11 #include "uct/internal.h"
13 #include "uct/policy/generic.h"
15 /* This implements the UCB1 policy with an extra AMAF heuristics. */
17 struct ucb1_policy_amaf
{
18 /* This is what the Modification of UCT with Patterns in Monte Carlo Go
19 * paper calls 'p'. Original UCB has this on 2, but this seems to
20 * produce way too wide searches; reduce this to get deeper and
21 * narrower readouts - try 0.2. */
23 /* First Play Urgency - if set to less than infinity (the MoGo paper
24 * above reports 1.0 as the best), new branches are explored only
25 * if none of the existing ones has higher urgency than fpu. */
27 unsigned int equiv_rave
;
30 /* Coefficient of local tree values embedded in RAVE. */
31 floating_t ltree_rave
;
32 /* Coefficient of criticality embedded in RAVE. */
34 int crit_min_playouts
;
40 static inline floating_t
fast_sqrt(unsigned int x
)
42 static const floating_t table
[] = {
43 0, 1, 1.41421356237309504880, 1.73205080756887729352,
44 2.00000000000000000000, 2.23606797749978969640,
45 2.44948974278317809819, 2.64575131106459059050,
46 2.82842712474619009760, 3.00000000000000000000,
47 3.16227766016837933199, 3.31662479035539984911,
48 3.46410161513775458705, 3.60555127546398929311,
49 3.74165738677394138558, 3.87298334620741688517,
50 4.00000000000000000000, 4.12310562561766054982,
51 4.24264068711928514640, 4.35889894354067355223,
52 4.47213595499957939281, 4.58257569495584000658,
53 4.69041575982342955456, 4.79583152331271954159,
54 4.89897948556635619639, 5.00000000000000000000,
55 5.09901951359278483002, 5.19615242270663188058,
56 5.29150262212918118100, 5.38516480713450403125,
57 5.47722557505166113456, 5.56776436283002192211,
58 5.65685424949238019520, 5.74456264653802865985,
59 5.83095189484530047087, 5.91607978309961604256,
60 6.00000000000000000000, 6.08276253029821968899,
61 6.16441400296897645025, 6.24499799839839820584,
62 6.32455532033675866399, 6.40312423743284868648,
63 6.48074069840786023096, 6.55743852430200065234,
64 6.63324958071079969822, 6.70820393249936908922,
65 6.78232998312526813906, 6.85565460040104412493,
66 6.92820323027550917410, 7.00000000000000000000,
67 7.07106781186547524400, 7.14142842854284999799,
68 7.21110255092797858623, 7.28010988928051827109,
69 7.34846922834953429459, 7.41619848709566294871,
70 7.48331477354788277116, 7.54983443527074969723,
71 7.61577310586390828566, 7.68114574786860817576,
72 7.74596669241483377035, 7.81024967590665439412,
73 7.87400787401181101968, 7.93725393319377177150,
75 if (x
< sizeof(table
) / sizeof(*table
)) {
82 #define URAVE_DEBUG if (0)
83 static floating_t
inline
84 ucb1rave_evaluate(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
)
86 struct ucb1_policy_amaf
*b
= p
->data
;
87 struct tree_node
*node
= descent
->node
;
88 struct tree_node
*lnode
= descent
->lnode
;
90 struct move_stats n
= node
->u
, r
= node
->amaf
;
91 if (p
->uct
->amaf_prior
) {
92 stats_merge(&r
, &node
->prior
);
94 stats_merge(&n
, &node
->prior
);
97 /* Local tree heuristics. */
98 assert(!lnode
|| lnode
->parent
);
99 if (p
->uct
->local_tree
&& b
->ltree_rave
> 0 && lnode
100 && (p
->uct
->local_tree_rootchoose
|| lnode
->parent
->parent
)) {
101 struct move_stats l
= lnode
->u
;
102 l
.playouts
= ((floating_t
) l
.playouts
) * b
->ltree_rave
/ LTREE_PLAYOUTS_MULTIPLIER
;
103 URAVE_DEBUG
fprintf(stderr
, "[ltree] adding [%s] %f%%%d to [%s] RAVE %f%%%d\n",
104 coord2sstr(node_coord(lnode
), tree
->board
), l
.value
, l
.playouts
,
105 coord2sstr(node_coord(node
), tree
->board
), r
.value
, r
.playouts
);
109 /* Criticality heuristics. */
110 if (b
->crit_rave
> 0 && node
->u
.playouts
> b
->crit_min_playouts
) {
111 floating_t crit
= tree_node_criticality(tree
, node
);
112 if (b
->crit_negative
|| crit
> 0) {
113 struct move_stats c
= {
114 .value
= tree_node_get_value(tree
, parity
, 1.0f
),
115 .playouts
= crit
* r
.playouts
* b
->crit_rave
117 URAVE_DEBUG
fprintf(stderr
, "[crit] adding %f%%%d to [%s] RAVE %f%%%d\n",
119 coord2sstr(node_coord(node
), tree
->board
), r
.value
, r
.playouts
);
125 floating_t value
= 0;
128 /* At the beginning, beta is at 1 and RAVE is used.
129 * At b->equiv_rate, beta is at 1/3 and gets steeper on. */
131 if (b
->sylvain_rave
) {
132 beta
= (floating_t
) r
.playouts
/ (r
.playouts
+ n
.playouts
133 + (floating_t
) n
.playouts
* r
.playouts
/ b
->equiv_rave
);
135 /* XXX: This can be cached in descend; but we don't use this by default. */
136 beta
= sqrt(b
->equiv_rave
/ (3 * node
->parent
->u
.playouts
+ b
->equiv_rave
));
139 value
= beta
* r
.value
+ (1.f
- beta
) * n
.value
;
140 URAVE_DEBUG
fprintf(stderr
, "\t%s value = %f * %f + (1 - %f) * %f (prior %f)\n",
141 coord2sstr(node_coord(node
), tree
->board
), beta
, r
.value
, beta
, n
.value
, node
->prior
.value
);
144 URAVE_DEBUG
fprintf(stderr
, "\t%s value = %f (prior %f)\n",
145 coord2sstr(node_coord(node
), tree
->board
), n
.value
, node
->prior
.value
);
147 } else if (r
.playouts
) {
149 URAVE_DEBUG
fprintf(stderr
, "\t%s value = rave %f (prior %f)\n",
150 coord2sstr(node_coord(node
), tree
->board
), r
.value
, node
->prior
.value
);
152 descent
->value
.playouts
= r
.playouts
+ n
.playouts
;
153 descent
->value
.value
= value
;
154 return tree_node_get_value(tree
, parity
, value
);
158 ucb1rave_descend(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
, bool allow_pass
)
160 struct ucb1_policy_amaf
*b
= p
->data
;
161 floating_t nconf
= 1.f
;
162 if (b
->explore_p
> 0)
163 nconf
= sqrt(log(descent
->node
->u
.playouts
+ descent
->node
->prior
.playouts
));
165 uctd_try_node_children(tree
, descent
, allow_pass
, parity
, p
->uct
->tenuki_d
, di
, urgency
) {
166 struct tree_node
*ni
= di
.node
;
167 urgency
= ucb1rave_evaluate(p
, tree
, &di
, parity
);
169 if (ni
->u
.playouts
> 0 && b
->explore_p
> 0) {
170 urgency
+= b
->explore_p
* nconf
/ fast_sqrt(ni
->u
.playouts
);
172 } else if (ni
->u
.playouts
+ ni
->amaf
.playouts
+ ni
->prior
.playouts
== 0) {
173 /* assert(!u->even_eqex); */
176 } uctd_set_best_child(di
, urgency
);
178 uctd_get_best_child(descent
);
183 ucb1amaf_update(struct uct_policy
*p
, struct tree
*tree
, struct tree_node
*node
,
184 enum stone node_color
, enum stone player_color
,
185 struct playout_amafmap
*map
, struct board
*final_board
,
188 struct ucb1_policy_amaf
*b
= p
->data
;
189 enum stone winner_color
= result
> 0.5 ? S_BLACK
: S_WHITE
;
190 enum stone child_color
= stone_other(node_color
);
193 struct board bb
; bb
.size
= 9+2;
194 for (struct tree_node
*ni
= node
; ni
; ni
= ni
->parent
)
195 fprintf(stderr
, "%s ", coord2sstr(node_coord(ni
), &bb
));
196 fprintf(stderr
, "[color %d] update result %d (color %d)\n",
197 node_color
, result
, player_color
);
201 if (node
->parent
== NULL
)
202 assert(tree
->root_color
== stone_other(child_color
));
204 if (!b
->crit_amaf
&& !is_pass(node_coord(node
))) {
205 stats_add_result(&node
->winner_owner
, board_at(final_board
, node_coord(node
)) == winner_color
? 1.0 : 0.0, 1);
206 stats_add_result(&node
->black_owner
, board_at(final_board
, node_coord(node
)) == S_BLACK
? 1.0 : 0.0, 1);
208 stats_add_result(&node
->u
, result
, 1);
209 if (!is_pass(node_coord(node
)) && amaf_nakade(map
->map
[node_coord(node
)]))
210 amaf_op(map
->map
[node_coord(node
)], -);
212 /* This loop ignores symmetry considerations, but they should
213 * matter only at a point when AMAF doesn't help much. */
214 assert(map
->game_baselen
>= 0);
215 for (struct tree_node
*ni
= node
->children
; ni
; ni
= ni
->sibling
) {
216 enum stone amaf_color
= map
->map
[node_coord(ni
)];
217 assert(amaf_color
!= S_OFFBOARD
);
218 if (amaf_color
== S_NONE
)
220 if (amaf_nakade(map
->map
[node_coord(ni
)])) {
221 if (!b
->check_nakade
)
224 for (i
= map
->game_baselen
; i
< map
->gamelen
; i
++)
225 if (map
->game
[i
].coord
== node_coord(ni
)
226 && map
->game
[i
].color
== child_color
)
228 if (i
== map
->gamelen
)
230 amaf_color
= child_color
;
233 floating_t nres
= result
;
234 if (amaf_color
!= child_color
) {
237 /* For child_color != player_color, we still want
238 * to record the result unmodified; in that case,
239 * we will correctly negate them at the descend phase. */
241 if (b
->crit_amaf
&& !is_pass(node_coord(node
))) {
242 stats_add_result(&ni
->winner_owner
, board_at(final_board
, node_coord(ni
)) == winner_color
? 1.0 : 0.0, 1);
243 stats_add_result(&ni
->black_owner
, board_at(final_board
, node_coord(ni
)) == S_BLACK
? 1.0 : 0.0, 1);
245 stats_add_result(&ni
->amaf
, nres
, 1);
248 struct board bb
; bb
.size
= 9+2;
249 fprintf(stderr
, "* %s<%"PRIhash
"> -> %s<%"PRIhash
"> [%d/%f => %d/%f]\n",
250 coord2sstr(node_coord(node
), &bb
), node
->hash
,
251 coord2sstr(node_coord(ni
), &bb
), ni
->hash
,
252 player_color
, result
, child_color
, nres
);
256 if (!is_pass(node_coord(node
))) {
259 node
= node
->parent
; child_color
= stone_other(child_color
);
265 policy_ucb1amaf_init(struct uct
*u
, char *arg
)
267 struct uct_policy
*p
= calloc2(1, sizeof(*p
));
268 struct ucb1_policy_amaf
*b
= calloc2(1, sizeof(*b
));
271 p
->choose
= uctp_generic_choose
;
272 p
->winner
= uctp_generic_winner
;
273 p
->evaluate
= ucb1rave_evaluate
;
274 p
->descend
= ucb1rave_descend
;
275 p
->update
= ucb1amaf_update
;
276 p
->wants_amaf
= true;
278 b
->explore_p
= 0; // 0.02 can be also good on 19x19 with prior=eqex=40
279 b
->equiv_rave
= 3000;
281 b
->check_nakade
= true;
282 b
->sylvain_rave
= true;
283 b
->ltree_rave
= 0.75f
;
286 b
->crit_min_playouts
= 2000;
287 b
->crit_negative
= 1;
291 char *optspec
, *next
= arg
;
294 next
+= strcspn(next
, ":");
295 if (*next
) { *next
++ = 0; } else { *next
= 0; }
297 char *optname
= optspec
;
298 char *optval
= strchr(optspec
, '=');
299 if (optval
) *optval
++ = 0;
301 if (!strcasecmp(optname
, "explore_p")) {
302 b
->explore_p
= atof(optval
);
303 } else if (!strcasecmp(optname
, "fpu") && optval
) {
304 b
->fpu
= atof(optval
);
305 } else if (!strcasecmp(optname
, "equiv_rave") && optval
) {
306 b
->equiv_rave
= atof(optval
);
307 } else if (!strcasecmp(optname
, "sylvain_rave")) {
308 b
->sylvain_rave
= !optval
|| *optval
== '1';
309 } else if (!strcasecmp(optname
, "check_nakade")) {
310 b
->check_nakade
= !optval
|| *optval
== '1';
311 } else if (!strcasecmp(optname
, "ltree_rave") && optval
) {
312 b
->ltree_rave
= atof(optval
);
313 } else if (!strcasecmp(optname
, "crit_rave") && optval
) {
314 b
->crit_rave
= atof(optval
);
315 } else if (!strcasecmp(optname
, "crit_min_playouts") && optval
) {
316 b
->crit_min_playouts
= atoi(optval
);
317 } else if (!strcasecmp(optname
, "crit_negative")) {
318 b
->crit_negative
= !optval
|| *optval
== '1';
319 } else if (!strcasecmp(optname
, "crit_amaf")) {
320 b
->crit_amaf
= !optval
|| *optval
== '1';
322 fprintf(stderr
, "ucb1amaf: Invalid policy argument %s or missing value\n",