. */
567 if (!strcasecmp(optval
, "sigmoid")) {
568 a
->adapter
= adapter_sigmoid
;
569 } else if (!strcasecmp(optval
, "linear")) {
570 a
->adapter
= adapter_linear
;
572 fprintf(stderr
, "UCT: Invalid adapter %s\n", optval
);
575 } else if (!strcasecmp(optname
, "adapt_base") && optval
) {
576 /* Adaptation base rate; see above. */
577 a
->adapt_base
= atof(optval
);
578 } else if (!strcasecmp(optname
, "adapt_rate") && optval
) {
579 /* Adaptation slope; see above. */
580 a
->adapt_rate
= atof(optval
);
581 } else if (!strcasecmp(optname
, "adapt_phase") && optval
) {
582 /* Adaptation phase shift; see above. */
583 a
->adapt_phase
= atof(optval
);
584 } else if (!strcasecmp(optname
, "adapt_moves") && optval
) {
585 /* Adaptation move amount; see above. */
586 a
->adapt_moves
= atoi(optval
);
587 } else if (!strcasecmp(optname
, "adapt_aport")) {
588 a
->adapt_aport
= !optval
|| atoi(optval
);
589 } else if (!strcasecmp(optname
, "adapt_dir") && optval
) {
590 /* Adaptation direction vector; see above. */
591 a
->adapt_dir
= atof(optval
);
594 fprintf(stderr
, "uct: Invalid dynkomi argument %s or missing value\n", optname
);