14 #include "playout/moggy.h"
15 #include "playout/old.h"
16 #include "playout/light.h"
18 #include "uct/internal.h"
22 struct uct_policy
*policy_ucb1_init(struct uct
*u
, char *arg
);
23 struct uct_policy
*policy_ucb1tuned_init(struct uct
*u
, char *arg
);
24 struct uct_policy
*policy_ucb1amaf_init(struct uct
*u
, char *arg
);
27 #define MC_GAMES 80000
28 #define MC_GAMELEN 400
32 progress_status(struct uct
*u
, struct tree
*t
, enum stone color
, int playouts
)
38 struct tree_node
*best
= u
->policy
->choose(u
->policy
, t
->root
, t
->board
, color
);
40 fprintf(stderr
, "... No moves left\n");
43 fprintf(stderr
, "[%d] ", playouts
);
44 fprintf(stderr
, "best %f ", best
->u
.value
);
47 fprintf(stderr
, "deepest % 2d ", t
->max_depth
- t
->root
->depth
);
50 fprintf(stderr
, "| seq ");
51 for (int depth
= 0; depth
< 6; depth
++) {
52 if (best
&& best
->u
.playouts
>= 25) {
53 fprintf(stderr
, "%3s ", coord2sstr(best
->coord
, t
->board
));
54 best
= u
->policy
->choose(u
->policy
, best
, t
->board
, color
);
61 fprintf(stderr
, "| can ");
63 struct tree_node
*can
[cans
];
64 memset(can
, 0, sizeof(can
));
65 best
= t
->root
->children
;
68 while ((!can
[c
] || best
->u
.playouts
> can
[c
]->u
.playouts
) && ++c
< cans
);
69 for (int d
= 0; d
< c
; d
++) can
[d
] = can
[d
+ 1];
70 if (c
> 0) can
[c
- 1] = best
;
75 fprintf(stderr
, "%3s(%.3f) ", coord2sstr(can
[cans
]->coord
, t
->board
), can
[cans
]->u
.value
);
81 fprintf(stderr
, "\n");
86 uct_playout(struct uct
*u
, struct board
*b
, enum stone color
, struct tree
*t
)
91 struct playout_amafmap
*amaf
= NULL
;
92 if (u
->policy
->wants_amaf
) {
93 amaf
= calloc(1, sizeof(*amaf
));
94 amaf
->map
= calloc(board_size2(&b2
) + 1, sizeof(*amaf
->map
));
95 amaf
->map
++; // -1 is pass
98 /* Walk the tree until we find a leaf, then expand it and do
99 * a random playout. */
100 struct tree_node
*n
= t
->root
;
101 enum stone orig_color
= color
;
103 int pass_limit
= (board_size(&b2
) - 2) * (board_size(&b2
) - 2) / 2;
104 int passes
= is_pass(b
->last_move
.coord
);
106 fprintf(stderr
, "--- UCT walk with color %d\n", color
);
107 for (; pass
; color
= stone_other(color
)) {
108 if (tree_leaf_node(n
)) {
109 if (n
->u
.playouts
>= u
->expand_p
)
110 tree_expand_node(t
, n
, &b2
, color
, u
->radar_d
, u
->policy
, (color
== orig_color
? 1 : -1));
112 result
= play_random_game(&b2
, color
, u
->gamelen
, u
->playout_amaf
? amaf
: NULL
, u
->playout
);
113 if (orig_color
!= color
&& result
>= 0)
116 fprintf(stderr
, "[%d..%d] %s random playout result %d\n", orig_color
, color
, coord2sstr(n
->coord
, t
->board
), result
);
118 /* Reset color to the @n color. */
119 color
= stone_other(color
);
123 n
= u
->policy
->descend(u
->policy
, t
, n
, (color
== orig_color
? 1 : -1), pass_limit
);
124 assert(n
== t
->root
|| n
->parent
);
126 fprintf(stderr
, "-- UCT sent us to [%s] %f\n", coord2sstr(n
->coord
, t
->board
), n
->u
.value
);
127 if (amaf
&& n
->coord
>= -1)
128 amaf
->map
[n
->coord
] = color
;
129 struct move m
= { n
->coord
, color
};
130 int res
= board_play(&b2
, &m
);
132 if (res
< 0 || (!is_pass(m
.coord
) && !group_at(&b2
, m
.coord
)) /* suicide */
133 || b2
.superko_violation
) {
135 for (struct tree_node
*ni
= n
; ni
; ni
= ni
->parent
)
136 fprintf(stderr
, "%s ", coord2sstr(ni
->coord
, t
->board
));
137 fprintf(stderr
, "deleting invalid %s node %d,%d res %d group %d spk %d\n",
138 stone2str(color
), coord_x(n
->coord
,b
), coord_y(n
->coord
,b
),
139 res
, group_at(&b2
, m
.coord
), b2
.superko_violation
);
141 tree_delete_node(t
, n
);
146 if (is_pass(n
->coord
)) {
149 float score
= board_official_score(&b2
);
150 result
= (orig_color
== S_BLACK
) ? score
< 0 : score
> 0;
152 fprintf(stderr
, "[%d..%d] %s p-p scoring playout result %d (W %f)\n", orig_color
, color
, coord2sstr(n
->coord
, t
->board
), result
, score
);
154 board_print(&b2
, stderr
);
162 assert(n
== t
->root
|| n
->parent
);
164 u
->policy
->update(u
->policy
, t
, n
, color
, amaf
, result
);
171 board_done_noalloc(&b2
);
176 prepare_move(struct engine
*e
, struct board
*b
, enum stone color
, coord_t promote
)
178 struct uct
*u
= e
->data
;
180 if (!b
->moves
&& u
->t
) {
181 /* Stale state from last game */
187 u
->t
= tree_init(b
, color
);
188 //board_print(b, stderr);
189 tree_load(u
->t
, b
, color
);
192 /* XXX: We hope that the opponent didn't suddenly play
193 * several moves in the row. */
194 if (!is_resign(promote
) && !tree_promote_at(u
->t
, b
, promote
)) {
195 fprintf(stderr
, "CANNOT FIND NODE TO PROMOTE!\n");
198 u
->t
= tree_init(b
, color
);
202 /* Set in main thread in case the playouts should stop. */
203 static volatile sig_atomic_t halt
= 0;
206 uct_playouts(struct uct
*u
, struct board
*b
, enum stone color
, struct tree
*t
)
208 int i
, games
= u
->games
- (t
->root
->u
.playouts
/ 1.5);
209 for (i
= 0; i
< games
; i
++) {
210 int result
= uct_playout(u
, b
, color
, t
);
212 /* Tree descent has hit invalid move. */
216 if (i
> 0 && !(i
% 10000)) {
217 progress_status(u
, t
, color
, i
);
220 if (i
> 0 && !(i
% 500)) {
221 struct tree_node
*best
= u
->policy
->choose(u
->policy
, t
->root
, b
, color
);
222 if (best
&& best
->u
.playouts
>= 1500 && best
->u
.value
>= u
->loss_threshold
)
228 fprintf(stderr
, "<halting early, %d games skipped>\n", games
- i
);
233 progress_status(u
, t
, color
, i
);
235 tree_dump(t
, u
->dumpthres
);
239 static pthread_mutex_t finish_mutex
= PTHREAD_MUTEX_INITIALIZER
;
240 static pthread_cond_t finish_cond
= PTHREAD_COND_INITIALIZER
;
241 static volatile int finish_thread
;
242 static pthread_mutex_t finish_serializer
= PTHREAD_MUTEX_INITIALIZER
;
255 spawn_helper(void *ctx_
)
257 struct spawn_ctx
*ctx
= ctx_
;
259 fast_srandom(ctx
->seed
);
261 ctx
->games
= uct_playouts(ctx
->u
, ctx
->b
, ctx
->color
, ctx
->t
);
263 pthread_mutex_lock(&finish_serializer
);
264 pthread_mutex_lock(&finish_mutex
);
265 finish_thread
= ctx
->tid
;
266 pthread_cond_signal(&finish_cond
);
267 pthread_mutex_unlock(&finish_mutex
);
272 uct_notify_play(struct engine
*e
, struct board
*b
, struct move
*m
)
274 prepare_move(e
, b
, stone_other(m
->color
), m
->coord
);
278 uct_genmove(struct engine
*e
, struct board
*b
, enum stone color
)
280 struct uct
*u
= e
->data
;
283 prepare_move(e
, b
, color
, resign
);
285 int played_games
= 0;
287 played_games
= uct_playouts(u
, b
, color
, u
->t
);
289 pthread_t threads
[u
->threads
];
292 pthread_mutex_lock(&finish_mutex
);
293 /* Spawn threads... */
294 for (int ti
= 0; ti
< u
->threads
; ti
++) {
295 struct spawn_ctx
*ctx
= malloc(sizeof(*ctx
));
296 ctx
->u
= u
; ctx
->b
= b
; ctx
->color
= color
;
297 ctx
->t
= tree_copy(u
->t
); ctx
->tid
= ti
;
298 ctx
->seed
= fast_random(65536) + ti
;
299 pthread_create(&threads
[ti
], NULL
, spawn_helper
, ctx
);
301 fprintf(stderr
, "Spawned thread %d\n", ti
);
303 /* ...and collect them back: */
304 while (joined
< u
->threads
) {
305 /* Wait for some thread to finish... */
306 pthread_cond_wait(&finish_cond
, &finish_mutex
);
307 /* ...and gather its remnants. */
308 struct spawn_ctx
*ctx
;
309 pthread_join(threads
[finish_thread
], (void **) &ctx
);
310 played_games
+= ctx
->games
;
312 tree_merge(u
->t
, ctx
->t
);
316 fprintf(stderr
, "Joined thread %d\n", finish_thread
);
317 /* Do not get stalled by slow threads. */
318 if (joined
>= u
->threads
/ 2)
320 pthread_mutex_unlock(&finish_serializer
);
322 pthread_mutex_unlock(&finish_mutex
);
326 tree_dump(u
->t
, u
->dumpthres
);
328 struct tree_node
*best
= u
->policy
->choose(u
->policy
, u
->t
->root
, b
, color
);
330 tree_done(u
->t
); u
->t
= NULL
;
331 return coord_copy(pass
);
334 fprintf(stderr
, "*** WINNER is %s (%d,%d) with score %1.4f (%d/%d:%d games)\n", coord2sstr(best
->coord
, b
), coord_x(best
->coord
, b
), coord_y(best
->coord
, b
), best
->u
.value
, best
->u
.playouts
, u
->t
->root
->u
.playouts
, played_games
);
335 if (best
->u
.value
< u
->resign_ratio
&& !is_pass(best
->coord
)) {
336 tree_done(u
->t
); u
->t
= NULL
;
337 return coord_copy(resign
);
339 tree_promote_node(u
->t
, best
);
340 return coord_copy(best
->coord
);
344 uct_genbook(struct engine
*e
, struct board
*b
, enum stone color
)
346 struct uct
*u
= e
->data
;
347 u
->t
= tree_init(b
, color
);
348 tree_load(u
->t
, b
, color
);
351 for (i
= 0; i
< u
->games
; i
++) {
352 int result
= uct_playout(u
, b
, color
, u
->t
);
354 /* Tree descent has hit invalid move. */
358 if (i
> 0 && !(i
% 10000)) {
359 progress_status(u
, u
->t
, color
, i
);
362 progress_status(u
, u
->t
, color
, i
);
364 tree_save(u
->t
, b
, u
->games
/ 100);
372 uct_dumpbook(struct engine
*e
, struct board
*b
, enum stone color
)
374 struct uct
*u
= e
->data
;
375 u
->t
= tree_init(b
, color
);
376 tree_load(u
->t
, b
, color
);
383 uct_state_init(char *arg
)
385 struct uct
*u
= calloc(1, sizeof(struct uct
));
389 u
->gamelen
= MC_GAMELEN
;
392 u
->playout_amaf
= true;
395 char *optspec
, *next
= arg
;
398 next
+= strcspn(next
, ",");
399 if (*next
) { *next
++ = 0; } else { *next
= 0; }
401 char *optname
= optspec
;
402 char *optval
= strchr(optspec
, '=');
403 if (optval
) *optval
++ = 0;
405 if (!strcasecmp(optname
, "debug")) {
407 u
->debug_level
= atoi(optval
);
410 } else if (!strcasecmp(optname
, "games") && optval
) {
411 u
->games
= atoi(optval
);
412 } else if (!strcasecmp(optname
, "gamelen") && optval
) {
413 u
->gamelen
= atoi(optval
);
414 } else if (!strcasecmp(optname
, "expand_p") && optval
) {
415 u
->expand_p
= atoi(optval
);
416 } else if (!strcasecmp(optname
, "radar_d") && optval
) {
417 /* For 19x19, it is good idea to set this to 3. */
418 u
->radar_d
= atoi(optval
);
419 } else if (!strcasecmp(optname
, "dumpthres") && optval
) {
420 u
->dumpthres
= atoi(optval
);
421 } else if (!strcasecmp(optname
, "playout_amaf")) {
422 /* Whether to include random playout moves in
423 * AMAF as well. (Otherwise, only tree moves
424 * are included in AMAF. Of course makes sense
425 * only in connection with an AMAF policy.) */
426 /* with-without: 55.5% (+-4.1) */
427 if (optval
&& *optval
== '0')
428 u
->playout_amaf
= false;
429 } else if (!strcasecmp(optname
, "policy") && optval
) {
430 char *policyarg
= strchr(optval
, ':');
433 if (!strcasecmp(optval
, "ucb1")) {
434 u
->policy
= policy_ucb1_init(u
, policyarg
);
435 } else if (!strcasecmp(optval
, "ucb1tuned")) {
436 u
->policy
= policy_ucb1tuned_init(u
, policyarg
);
437 } else if (!strcasecmp(optval
, "ucb1amaf")) {
438 u
->policy
= policy_ucb1amaf_init(u
, policyarg
);
440 fprintf(stderr
, "UCT: Invalid tree policy %s\n", optval
);
442 } else if (!strcasecmp(optname
, "playout") && optval
) {
443 char *playoutarg
= strchr(optval
, ':');
446 if (!strcasecmp(optval
, "old")) {
447 u
->playout
= playout_old_init(playoutarg
);
448 } else if (!strcasecmp(optval
, "moggy")) {
449 u
->playout
= playout_moggy_init(playoutarg
);
450 } else if (!strcasecmp(optval
, "light")) {
451 u
->playout
= playout_light_init(playoutarg
);
453 fprintf(stderr
, "UCT: Invalid playout policy %s\n", optval
);
455 } else if (!strcasecmp(optname
, "threads") && optval
) {
456 u
->threads
= atoi(optval
);
458 fprintf(stderr
, "uct: Invalid engine argument %s or missing value\n", optname
);
463 u
->resign_ratio
= 0.2; /* Resign when most games are lost. */
464 u
->loss_threshold
= 0.85; /* Stop reading if after at least 1500 playouts this is best value. */
466 u
->policy
= policy_ucb1amaf_init(u
, NULL
);
469 u
->playout
= playout_moggy_init(NULL
);
470 u
->playout
->debug_level
= u
->debug_level
;
477 engine_uct_init(char *arg
)
479 struct uct
*u
= uct_state_init(arg
);
480 struct engine
*e
= calloc(1, sizeof(struct engine
));
481 e
->name
= "UCT Engine";
482 e
->comment
= "I'm playing UCT. When we both pass, I will consider all the stones on the board alive. If you are reading this, write 'yes'. Please bear with me at the game end, I need to fill the whole board; if you help me, we will both be happier. Filling the board will not lose points (NZ rules).";
483 e
->genmove
= uct_genmove
;
484 e
->notify_play
= uct_notify_play
;