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 /* First Play Urgency - if set to less than infinity (the MoGo paper
26 * above reports 1.0 as the best), new branches are explored only
27 * if none of the existing ones has higher urgency than fpu. */
29 unsigned int equiv_rave
;
31 /* Give more weight to moves played earlier. */
33 /* Coefficient of local tree values embedded in RAVE. */
34 floating_t ltree_rave
;
35 /* Coefficient of criticality embedded in RAVE. */
37 int crit_min_playouts
;
38 floating_t crit_plthres_coef
;
46 static inline floating_t
fast_sqrt(unsigned int x
)
48 static const floating_t table
[] = {
49 0, 1, 1.41421356237309504880, 1.73205080756887729352,
50 2.00000000000000000000, 2.23606797749978969640,
51 2.44948974278317809819, 2.64575131106459059050,
52 2.82842712474619009760, 3.00000000000000000000,
53 3.16227766016837933199, 3.31662479035539984911,
54 3.46410161513775458705, 3.60555127546398929311,
55 3.74165738677394138558, 3.87298334620741688517,
56 4.00000000000000000000, 4.12310562561766054982,
57 4.24264068711928514640, 4.35889894354067355223,
58 4.47213595499957939281, 4.58257569495584000658,
59 4.69041575982342955456, 4.79583152331271954159,
60 4.89897948556635619639, 5.00000000000000000000,
61 5.09901951359278483002, 5.19615242270663188058,
62 5.29150262212918118100, 5.38516480713450403125,
63 5.47722557505166113456, 5.56776436283002192211,
64 5.65685424949238019520, 5.74456264653802865985,
65 5.83095189484530047087, 5.91607978309961604256,
66 6.00000000000000000000, 6.08276253029821968899,
67 6.16441400296897645025, 6.24499799839839820584,
68 6.32455532033675866399, 6.40312423743284868648,
69 6.48074069840786023096, 6.55743852430200065234,
70 6.63324958071079969822, 6.70820393249936908922,
71 6.78232998312526813906, 6.85565460040104412493,
72 6.92820323027550917410, 7.00000000000000000000,
73 7.07106781186547524400, 7.14142842854284999799,
74 7.21110255092797858623, 7.28010988928051827109,
75 7.34846922834953429459, 7.41619848709566294871,
76 7.48331477354788277116, 7.54983443527074969723,
77 7.61577310586390828566, 7.68114574786860817576,
78 7.74596669241483377035, 7.81024967590665439412,
79 7.87400787401181101968, 7.93725393319377177150,
81 if (x
< sizeof(table
) / sizeof(*table
)) {
88 #define URAVE_DEBUG if (0)
89 static floating_t
inline
90 ucb1rave_evaluate(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
)
92 struct ucb1_policy_amaf
*b
= p
->data
;
93 struct tree_node
*node
= descent
->node
;
94 struct tree_node
*lnode
= descent
->lnode
;
96 struct move_stats n
= node
->u
, r
= node
->amaf
;
97 if (p
->uct
->amaf_prior
) {
98 stats_merge(&r
, &node
->prior
);
100 stats_merge(&n
, &node
->prior
);
103 /* Local tree heuristics. */
104 assert(!lnode
|| lnode
->parent
);
105 if (p
->uct
->local_tree
&& b
->ltree_rave
> 0 && lnode
106 && (p
->uct
->local_tree_rootchoose
|| lnode
->parent
->parent
)) {
107 struct move_stats l
= lnode
->u
;
108 l
.playouts
= ((floating_t
) l
.playouts
) * b
->ltree_rave
/ LTREE_PLAYOUTS_MULTIPLIER
;
109 URAVE_DEBUG
fprintf(stderr
, "[ltree] adding [%s] %f%%%d to [%s] RAVE %f%%%d\n",
110 coord2sstr(node_coord(lnode
), tree
->board
), l
.value
, l
.playouts
,
111 coord2sstr(node_coord(node
), tree
->board
), r
.value
, r
.playouts
);
115 /* Criticality heuristics. */
116 if (b
->crit_rave
> 0 && (b
->crit_plthres_coef
> 0
117 ? node
->u
.playouts
> tree
->root
->u
.playouts
* b
->crit_plthres_coef
118 : node
->u
.playouts
> b
->crit_min_playouts
)) {
119 floating_t crit
= tree_node_criticality(tree
, node
);
120 if (b
->crit_negative
|| crit
> 0) {
121 floating_t val
= 1.0f
;
122 if (b
->crit_negflip
&& crit
< 0) {
126 struct move_stats c
= {
127 .value
= tree_node_get_value(tree
, parity
, val
),
128 .playouts
= crit
* r
.playouts
* b
->crit_rave
130 URAVE_DEBUG
fprintf(stderr
, "[crit] adding %f%%%d to [%s] RAVE %f%%%d\n",
132 coord2sstr(node_coord(node
), tree
->board
), r
.value
, r
.playouts
);
138 floating_t value
= 0;
141 /* At the beginning, beta is at 1 and RAVE is used.
142 * At b->equiv_rate, beta is at 1/3 and gets steeper on. */
144 if (b
->sylvain_rave
) {
145 beta
= (floating_t
) r
.playouts
/ (r
.playouts
+ n
.playouts
146 + (floating_t
) n
.playouts
* r
.playouts
/ b
->equiv_rave
);
148 /* XXX: This can be cached in descend; but we don't use this by default. */
149 beta
= sqrt(b
->equiv_rave
/ (3 * node
->parent
->u
.playouts
+ b
->equiv_rave
));
152 value
= beta
* r
.value
+ (1.f
- beta
) * n
.value
;
153 URAVE_DEBUG
fprintf(stderr
, "\t%s value = %f * %f + (1 - %f) * %f (prior %f)\n",
154 coord2sstr(node_coord(node
), tree
->board
), beta
, r
.value
, beta
, n
.value
, node
->prior
.value
);
157 URAVE_DEBUG
fprintf(stderr
, "\t%s value = %f (prior %f)\n",
158 coord2sstr(node_coord(node
), tree
->board
), n
.value
, node
->prior
.value
);
160 } else if (r
.playouts
) {
162 URAVE_DEBUG
fprintf(stderr
, "\t%s value = rave %f (prior %f)\n",
163 coord2sstr(node_coord(node
), tree
->board
), r
.value
, node
->prior
.value
);
165 descent
->value
.playouts
= r
.playouts
+ n
.playouts
;
166 descent
->value
.value
= value
;
167 return tree_node_get_value(tree
, parity
, value
);
171 ucb1rave_descend(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
, bool allow_pass
)
173 struct ucb1_policy_amaf
*b
= p
->data
;
174 floating_t nconf
= 1.f
;
175 if (b
->explore_p
> 0)
176 nconf
= sqrt(log(descent
->node
->u
.playouts
+ descent
->node
->prior
.playouts
));
178 uctd_try_node_children(tree
, descent
, allow_pass
, parity
, p
->uct
->tenuki_d
, di
, urgency
) {
179 struct tree_node
*ni
= di
.node
;
180 urgency
= ucb1rave_evaluate(p
, tree
, &di
, parity
);
182 if (ni
->u
.playouts
> 0 && b
->explore_p
> 0) {
183 urgency
+= b
->explore_p
* nconf
/ fast_sqrt(ni
->u
.playouts
);
185 } else if (ni
->u
.playouts
+ ni
->amaf
.playouts
+ ni
->prior
.playouts
== 0) {
186 /* assert(!u->even_eqex); */
189 } uctd_set_best_child(di
, urgency
);
191 uctd_get_best_child(descent
);
196 ucb1amaf_update(struct uct_policy
*p
, struct tree
*tree
, struct tree_node
*node
,
197 enum stone node_color
, enum stone player_color
,
198 struct playout_amafmap
*map
, struct board
*final_board
,
201 struct ucb1_policy_amaf
*b
= p
->data
;
202 enum stone winner_color
= result
> 0.5 ? S_BLACK
: S_WHITE
;
204 /* Record of the random playout - for each intersection coord,
205 * first_move[coord] is the index map->game of the first move
206 * at this coordinate, or INT_MAX if the move was not played.
207 * The parity gives the color of this move.
209 int first_map
[board_size2(final_board
)+1];
210 int *first_move
= &first_map
[1]; // +1 for pass
213 struct board bb
; bb
.size
= 9+2;
214 for (struct tree_node
*ni
= node
; ni
; ni
= ni
->parent
)
215 fprintf(stderr
, "%s ", coord2sstr(node_coord(ni
), &bb
));
216 fprintf(stderr
, "[color %d] update result %d (color %d)\n",
217 node_color
, result
, player_color
);
220 /* Initialize first_move */
221 for (int i
= pass
; i
< board_size2(final_board
); i
++) first_move
[i
] = INT_MAX
;
223 assert(map
->gamelen
> 0);
224 for (move
= map
->gamelen
- 1; move
>= map
->game_baselen
; move
--)
225 first_move
[map
->game
[move
]] = move
;
228 if (!b
->crit_amaf
&& !is_pass(node_coord(node
))) {
229 stats_add_result(&node
->winner_owner
, board_local_value(b
->crit_lvalue
, final_board
, node_coord(node
), winner_color
), 1);
230 stats_add_result(&node
->black_owner
, board_local_value(b
->crit_lvalue
, final_board
, node_coord(node
), S_BLACK
), 1);
232 stats_add_result(&node
->u
, result
, 1);
234 /* This loop ignores symmetry considerations, but they should
235 * matter only at a point when AMAF doesn't help much. */
236 assert(map
->game_baselen
>= 0);
237 for (struct tree_node
*ni
= node
->children
; ni
; ni
= ni
->sibling
) {
238 if (is_pass(node_coord(ni
))) continue;
240 /* Use the child move only if it was first played by the same color. */
241 int first
= first_move
[node_coord(ni
)];
242 if (first
== INT_MAX
) continue;
243 assert(first
> move
&& first
< map
->gamelen
);
244 int distance
= first
- (move
+ 1);
245 if (distance
& 1) continue;
247 /* Give more weight to moves played earlier */
249 if (b
->distance_rave
!= 0) {
250 weight
+= b
->distance_rave
* (map
->gamelen
- first
) / (map
->gamelen
- move
);
252 stats_add_result(&ni
->amaf
, result
, weight
);
255 stats_add_result(&ni
->winner_owner
, board_local_value(b
->crit_lvalue
, final_board
, node_coord(ni
), winner_color
), 1);
256 stats_add_result(&ni
->black_owner
, board_local_value(b
->crit_lvalue
, final_board
, node_coord(ni
), S_BLACK
), 1);
259 struct board bb
; bb
.size
= 9+2;
260 fprintf(stderr
, "* %s<%"PRIhash
"> -> %s<%"PRIhash
"> [%d/%f => %d/%f]\n",
261 coord2sstr(node_coord(node
), &bb
), node
->hash
,
262 coord2sstr(node_coord(ni
), &bb
), ni
->hash
,
263 player_color
, result
, move
, res
);
267 assert(move
>= 0 && map
->game
[move
] == node_coord(node
) && first_move
[node_coord(node
)] > move
);
268 first_move
[node_coord(node
)] = move
;
277 policy_ucb1amaf_init(struct uct
*u
, char *arg
)
279 struct uct_policy
*p
= calloc2(1, sizeof(*p
));
280 struct ucb1_policy_amaf
*b
= calloc2(1, sizeof(*b
));
283 p
->choose
= uctp_generic_choose
;
284 p
->winner
= uctp_generic_winner
;
285 p
->evaluate
= ucb1rave_evaluate
;
286 p
->descend
= ucb1rave_descend
;
287 p
->update
= ucb1amaf_update
;
288 p
->wants_amaf
= true;
290 b
->explore_p
= 0; // 0.02 can be also good on 19x19 with prior=eqex=40
291 b
->equiv_rave
= 3000;
293 b
->sylvain_rave
= true;
294 b
->ltree_rave
= 0.75f
;
297 b
->crit_min_playouts
= 2000;
298 b
->crit_negative
= 1;
302 char *optspec
, *next
= arg
;
305 next
+= strcspn(next
, ":");
306 if (*next
) { *next
++ = 0; } else { *next
= 0; }
308 char *optname
= optspec
;
309 char *optval
= strchr(optspec
, '=');
310 if (optval
) *optval
++ = 0;
312 if (!strcasecmp(optname
, "explore_p")) {
313 b
->explore_p
= atof(optval
);
314 } else if (!strcasecmp(optname
, "fpu") && optval
) {
315 b
->fpu
= atof(optval
);
316 } else if (!strcasecmp(optname
, "equiv_rave") && optval
) {
317 b
->equiv_rave
= atof(optval
);
318 } else if (!strcasecmp(optname
, "sylvain_rave")) {
319 b
->sylvain_rave
= !optval
|| *optval
== '1';
320 } else if (!strcasecmp(optname
, "distance_rave") && optval
) {
321 b
->distance_rave
= atoi(optval
);
322 } else if (!strcasecmp(optname
, "ltree_rave") && optval
) {
323 b
->ltree_rave
= atof(optval
);
324 } else if (!strcasecmp(optname
, "crit_rave") && optval
) {
325 b
->crit_rave
= atof(optval
);
326 } else if (!strcasecmp(optname
, "crit_min_playouts") && optval
) {
327 b
->crit_min_playouts
= atoi(optval
);
328 } else if (!strcasecmp(optname
, "crit_plthres_coef") && optval
) {
329 b
->crit_plthres_coef
= atof(optval
);
330 } else if (!strcasecmp(optname
, "crit_negative")) {
331 b
->crit_negative
= !optval
|| *optval
== '1';
332 } else if (!strcasecmp(optname
, "crit_negflip")) {
333 b
->crit_negflip
= !optval
|| *optval
== '1';
334 } else if (!strcasecmp(optname
, "crit_amaf")) {
335 b
->crit_amaf
= !optval
|| *optval
== '1';
336 } else if (!strcasecmp(optname
, "crit_lvalue")) {
337 b
->crit_lvalue
= !optval
|| *optval
== '1';
339 fprintf(stderr
, "ucb1amaf: Invalid policy argument %s or missing value\n",