playout.h

   1 #ifndef PACHI_PLAYOUT_H
   2 #define PACHI_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         /* By default, with setboard set we will refuse to make (random)
  53          * moves outside of the *choose routine in order not to mess up
  54          * state tracking. If you use *setboard but do not track state
  55          * (e.g. you just initialize some per-playout data, like the Moggy
  56          * policy), set setboard_randomok too. */
  57         bool setboard_randomok;
  58         /* Particular playout policy's internal data. */
  59         void *data;
  60 };
  61
  62
  63 /** Playout engine interface: */
  64
  65 /* Engine hook for forcing moves before doing policy decision.
  66  * Return pass to forward to policy. */
  67 typedef coord_t (*playouth_prepolicy)(struct playout_policy *playout_policy, struct playout_setup *setup, struct board *b, enum stone color);
  68
  69 /* Engine hook for choosing moves in case policy did not choose
  70  * a move.
  71  * Return pass to forward to uniformly random selection. */
  72 typedef coord_t (*playouth_postpolicy)(struct playout_policy *playout_policy, struct playout_setup *setup, struct board *b, enum stone color);
  73
  74 struct playout_setup {
  75         unsigned int gamelen; /* Maximal # of moves in playout. */
  76         /* Minimal difference between captures to terminate the playout.
  77          * 0 means don't check. */
  78         int mercymin;
  79
  80         void *hook_data; // for hook to reference its state
  81         playouth_prepolicy prepolicy_hook;
  82         playouth_postpolicy postpolicy_hook;
  83 };
  84
  85
  86 struct playout_amafmap {
  87         /* We keep record of the game so that we can
  88          * examine nakade moves; really going out of our way to
  89          * implement nakade AMAF properly turns out to be crucial
  90          * when reading some tactical positions in depth (even if
  91          * they are just one-stone-snapback). */
  92         coord_t game[MAX_GAMELEN];
  93         bool is_ko_capture[MAX_GAMELEN];
  94         int gamelen;
  95         /* Our current position in the game sequence; in AMAF, we search
  96          * the range [game_baselen, gamelen[ */
  97         int game_baselen;
  98 };
  99
 100
 101 /* >0: starting_color wins, <0: starting_color loses; the actual
 102  * number is a DOUBLE of the score difference
 103  * 0: superko inside the game tree (XXX: jigo not handled) */
 104 int play_random_game(struct playout_setup *setup,
 105                      struct board *b, enum stone starting_color,
 106                      struct playout_amafmap *amafmap,
 107                      struct board_ownermap *ownermap,
 108                      struct playout_policy *policy);
 109
 110 coord_t play_random_move(struct playout_setup *setup,
 111                          struct board *b, enum stone color,
 112                          struct playout_policy *policy);
 113
 114 void playout_policy_done(struct playout_policy *p);
 115
 116 #endif