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 struct board_ownermap;
  11
  12
  13 /** Playout policy interface: */
  14
  15 struct playout_policy;
  16 struct playout_setup;
  17
  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);
  25
  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);
  28
  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);
  34
  35 /* Allow play of randomly selected move. */
  36 typedef bool (*playoutp_permit)(struct playout_policy *playout_policy, struct board *b, struct move *m);
  37
  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);
  40
  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;
  54 };
  55
  56
  57 /** Playout engine interface: */
  58
  59 /* Engine hook for forcing moves before doing policy decision.
  60  * Return pass to forward to policy. */
  61 typedef coord_t (*playouth_prepolicy)(struct playout_policy *playout_policy, struct playout_setup *setup, struct board *b, enum stone color);
  62
  63 /* Engine hook for choosing moves in case policy did not choose
  64  * a move.
  65  * Return pass to forward to uniformly random selection. */
  66 typedef coord_t (*playouth_postpolicy)(struct playout_policy *playout_policy, struct playout_setup *setup, struct board *b, enum stone color);
  67
  68 struct playout_setup {
  69         unsigned int gamelen; /* Maximal # of moves in playout. */
  70         /* Minimal difference between captures to terminate the playout.
  71          * 0 means don't check. */
  72         unsigned int mercymin;
  73
  74         void *hook_data; // for hook to reference its state
  75         playouth_prepolicy prepolicy_hook;
  76         playouth_postpolicy postpolicy_hook;
  77 };
  78
  79
  80 struct playout_amafmap {
  81         /* Record of the random playout - for each intersection:
  82          * S_NONE: This move was never played
  83          * S_BLACK: This move was played by black first
  84          * S_WHITE: This move was played by white first
  85          */
  86         enum stone *map; // [board_size2(b)]
  87
  88         /* the lowest &0xf is the enum stone, upper bits are nakade
  89          * counter - in case of nakade, we record only color of the
  90          * first stone played inside, but count further throwins
  91          * and ignore AMAF value after these. */
  92 #define amaf_nakade(item_) (item_ >> 8)
  93 #define amaf_op(item_, op_) do { \
  94                 int mi_ = item_; \
  95                 item_ = (mi_ & 0xf) | ((amaf_nakade(mi_) op_ 1) << 8); \
  96 } while (0)
  97
  98         /* Additionally, we keep record of the game so that we can
  99          * examine nakade moves; really going out of our way to
 100          * implement nakade AMAF properly turns out to be crucial
 101          * when reading some tactical positions in depth (even if
 102          * they are just one-stone-snapback). */
 103         struct move game[MAX_GAMELEN + 1];
 104         unsigned int gamelen;
 105         /* Our current position in the game sequence; in AMAF, we search
 106          * the range [game_baselen, gamelen]. */
 107         unsigned int game_baselen;
 108
 109         /* Whether to record the nakade moves (true) or just completely
 110          * ignore them (false; just the first color on the intersection
 111          * is stored in the map, nakade counter is not incremented; game
 112          * record is still kept). */
 113         bool record_nakade;
 114 };
 115
 116
 117 /* >0: starting_color wins, <0: starting_color loses; the actual
 118  * number is a DOUBLE of the score difference
 119  * 0: superko inside the game tree (XXX: jigo not handled) */
 120 int play_random_game(struct playout_setup *setup,
 121                      struct board *b, enum stone starting_color,
 122                      struct playout_amafmap *amafmap,
 123                      struct board_ownermap *ownermap,
 124                      struct playout_policy *policy);
 125
 126 #endif