With TD_GAMES, perform an exact number of playouts
[pachi/derm.git] / playout.h
blob4bd2101daa580ffc3a45550c80627d6b565f6a91
1 #ifndef ZZGO_PLAYOUT_H
2 #define ZZGO_PLAYOUT_H
4 #define MAX_GAMELEN 600
6 struct board;
7 struct move;
8 enum stone;
9 struct prior_map;
10 struct board_ownermap;
13 /** Playout policy interface: */
15 struct playout_policy;
16 struct playout_setup;
18 /* Initialize policy data structures for new playout; subsequent choose calls
19 * (but not assess/permit calls!) will all be made on the same board; if
20 * setboard is used, it is guaranteed that choose will pick all moves played
21 * on the board subsequently. The routine is expected to initialize b->ps
22 * with internal data. At the playout end, b->ps will be simply free()d,
23 * so make sure all data is within single allocated block. */
24 typedef void (*playoutp_setboard)(struct playout_policy *playout_policy, struct board *b);
26 /* Pick the next playout simulation move. */
27 typedef coord_t (*playoutp_choose)(struct playout_policy *playout_policy, struct playout_setup *playout_setup, struct board *b, enum stone to_play);
29 /* Set number of won (>0) or lost (<0) games for each considerable
30 * move (usually a proportion of @games); can leave some untouched
31 * if policy has no opinion. The number must have proper parity;
32 * just use uct/prior.h:add_prior_value(). */
33 typedef void (*playoutp_assess)(struct playout_policy *playout_policy, struct prior_map *map, int games);
35 /* Allow play of randomly selected move. */
36 typedef bool (*playoutp_permit)(struct playout_policy *playout_policy, struct board *b, struct move *m);
38 /* Tear down the policy state; policy and policy->data will be free()d by caller. */
39 typedef void (*playoutp_done)(struct playout_policy *playout_policy);
41 struct playout_policy {
42 int debug_level;
43 /* We call setboard when we start new playout.
44 * We call choose when we ask policy about next move.
45 * We call assess when we ask policy about how good given move is.
46 * We call permit when we ask policy if we can make a randomly chosen move. */
47 playoutp_setboard setboard;
48 playoutp_choose choose;
49 playoutp_assess assess;
50 playoutp_permit permit;
51 playoutp_done done;
52 /* Particular playout policy's internal data. */
53 void *data;
57 /** Playout engine interface: */
59 struct playout_setup {
60 unsigned int gamelen; /* Maximal # of moves in playout. */
61 /* Minimal difference between captures to terminate the playout.
62 * 0 means don't check. */
63 unsigned int mercymin;
65 /* XXX: We used to have more, perhaps we will again have more
66 * in the future. */
70 struct playout_amafmap {
71 /* Record of the random playout - for each intersection:
72 * S_NONE: This move was never played
73 * S_BLACK: This move was played by black first
74 * S_WHITE: This move was played by white first
76 enum stone *map; // [board_size2(b)]
78 /* the lowest &0xf is the enum stone, upper bits are nakade
79 * counter - in case of nakade, we record only color of the
80 * first stone played inside, but count further throwins
81 * and ignore AMAF value after these. */
82 #define amaf_nakade(item_) (item_ >> 8)
83 #define amaf_op(item_, op_) do { \
84 int mi_ = item_; \
85 item_ = (mi_ & 0xf) | ((amaf_nakade(mi_) op_ 1) << 8); \
86 } while (0)
88 /* Additionally, we keep record of the game so that we can
89 * examine nakade moves; really going out of our way to
90 * implement nakade AMAF properly turns out to be crucial
91 * when reading some tactical positions in depth (even if
92 * they are just one-stone-snapback). */
93 struct move game[MAX_GAMELEN + 1];
94 unsigned int gamelen;
95 /* Our current position in the game sequence; in AMAF, we search
96 * the range [game_baselen, gamelen]. */
97 unsigned int game_baselen;
99 /* Whether to record the nakade moves (true) or just completely
100 * ignore them (false; just the first color on the intersection
101 * is stored in the map, nakade counter is not incremented; game
102 * record is still kept). */
103 bool record_nakade;
107 /* >0: starting_color wins, <0: starting_color loses; the actual
108 * number is a DOUBLE of the score difference
109 * 0: superko inside the game tree (XXX: jigo not handled) */
110 int play_random_game(struct playout_setup *setup,
111 struct board *b, enum stone starting_color,
112 struct playout_amafmap *amafmap,
113 struct board_ownermap *ownermap,
114 struct playout_policy *policy);
116 #endif