12 #include "tactics/util.h"
13 #include "uct/internal.h"
15 #include "uct/policy/generic.h"
17 /* This implements the UCB1 policy with an extra AMAF heuristics. */
19 struct ucb1_policy_amaf
{
20 /* This is what the Modification of UCT with Patterns in Monte Carlo Go
21 * paper calls 'p'. Original UCB has this on 2, but this seems to
22 * produce way too wide searches; reduce this to get deeper and
23 * narrower readouts - try 0.2. */
25 /* In distributed mode, encourage different slaves to work on different
26 * parts of the tree by adding virtual wins to different nodes. */
29 int vwin_min_playouts
;
30 /* First Play Urgency - if set to less than infinity (the MoGo paper
31 * above reports 1.0 as the best), new branches are explored only
32 * if none of the existing ones has higher urgency than fpu. */
34 unsigned int equiv_rave
;
36 /* Give more weight to moves played earlier. */
38 /* Give 0 or negative rave bonus to ko threats before taking the ko.
39 1=normal bonus, 0=no bonus, -1=invert rave bonus, -2=double penalty... */
41 /* Coefficient of local tree values embedded in RAVE. */
42 floating_t ltree_rave
;
43 /* Coefficient of criticality embedded in RAVE. */
45 int crit_min_playouts
;
46 floating_t crit_plthres_coef
;
54 static inline floating_t
fast_sqrt(unsigned int x
)
56 static const floating_t table
[] = {
57 0, 1, 1.41421356237309504880, 1.73205080756887729352,
58 2.00000000000000000000, 2.23606797749978969640,
59 2.44948974278317809819, 2.64575131106459059050,
60 2.82842712474619009760, 3.00000000000000000000,
61 3.16227766016837933199, 3.31662479035539984911,
62 3.46410161513775458705, 3.60555127546398929311,
63 3.74165738677394138558, 3.87298334620741688517,
64 4.00000000000000000000, 4.12310562561766054982,
65 4.24264068711928514640, 4.35889894354067355223,
66 4.47213595499957939281, 4.58257569495584000658,
67 4.69041575982342955456, 4.79583152331271954159,
68 4.89897948556635619639, 5.00000000000000000000,
69 5.09901951359278483002, 5.19615242270663188058,
70 5.29150262212918118100, 5.38516480713450403125,
71 5.47722557505166113456, 5.56776436283002192211,
72 5.65685424949238019520, 5.74456264653802865985,
73 5.83095189484530047087, 5.91607978309961604256,
74 6.00000000000000000000, 6.08276253029821968899,
75 6.16441400296897645025, 6.24499799839839820584,
76 6.32455532033675866399, 6.40312423743284868648,
77 6.48074069840786023096, 6.55743852430200065234,
78 6.63324958071079969822, 6.70820393249936908922,
79 6.78232998312526813906, 6.85565460040104412493,
80 6.92820323027550917410, 7.00000000000000000000,
81 7.07106781186547524400, 7.14142842854284999799,
82 7.21110255092797858623, 7.28010988928051827109,
83 7.34846922834953429459, 7.41619848709566294871,
84 7.48331477354788277116, 7.54983443527074969723,
85 7.61577310586390828566, 7.68114574786860817576,
86 7.74596669241483377035, 7.81024967590665439412,
87 7.87400787401181101968, 7.93725393319377177150,
89 if (x
< sizeof(table
) / sizeof(*table
)) {
96 #define URAVE_DEBUG if (0)
97 static inline floating_t
98 ucb1rave_evaluate(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
)
100 struct ucb1_policy_amaf
*b
= p
->data
;
101 struct tree_node
*node
= descent
->node
;
102 struct tree_node
*lnode
= descent
->lnode
;
104 struct move_stats n
= node
->u
, r
= node
->amaf
;
105 if (p
->uct
->amaf_prior
) {
106 stats_merge(&r
, &node
->prior
);
108 stats_merge(&n
, &node
->prior
);
111 /* Local tree heuristics. */
112 assert(!lnode
|| lnode
->parent
);
113 if (p
->uct
->local_tree
&& b
->ltree_rave
> 0 && lnode
114 && (p
->uct
->local_tree_rootchoose
|| lnode
->parent
->parent
)) {
115 struct move_stats l
= lnode
->u
;
116 l
.playouts
= ((floating_t
) l
.playouts
) * b
->ltree_rave
/ LTREE_PLAYOUTS_MULTIPLIER
;
117 URAVE_DEBUG
fprintf(stderr
, "[ltree] adding [%s] %f%%%d to [%s] RAVE %f%%%d\n",
118 coord2sstr(node_coord(lnode
), tree
->board
), l
.value
, l
.playouts
,
119 coord2sstr(node_coord(node
), tree
->board
), r
.value
, r
.playouts
);
123 /* Criticality heuristics. */
124 if (b
->crit_rave
> 0 && (b
->crit_plthres_coef
> 0
125 ? node
->u
.playouts
> tree
->root
->u
.playouts
* b
->crit_plthres_coef
126 : node
->u
.playouts
> b
->crit_min_playouts
)) {
127 floating_t crit
= tree_node_criticality(tree
, node
);
128 if (b
->crit_negative
|| crit
> 0) {
129 floating_t val
= 1.0f
;
130 if (b
->crit_negflip
&& crit
< 0) {
134 struct move_stats c
= {
135 .value
= tree_node_get_value(tree
, parity
, val
),
136 .playouts
= crit
* r
.playouts
* b
->crit_rave
138 URAVE_DEBUG
fprintf(stderr
, "[crit] adding %f%%%d to [%s] RAVE %f%%%d\n",
140 coord2sstr(node_coord(node
), tree
->board
), r
.value
, r
.playouts
);
146 floating_t value
= 0;
149 /* At the beginning, beta is at 1 and RAVE is used.
150 * At b->equiv_rate, beta is at 1/3 and gets steeper on. */
152 if (b
->sylvain_rave
) {
153 beta
= (floating_t
) r
.playouts
/ (r
.playouts
+ n
.playouts
154 + (floating_t
) n
.playouts
* r
.playouts
/ b
->equiv_rave
);
156 /* XXX: This can be cached in descend; but we don't use this by default. */
157 beta
= sqrt(b
->equiv_rave
/ (3 * node
->parent
->u
.playouts
+ b
->equiv_rave
));
160 value
= beta
* r
.value
+ (1.f
- beta
) * n
.value
;
161 URAVE_DEBUG
fprintf(stderr
, "\t%s value = %f * %f + (1 - %f) * %f (prior %f)\n",
162 coord2sstr(node_coord(node
), tree
->board
), beta
, r
.value
, beta
, n
.value
, node
->prior
.value
);
165 URAVE_DEBUG
fprintf(stderr
, "\t%s value = %f (prior %f)\n",
166 coord2sstr(node_coord(node
), tree
->board
), n
.value
, node
->prior
.value
);
168 } else if (r
.playouts
) {
170 URAVE_DEBUG
fprintf(stderr
, "\t%s value = rave %f (prior %f)\n",
171 coord2sstr(node_coord(node
), tree
->board
), r
.value
, node
->prior
.value
);
173 descent
->value
.playouts
= r
.playouts
+ n
.playouts
;
174 descent
->value
.value
= value
;
175 return tree_node_get_value(tree
, parity
, value
);
179 ucb1rave_descend(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
, bool allow_pass
)
181 struct ucb1_policy_amaf
*b
= p
->data
;
182 floating_t nconf
= 1.f
;
183 if (b
->explore_p
> 0)
184 nconf
= sqrt(log(descent
->node
->u
.playouts
+ descent
->node
->prior
.playouts
));
185 struct uct
*u
= p
->uct
;
187 if (u
->max_slaves
> 0 && u
->slave_index
>= 0)
188 vwin
= descent
->node
== tree
->root
? b
->root_virtual_win
: b
->virtual_win
;
191 uctd_try_node_children(tree
, descent
, allow_pass
, parity
, u
->tenuki_d
, di
, urgency
) {
192 struct tree_node
*ni
= di
.node
;
193 urgency
= ucb1rave_evaluate(p
, tree
, &di
, parity
);
195 /* In distributed mode, encourage different slaves to work on different
196 * parts of the tree. We rely on the fact that children (if they exist)
197 * are the same and in the same order in all slaves. */
198 if (vwin
> 0 && ni
->u
.playouts
> b
->vwin_min_playouts
&& (child
- u
->slave_index
) % u
->max_slaves
== 0)
199 urgency
+= vwin
/ (ni
->u
.playouts
+ vwin
);
201 if (ni
->u
.playouts
> 0 && b
->explore_p
> 0) {
202 urgency
+= b
->explore_p
* nconf
/ fast_sqrt(ni
->u
.playouts
);
204 } else if (ni
->u
.playouts
+ ni
->amaf
.playouts
+ ni
->prior
.playouts
== 0) {
205 /* assert(!u->even_eqex); */
208 } uctd_set_best_child(di
, urgency
);
210 uctd_get_best_child(descent
);
214 /* Return the length of the current ko (number of moves up to to the last ko capture),
215 * 0 if the sequence is empty or doesn't start with a ko capture.
217 * W plays a ko threat
218 * B answers ko threat
219 * W re-captures the ko <- return 4
220 * B plays a ko threat
221 * W connects the ko */
222 static inline int ko_length(bool *ko_capture_map
, int map_length
)
224 if (map_length
<= 0 || !ko_capture_map
[0]) return 0;
226 while (length
+ 2 < map_length
&& ko_capture_map
[length
+ 2]) length
+= 3;
231 ucb1amaf_update(struct uct_policy
*p
, struct tree
*tree
, struct tree_node
*node
,
232 enum stone node_color
, enum stone player_color
,
233 struct playout_amafmap
*map
, struct board
*final_board
,
236 struct ucb1_policy_amaf
*b
= p
->data
;
237 enum stone winner_color
= result
> 0.5 ? S_BLACK
: S_WHITE
;
239 /* Record of the random playout - for each intersection coord,
240 * first_move[coord] is the index map->game of the first move
241 * at this coordinate, or INT_MAX if the move was not played.
242 * The parity gives the color of this move.
244 int first_map
[board_size2(final_board
)+1];
245 int *first_move
= &first_map
[1]; // +1 for pass
248 struct board bb
; bb
.size
= 9+2;
249 for (struct tree_node
*ni
= node
; ni
; ni
= ni
->parent
)
250 fprintf(stderr
, "%s ", coord2sstr(node_coord(ni
), &bb
));
251 fprintf(stderr
, "[color %d] update result %d (color %d)\n",
252 node_color
, result
, player_color
);
255 /* Initialize first_move */
256 for (int i
= pass
; i
< board_size2(final_board
); i
++) first_move
[i
] = INT_MAX
;
258 assert(map
->gamelen
> 0);
259 for (move
= map
->gamelen
- 1; move
>= map
->game_baselen
; move
--)
260 first_move
[map
->game
[move
]] = move
;
263 if (!b
->crit_amaf
&& !is_pass(node_coord(node
))) {
264 stats_add_result(&node
->winner_owner
, board_local_value(b
->crit_lvalue
, final_board
, node_coord(node
), winner_color
), 1);
265 stats_add_result(&node
->black_owner
, board_local_value(b
->crit_lvalue
, final_board
, node_coord(node
), S_BLACK
), 1);
267 stats_add_result(&node
->u
, result
, 1);
269 bool *ko_capture_map
= &map
->is_ko_capture
[move
+1];
270 int max_threat_dist
= b
->threat_rave
<= 0 ? ko_length(ko_capture_map
, map
->gamelen
- (move
+1)) : -1;
272 /* This loop ignores symmetry considerations, but they should
273 * matter only at a point when AMAF doesn't help much. */
274 assert(map
->game_baselen
>= 0);
275 for (struct tree_node
*ni
= node
->children
; ni
; ni
= ni
->sibling
) {
276 if (is_pass(node_coord(ni
))) continue;
278 /* Use the child move only if it was first played by the same color. */
279 int first
= first_move
[node_coord(ni
)];
280 if (first
== INT_MAX
) continue;
281 assert(first
> move
&& first
< map
->gamelen
);
282 int distance
= first
- (move
+ 1);
283 if (distance
& 1) continue;
286 floating_t res
= result
;
288 /* Don't give amaf bonus to a ko threat before taking the ko.
289 * http://www.grappa.univ-lille3.fr/~coulom/Aja_PhD_Thesis.pdf
291 if (distance
<= max_threat_dist
&& distance
% 6 == 4) {
292 weight
= - b
->threat_rave
;
294 } else if (b
->distance_rave
!= 0) {
295 /* Give more weight to moves played earlier */
296 weight
+= b
->distance_rave
* (map
->gamelen
- first
) / (map
->gamelen
- move
);
298 stats_add_result(&ni
->amaf
, res
, weight
);
301 stats_add_result(&ni
->winner_owner
, board_local_value(b
->crit_lvalue
, final_board
, node_coord(ni
), winner_color
), 1);
302 stats_add_result(&ni
->black_owner
, board_local_value(b
->crit_lvalue
, final_board
, node_coord(ni
), S_BLACK
), 1);
305 struct board bb
; bb
.size
= 9+2;
306 fprintf(stderr
, "* %s<%"PRIhash
"> -> %s<%"PRIhash
"> [%d/%f => %d/%f]\n",
307 coord2sstr(node_coord(node
), &bb
), node
->hash
,
308 coord2sstr(node_coord(ni
), &bb
), ni
->hash
,
309 player_color
, result
, move
, res
);
313 assert(move
>= 0 && map
->game
[move
] == node_coord(node
) && first_move
[node_coord(node
)] > move
);
314 first_move
[node_coord(node
)] = move
;
323 policy_ucb1amaf_init(struct uct
*u
, char *arg
, struct board
*board
)
325 struct uct_policy
*p
= calloc2(1, sizeof(*p
));
326 struct ucb1_policy_amaf
*b
= calloc2(1, sizeof(*b
));
329 p
->choose
= uctp_generic_choose
;
330 p
->winner
= uctp_generic_winner
;
331 p
->evaluate
= ucb1rave_evaluate
;
332 p
->descend
= ucb1rave_descend
;
333 p
->update
= ucb1amaf_update
;
334 p
->wants_amaf
= true;
337 b
->equiv_rave
= board_large(board
) ? 4000 : 3000;
339 b
->sylvain_rave
= true;
340 b
->distance_rave
= 3;
342 b
->ltree_rave
= 0.75f
;
345 b
->crit_min_playouts
= 2000;
346 b
->crit_negative
= 1;
349 b
->root_virtual_win
= -1;
350 b
->vwin_min_playouts
= 1000;
353 char *optspec
, *next
= arg
;
356 next
+= strcspn(next
, ":");
357 if (*next
) { *next
++ = 0; } else { *next
= 0; }
359 char *optname
= optspec
;
360 char *optval
= strchr(optspec
, '=');
361 if (optval
) *optval
++ = 0;
363 if (!strcasecmp(optname
, "explore_p")) {
364 b
->explore_p
= atof(optval
);
365 } else if (!strcasecmp(optname
, "fpu") && optval
) {
366 b
->fpu
= atof(optval
);
367 } else if (!strcasecmp(optname
, "equiv_rave") && optval
) {
368 b
->equiv_rave
= atof(optval
);
369 } else if (!strcasecmp(optname
, "sylvain_rave")) {
370 b
->sylvain_rave
= !optval
|| *optval
== '1';
371 } else if (!strcasecmp(optname
, "distance_rave") && optval
) {
372 b
->distance_rave
= atoi(optval
);
373 } else if (!strcasecmp(optname
, "threat_rave") && optval
) {
374 b
->threat_rave
= atoi(optval
);
375 } else if (!strcasecmp(optname
, "ltree_rave") && optval
) {
376 b
->ltree_rave
= atof(optval
);
377 } else if (!strcasecmp(optname
, "crit_rave") && optval
) {
378 b
->crit_rave
= atof(optval
);
379 } else if (!strcasecmp(optname
, "crit_min_playouts") && optval
) {
380 b
->crit_min_playouts
= atoi(optval
);
381 } else if (!strcasecmp(optname
, "crit_plthres_coef") && optval
) {
382 b
->crit_plthres_coef
= atof(optval
);
383 } else if (!strcasecmp(optname
, "crit_negative")) {
384 b
->crit_negative
= !optval
|| *optval
== '1';
385 } else if (!strcasecmp(optname
, "crit_negflip")) {
386 b
->crit_negflip
= !optval
|| *optval
== '1';
387 } else if (!strcasecmp(optname
, "crit_amaf")) {
388 b
->crit_amaf
= !optval
|| *optval
== '1';
389 } else if (!strcasecmp(optname
, "crit_lvalue")) {
390 b
->crit_lvalue
= !optval
|| *optval
== '1';
391 } else if (!strcasecmp(optname
, "virtual_win") && optval
) {
392 b
->virtual_win
= atoi(optval
);
393 } else if (!strcasecmp(optname
, "root_virtual_win") && optval
) {
394 b
->root_virtual_win
= atoi(optval
);
395 } else if (!strcasecmp(optname
, "vwin_min_playouts") && optval
) {
396 b
->vwin_min_playouts
= atoi(optval
);
398 fprintf(stderr
, "ucb1amaf: Invalid policy argument %s or missing value\n",
404 if (b
->root_virtual_win
< 0)
405 b
->root_virtual_win
= b
->virtual_win
;