playout.h

   1 #ifndef ZZGO_PLAYOUT_H
   2 #define ZZGO_PLAYOUT_H
   3
   4 struct board;
   5 struct move;
   6 enum stone;
   7 struct prior_map;
   8
   9
  10 struct playout_policy;
  11 typedef coord_t (*playoutp_choose)(struct playout_policy *playout_policy, struct board *b, enum stone to_play);
  12 /* Set number of won (>0) or lost (<0) games for each considerable
  13  * move (usually a proportion of @games); can leave some untouched
  14  * if policy has no opinion. The number must have proper parity;
  15  * just use uct/prior.h:add_prior_value(). */
  16 typedef void (*playoutp_assess)(struct playout_policy *playout_policy, struct prior_map *map, int games);
  17 typedef bool (*playoutp_permit)(struct playout_policy *playout_policy, struct board *b, struct move *m);
  18
  19 struct playout_policy {
  20         int debug_level;
  21         /* We call choose when we ask policy about next move.
  22          * We call assess when we ask policy about how good given move is.
  23          * We call permit when we ask policy if we can make a randomly chosen move. */
  24         playoutp_choose choose;
  25         playoutp_assess assess;
  26         playoutp_permit permit;
  27         void *data;
  28 };
  29
  30
  31 struct playout_amafmap {
  32         /* Record of the random playout - for each intersection:
  33          * S_NONE: This move was never played
  34          * S_BLACK: This move was played by black first
  35          * S_WHITE: This move was played by white first
  36          */
  37         enum stone *map; // [board_size2(b)]
  38
  39         /* the lowest &0xf is the enum stone, upper bits are nakade
  40          * counter - in case of nakade, we record only color of the
  41          * first stone played inside, but count further throwins
  42          * and ignore AMAF value after these. */
  43 #define amaf_nakade(item_) (item_ >> 8)
  44 #define amaf_op(item_, op_) do { \
  45                 int mi_ = item_; \
  46                 item_ = (mi_ & 0xf) | ((amaf_nakade(mi_) op_ 1) << 8); \
  47 } while (0)
  48
  49         /* Additionally, we keep record of the game so that we can
  50          * examine nakade moves; really going out of our way to
  51          * implement nakade AMAF properly turns out to be crucial
  52          * when reading some tactical positions in depth (even if
  53          * they are just one-stone-snapback). */
  54         struct move game[512];
  55         int gamelen;
  56         /* Our current position in the game sequence; in AMAF, we search
  57          * the range [game_baselen, gamelen]. */
  58         int game_baselen;
  59
  60         /* Whether to record the nakade moves (true) or just completely
  61          * ignore them (false; just the first color on the intersection
  62          * is stored in the map, nakade counter is not incremented; game
  63          * record is still kept). */
  64         bool record_nakade;
  65 };
  66
  67
  68 /* 1: starting_color wins, 0: starting_color loses
  69  * -1: superko inside the game tree */
  70 int play_random_game(struct board *b, enum stone starting_color, int gamelen, struct playout_amafmap *amafmap, struct playout_policy *policy);
  71
  72 #endif