playout.h

   1 #ifndef ZZGO_PLAYOUT_H
   2 #define ZZGO_PLAYOUT_H
   3
   4 #define MAX_GAMELEN 600
   5
   6 struct board;
   7 struct move;
   8 enum stone;
   9 struct prior_map;
  10
  11
  12 struct playout_policy;
  13 typedef coord_t (*playoutp_choose)(struct playout_policy *playout_policy, struct board *b, enum stone to_play);
  14 /* Set number of won (>0) or lost (<0) games for each considerable
  15  * move (usually a proportion of @games); can leave some untouched
  16  * if policy has no opinion. The number must have proper parity;
  17  * just use uct/prior.h:add_prior_value(). */
  18 typedef void (*playoutp_assess)(struct playout_policy *playout_policy, struct prior_map *map, int games);
  19 typedef bool (*playoutp_permit)(struct playout_policy *playout_policy, struct board *b, struct move *m);
  20
  21 struct playout_policy {
  22         int debug_level;
  23         /* We call choose when we ask policy about next move.
  24          * We call assess when we ask policy about how good given move is.
  25          * We call permit when we ask policy if we can make a randomly chosen move. */
  26         playoutp_choose choose;
  27         playoutp_assess assess;
  28         playoutp_permit permit;
  29         void *data;
  30 };
  31
  32
  33 struct playout_amafmap {
  34         /* Record of the random playout - for each intersection:
  35          * S_NONE: This move was never played
  36          * S_BLACK: This move was played by black first
  37          * S_WHITE: This move was played by white first
  38          */
  39         enum stone *map; // [board_size2(b)]
  40
  41         /* the lowest &0xf is the enum stone, upper bits are nakade
  42          * counter - in case of nakade, we record only color of the
  43          * first stone played inside, but count further throwins
  44          * and ignore AMAF value after these. */
  45 #define amaf_nakade(item_) (item_ >> 8)
  46 #define amaf_op(item_, op_) do { \
  47                 int mi_ = item_; \
  48                 item_ = (mi_ & 0xf) | ((amaf_nakade(mi_) op_ 1) << 8); \
  49 } while (0)
  50
  51         /* Additionally, we keep record of the game so that we can
  52          * examine nakade moves; really going out of our way to
  53          * implement nakade AMAF properly turns out to be crucial
  54          * when reading some tactical positions in depth (even if
  55          * they are just one-stone-snapback). */
  56         struct move game[MAX_GAMELEN + 1];
  57         int gamelen;
  58         /* Our current position in the game sequence; in AMAF, we search
  59          * the range [game_baselen, gamelen]. */
  60         int game_baselen;
  61
  62         /* Whether to record the nakade moves (true) or just completely
  63          * ignore them (false; just the first color on the intersection
  64          * is stored in the map, nakade counter is not incremented; game
  65          * record is still kept). */
  66         bool record_nakade;
  67 };
  68
  69
  70 /* 1: starting_color wins, 0: starting_color loses
  71  * -1: superko inside the game tree */
  72 int play_random_game(struct board *b, enum stone starting_color, int gamelen, struct playout_amafmap *amafmap, struct playout_policy *policy);
  73
  74 #endif