1 #ifndef ZZGO_UCT_INTERNAL_H
2 #define ZZGO_UCT_INTERNAL_H
4 /* Internal UCT structures */
7 #include "distributed/distributed.h"
19 /* How big proportion of ownermap counts must be of one color to consider
22 /* How many games to consider at minimum before judging groups. */
23 #define GJ_MINGAMES 500
25 /* Distributed stats for each child of the root node. */
27 struct move_stats2 last_sent_own
;
28 struct move_stats2 added_from_others
;
29 struct tree_node
*node
;
32 /* Internal engine state. */
38 double best2_ratio
, bestr_ratio
;
40 bool territory_scoring
;
42 bool playout_amaf
, playout_amaf_nakade
;
44 int playout_amaf_cutoff
;
49 unsigned long max_tree_size
;
53 enum uct_thread_model
{
54 TM_ROOT
, /* Root parallelization. */
55 TM_TREE
, /* Tree parallelization w/o virtual loss. */
56 TM_TREEVL
, /* Tree parallelization with virtual loss. */
60 bool pondering_opt
; /* User wants pondering */
61 bool pondering
; /* Actually pondering now */
62 bool slave
; /* Act as slave in distributed engine. */
70 struct uct_dynkomi
*dynkomi
;
76 int random_policy_chance
;
79 float local_tree_aging
;
80 bool local_tree_allseq
;
81 /* Playout-localtree integration. */
82 bool local_tree_playout
; // can be true only if ELO playout
83 bool local_tree_pseqroot
;
87 struct uct_policy
*policy
;
88 struct uct_policy
*random_policy
;
89 struct playout_policy
*playout
;
90 struct uct_prior
*prior
;
92 /* Used within frame of single genmove. */
93 struct board_ownermap ownermap
;
94 /* Used for coordination among slaves of the distributed engine. */
95 struct node_stats
*stats
;
97 int played_all
; /* games played by all slaves */
99 /* Game state - maintained by setup_state(), reset_state(). */
103 #define UDEBUGL(n) DEBUGL_(u->debug_level, n)
105 bool uct_pass_is_safe(struct uct
*u
, struct board
*b
, enum stone color
, bool pass_all_alive
);
107 void uct_prepare_move(struct uct
*u
, struct board
*b
, enum stone color
);
108 void uct_genmove_setup(struct uct
*u
, struct board
*b
, enum stone color
);
111 /* This is the state used for descending the tree; we use this wrapper
112 * structure in order to be able to easily descend in multiple trees
113 * in parallel (e.g. main tree and local tree) or compute cummulative
114 * "path value" throughout the tree descent. */
116 /* Active tree nodes: */
117 struct tree_node
*node
; /* Main tree. */
118 struct tree_node
*lnode
; /* Local tree. */
119 /* Value of main tree node (with all value factors, but unbiased
120 * - without exploration factor), from black's perspective. */
121 struct move_stats value
;
125 typedef struct tree_node
*(*uctp_choose
)(struct uct_policy
*p
, struct tree_node
*node
, struct board
*b
, enum stone color
, coord_t exclude
);
126 typedef float (*uctp_evaluate
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
);
127 typedef void (*uctp_descend
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
, int parity
, bool allow_pass
);
128 typedef void (*uctp_winner
)(struct uct_policy
*p
, struct tree
*tree
, struct uct_descent
*descent
);
129 typedef void (*uctp_prior
)(struct uct_policy
*p
, struct tree
*tree
, struct tree_node
*node
, struct board
*b
, enum stone color
, int parity
);
130 typedef void (*uctp_update
)(struct uct_policy
*p
, struct tree
*tree
, struct tree_node
*node
, enum stone node_color
, enum stone player_color
, struct playout_amafmap
*amaf
, float result
);
136 uctp_evaluate evaluate
;
137 uctp_descend descend
;