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 /* In distributed mode, encourage different slaves to work on different
24 * parts of the tree by adding virtual wins to different nodes. */
27 int vwin_min_playouts
;
28 /* First Play Urgency - if set to less than infinity (the MoGo paper
29 * above reports 1.0 as the best), new branches are explored only
30 * if none of the existing ones has higher urgency than fpu. */
32 unsigned int equiv_rave
;
35 /* Coefficient of local tree values embedded in RAVE. */
36 floating_t ltree_rave
;
37 /* Coefficient of criticality embedded in RAVE. */
39 int crit_min_playouts
;
45 static inline floating_t
fast_sqrt(unsigned int x
)
47 static const floating_t table
[] = {
48 0, 1, 1.41421356237309504880, 1.73205080756887729352,
49 2.00000000000000000000, 2.23606797749978969640,
50 2.44948974278317809819, 2.64575131106459059050,
51 2.82842712474619009760, 3.00000000000000000000,
52 3.16227766016837933199, 3.31662479035539984911,
53 3.46410161513775458705, 3.60555127546398929311,
54 3.74165738677394138558, 3.87298334620741688517,
55 4.00000000000000000000, 4.12310562561766054982,
56 4.24264068711928514640, 4.35889894354067355223,
57 4.47213595499957939281, 4.58257569495584000658,
58 4.69041575982342955456, 4.79583152331271954159,
59 4.89897948556635619639, 5.00000000000000000000,
60 5.09901951359278483002, 5.19615242270663188058,
61 5.29150262212918118100, 5.38516480713450403125,
62 5.47722557505166113456, 5.56776436283002192211,
63 5.65685424949238019520, 5.74456264653802865985,
64 5.83095189484530047087, 5.91607978309961604256,
65 6.00000000000000000000, 6.08276253029821968899,
66 6.16441400296897645025, 6.24499799839839820584,
67 6.32455532033675866399, 6.40312423743284868648,
68 6.48074069840786023096, 6.55743852430200065234,
69 6.63324958071079969822, 6.70820393249936908922,
70 6.78232998312526813906, 6.85565460040104412493,
71 6.92820323027550917410, 7.00000000000000000000,
72 7.07106781186547524400, 7.14142842854284999799,
73 7.21110255092797858623, 7.28010988928051827109,
74 7.34846922834953429459, 7.41619848709566294871,
75 7.48331477354788277116, 7.54983443527074969723,
76 7.61577310586390828566, 7.68114574786860817576,
77 7.74596669241483377035, 7.81024967590665439412,
78 7.87400787401181101968, 7.93725393319377177150,
80 if (x
< sizeof(table
) / sizeof(*table
)) {
87 #define URAVE_DEBUG if (0)
88 static floating_t
inline
89 ucb1rave_evaluate(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
)
91 struct ucb1_policy_amaf
*b
= p
->data
;
92 struct tree_node
*node
= descent
->node
;
93 struct tree_node
*lnode
= descent
->lnode
;
95 struct move_stats n
= node
->u
, r
= node
->amaf
;
96 if (p
->uct
->amaf_prior
) {
97 stats_merge(&r
, &node
->prior
);
99 stats_merge(&n
, &node
->prior
);
102 /* Local tree heuristics. */
103 assert(!lnode
|| lnode
->parent
);
104 if (p
->uct
->local_tree
&& b
->ltree_rave
> 0 && lnode
105 && (p
->uct
->local_tree_rootchoose
|| lnode
->parent
->parent
)) {
106 struct move_stats l
= lnode
->u
;
107 l
.playouts
= ((floating_t
) l
.playouts
) * b
->ltree_rave
/ LTREE_PLAYOUTS_MULTIPLIER
;
108 URAVE_DEBUG
fprintf(stderr
, "[ltree] adding [%s] %f%%%d to [%s] RAVE %f%%%d\n",
109 coord2sstr(lnode
->coord
, tree
->board
), l
.value
, l
.playouts
,
110 coord2sstr(node
->coord
, tree
->board
), r
.value
, r
.playouts
);
114 /* Criticality heuristics. */
115 if (b
->crit_rave
> 0 && node
->u
.playouts
> b
->crit_min_playouts
) {
116 floating_t crit
= tree_node_criticality(tree
, node
);
117 if (b
->crit_negative
|| crit
> 0) {
118 struct move_stats c
= {
119 .value
= tree_node_get_value(tree
, parity
, 1.0f
),
120 .playouts
= crit
* r
.playouts
* b
->crit_rave
122 URAVE_DEBUG
fprintf(stderr
, "[crit] adding %f%%%d to [%s] RAVE %f%%%d\n",
124 coord2sstr(node
->coord
, tree
->board
), r
.value
, r
.playouts
);
130 floating_t value
= 0;
133 /* At the beginning, beta is at 1 and RAVE is used.
134 * At b->equiv_rate, beta is at 1/3 and gets steeper on. */
136 if (b
->sylvain_rave
) {
137 beta
= (floating_t
) r
.playouts
/ (r
.playouts
+ n
.playouts
138 + (floating_t
) n
.playouts
* r
.playouts
/ b
->equiv_rave
);
140 /* XXX: This can be cached in descend; but we don't use this by default. */
141 beta
= sqrt(b
->equiv_rave
/ (3 * node
->parent
->u
.playouts
+ b
->equiv_rave
));
144 value
= beta
* r
.value
+ (1.f
- beta
) * n
.value
;
145 URAVE_DEBUG
fprintf(stderr
, "\t%s value = %f * %f + (1 - %f) * %f (prior %f)\n",
146 coord2sstr(node
->coord
, tree
->board
), beta
, r
.value
, beta
, n
.value
, node
->prior
.value
);
149 URAVE_DEBUG
fprintf(stderr
, "\t%s value = %f (prior %f)\n",
150 coord2sstr(node
->coord
, tree
->board
), n
.value
, node
->prior
.value
);
152 } else if (r
.playouts
) {
154 URAVE_DEBUG
fprintf(stderr
, "\t%s value = rave %f (prior %f)\n",
155 coord2sstr(node
->coord
, tree
->board
), r
.value
, node
->prior
.value
);
157 descent
->value
.playouts
= r
.playouts
+ n
.playouts
;
158 descent
->value
.value
= value
;
159 return tree_node_get_value(tree
, parity
, value
);
163 ucb1rave_descend(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
, bool allow_pass
)
165 struct ucb1_policy_amaf
*b
= p
->data
;
166 floating_t nconf
= 1.f
;
167 if (b
->explore_p
> 0)
168 nconf
= sqrt(log(descent
->node
->u
.playouts
+ descent
->node
->prior
.playouts
));
169 struct uct
*u
= p
->uct
;
171 if (u
->max_slaves
> 0 && u
->slave_index
>= 0)
172 vwin
= descent
->node
== tree
->root
? b
->root_virtual_win
: b
->virtual_win
;
175 uctd_try_node_children(tree
, descent
, allow_pass
, parity
, u
->tenuki_d
, di
, urgency
) {
176 struct tree_node
*ni
= di
.node
;
177 urgency
= ucb1rave_evaluate(p
, tree
, &di
, parity
);
179 /* In distributed mode, encourage different slaves to work on different
180 * parts of the tree. We rely on the fact that children (if they exist)
181 * are the same and in the same order in all slaves. */
182 if (vwin
> 0 && ni
->u
.playouts
> b
->vwin_min_playouts
&& (child
- u
->slave_index
) % u
->max_slaves
== 0)
183 urgency
+= vwin
/ (ni
->u
.playouts
+ vwin
);
185 if (ni
->u
.playouts
> 0 && b
->explore_p
> 0) {
186 urgency
+= b
->explore_p
* nconf
/ fast_sqrt(ni
->u
.playouts
);
188 } else if (ni
->u
.playouts
+ ni
->amaf
.playouts
+ ni
->prior
.playouts
== 0) {
189 /* assert(!u->even_eqex); */
192 } uctd_set_best_child(di
, urgency
);
194 uctd_get_best_child(descent
);
199 ucb1amaf_update(struct uct_policy
*p
, struct tree
*tree
, struct tree_node
*node
,
200 enum stone node_color
, enum stone player_color
,
201 struct playout_amafmap
*map
, struct board
*final_board
,
204 struct ucb1_policy_amaf
*b
= p
->data
;
205 enum stone winner_color
= result
> 0.5 ? S_BLACK
: S_WHITE
;
206 enum stone child_color
= stone_other(node_color
);
209 struct board bb
; bb
.size
= 9+2;
210 for (struct tree_node
*ni
= node
; ni
; ni
= ni
->parent
)
211 fprintf(stderr
, "%s ", coord2sstr(ni
->coord
, &bb
));
212 fprintf(stderr
, "[color %d] update result %d (color %d)\n",
213 node_color
, result
, player_color
);
217 if (node
->parent
== NULL
)
218 assert(tree
->root_color
== stone_other(child_color
));
220 if (!b
->crit_amaf
&& !is_pass(node
->coord
)) {
221 stats_add_result(&node
->winner_owner
, board_at(final_board
, node
->coord
) == winner_color
? 1.0 : 0.0, 1);
222 stats_add_result(&node
->black_owner
, board_at(final_board
, node
->coord
) == S_BLACK
? 1.0 : 0.0, 1);
224 stats_add_result(&node
->u
, result
, 1);
225 if (amaf_nakade(map
->map
[node
->coord
]))
226 amaf_op(map
->map
[node
->coord
], -);
228 /* This loop ignores symmetry considerations, but they should
229 * matter only at a point when AMAF doesn't help much. */
230 assert(map
->game_baselen
>= 0);
231 for (struct tree_node
*ni
= node
->children
; ni
; ni
= ni
->sibling
) {
232 enum stone amaf_color
= map
->map
[ni
->coord
];
233 assert(amaf_color
!= S_OFFBOARD
);
234 if (amaf_color
== S_NONE
)
236 if (amaf_nakade(map
->map
[ni
->coord
])) {
237 if (!b
->check_nakade
)
240 for (i
= map
->game_baselen
; i
< map
->gamelen
; i
++)
241 if (map
->game
[i
].coord
== ni
->coord
242 && map
->game
[i
].color
== child_color
)
244 if (i
== map
->gamelen
)
246 amaf_color
= child_color
;
249 floating_t nres
= result
;
250 if (amaf_color
!= child_color
) {
253 /* For child_color != player_color, we still want
254 * to record the result unmodified; in that case,
255 * we will correctly negate them at the descend phase. */
257 if (b
->crit_amaf
&& !is_pass(node
->coord
)) {
258 stats_add_result(&ni
->winner_owner
, board_at(final_board
, ni
->coord
) == winner_color
? 1.0 : 0.0, 1);
259 stats_add_result(&ni
->black_owner
, board_at(final_board
, ni
->coord
) == S_BLACK
? 1.0 : 0.0, 1);
261 stats_add_result(&ni
->amaf
, nres
, 1);
264 struct board bb
; bb
.size
= 9+2;
265 fprintf(stderr
, "* %s<%"PRIhash
"> -> %s<%"PRIhash
"> [%d/%f => %d/%f]\n",
266 coord2sstr(node
->coord
, &bb
), node
->hash
,
267 coord2sstr(ni
->coord
, &bb
), ni
->hash
,
268 player_color
, result
, child_color
, nres
);
272 if (!is_pass(node
->coord
)) {
275 node
= node
->parent
; child_color
= stone_other(child_color
);
281 policy_ucb1amaf_init(struct uct
*u
, char *arg
)
283 struct uct_policy
*p
= calloc2(1, sizeof(*p
));
284 struct ucb1_policy_amaf
*b
= calloc2(1, sizeof(*b
));
287 p
->choose
= uctp_generic_choose
;
288 p
->winner
= uctp_generic_winner
;
289 p
->evaluate
= ucb1rave_evaluate
;
290 p
->descend
= ucb1rave_descend
;
291 p
->update
= ucb1amaf_update
;
292 p
->wants_amaf
= true;
294 b
->explore_p
= 0; // 0.02 can be also good on 19x19 with prior=eqex=40
295 b
->equiv_rave
= 3000;
297 b
->check_nakade
= true;
298 b
->sylvain_rave
= true;
299 b
->ltree_rave
= 0.75f
;
302 b
->crit_min_playouts
= 2000;
303 b
->crit_negative
= 1;
306 b
->root_virtual_win
= -1;
307 b
->vwin_min_playouts
= 1000;
310 char *optspec
, *next
= arg
;
313 next
+= strcspn(next
, ":");
314 if (*next
) { *next
++ = 0; } else { *next
= 0; }
316 char *optname
= optspec
;
317 char *optval
= strchr(optspec
, '=');
318 if (optval
) *optval
++ = 0;
320 if (!strcasecmp(optname
, "explore_p")) {
321 b
->explore_p
= atof(optval
);
322 } else if (!strcasecmp(optname
, "fpu") && optval
) {
323 b
->fpu
= atof(optval
);
324 } else if (!strcasecmp(optname
, "equiv_rave") && optval
) {
325 b
->equiv_rave
= atof(optval
);
326 } else if (!strcasecmp(optname
, "sylvain_rave")) {
327 b
->sylvain_rave
= !optval
|| *optval
== '1';
328 } else if (!strcasecmp(optname
, "check_nakade")) {
329 b
->check_nakade
= !optval
|| *optval
== '1';
330 } else if (!strcasecmp(optname
, "ltree_rave") && optval
) {
331 b
->ltree_rave
= atof(optval
);
332 } else if (!strcasecmp(optname
, "crit_rave") && optval
) {
333 b
->crit_rave
= atof(optval
);
334 } else if (!strcasecmp(optname
, "crit_min_playouts") && optval
) {
335 b
->crit_min_playouts
= atoi(optval
);
336 } else if (!strcasecmp(optname
, "crit_negative")) {
337 b
->crit_negative
= !optval
|| *optval
== '1';
338 } else if (!strcasecmp(optname
, "crit_amaf")) {
339 b
->crit_amaf
= !optval
|| *optval
== '1';
340 } else if (!strcasecmp(optname
, "virtual_win") && optval
) {
341 b
->virtual_win
= atoi(optval
);
342 } else if (!strcasecmp(optname
, "root_virtual_win") && optval
) {
343 b
->root_virtual_win
= atoi(optval
);
344 } else if (!strcasecmp(optname
, "vwin_min_playouts") && optval
) {
345 b
->vwin_min_playouts
= atoi(optval
);
347 fprintf(stderr
, "ucb1amaf: Invalid policy argument %s or missing value\n",
353 if (b
->root_virtual_win
< 0)
354 b
->root_virtual_win
= b
->virtual_win
;